lexicon/terms/math-core.json — lexicon

RENDERED FROM STRUCTURED DATA — this page is not a markdown file. Source: lexicon/terms/math-core.json in uni-cookbook at 575fc93d9d31. The sha256 recorded for this page is the digest of that JSON, not of this text. The renderer is generators/derive_docs.cjs; it copies fields and adds no facts. Every value below is a field of the source, verbatim, including the ones that read NOT RUN and PENDING.

The five-register lexicon entries (Sanskrit / Latin / English / Hindi / Spanish) for the concepts whose CONCEPTS.json domain is 'active-inference-math'. Each entry renders ONE concept fixed by its definition in lexicon/CONCEPTS.json, which is the join key and the fence. NOTHING HERE IS CLAIMED AS A PERFECT TRANSLATION. Per rule 18.14 of the operator's charter ('Do not claim perfect translation') and the twin admission tests, every rendering enters PENDING / SUB EXAMINE and is admitted only against a cited dictionary headword plus a passing back-translation test. NO back-translation test in this file has been run: every back_translation_test is a PENDING falsifier, not a discharged gate, and no entry may be read as admitted to the Tree. This file raises no UNI rung, discharges no gate, and states no fact about nature. A nature citation is never a UNI gate.

authored 2026-07-15 · reviewer: NONE - PENDING EXPERT REVIEW. No Sanskritist, Latinist, Hindi or Spanish lexicographer has read any entry in this file. Every 'confidence' value below is the author's own and is therefore Class-G own narrative, which M9 ranks BELOW Class-B tool state; where a status says CITED, believe the quoted dictionary text (Class-B), not the confidence field. · counts: {"entries":29,"source_domain":"active-inference-math","source_domain_count_in_CONCEPTS":29,"status_census":{"sanskrit":{"COINED":28,"CITED":1},"latin":{"COINED":29},"hindi":{"PENDING-CITATION":29},"spanish":{"PENDING-CITATION":29}},"coined_registers_lacking_derivation":[]}

Falsifiers for this file

• Any register marked CITED whose named authority does not carry that headword, or does not carry it in the claimed sense. That entry is a pseudopropheta and must be struck to PENDING-CITATION immediately.
• Any COINED entry lacking a morphological derivation, or whose derivation invokes a morpheme that its authority does not carry in the claimed sense.
• Any back_translation_test being reported as PASSED. None has been run; every one is PENDING.
• Any entry whose rendering lets a doctrinal metaphysics (Samkhya, Vedanta, Nyaya, Yoga) enter silently without an ambiguity_notes record of the freight.
• Any entry that renders VARIATIONAL_FREE_ENERGY or EXPECTED_FREE_ENERGY through a heat/energy root in any register, which imports a false physical claim the concept explicitly forbids.
• Any entry that renders GENERATIVE_MODEL and GENERATIVE_PROCESS with one term, which CONCEPTS.json registers as a falsifier of the whole lexicon.
• Any entry that renders the technical 'true' of TRUE_POSTERIOR with the term used for the TRUE store, which would make 'a True signal can be wrong' unsayable.

ACTIVE_INFERENCE — active inference

Store: TRUE · Domain: active-inference-math · Confidence: LOW. The Sanskrit head morpheme is CITED but the compound is a coinage with no attested usage and unverified modifier morphemes; the Latin is a coinage with NO verified component. Class-G own narrative. · Reviewer: NONE - PENDING EXPERT REVIEW

A scheme in which perception minimizes variational free energy over beliefs and action minimizes expected free energy over policies, so that perceiving and acting are one operation applied through the two halves of a blanket.

register	form	literal gloss	status
sanskrit	सक्रियानुमान (sakriyānumāna)	action-accompanied inferring; inferring that is carried on with action	COINED
latin	inferentia activa	active inferring	COINED
english	active inference	A scheme in which perception minimizes variational free energy over beliefs and action minimizes expected free energy over policies, so that perceiving and acting are one operation applied through the two halves of a Markov blanket.	—
hindi	सक्रिय अनुमान ()	active inference	PENDING-CITATION
spanish	inferencia activa	active inference	PENDING-CITATION

Ambiguity notes

• 'Active' does NOT mean energetic, busy, or vigorous. It means 'through the active states of the blanket' -- the half of the blanket the interior writes to. A register that renders it with a vigour/energy root has named the wrong thing.
• The English name misleads even in English: it reads as though 'inference' were the noun and 'active' a flavour of it. The concept is the reverse claim -- that action IS inference, not that there is a lively kind of inference.
• Sanskrit anumāna carries heavy Nyāya freight: it is one of the pramāṇas (means of valid knowledge) and in that system it is a syllogistic, five-membered public procedure with a fixed structure (pratijñā, hetu, udāharaṇa, upanaya, nigamana). Active inference is a continuous variational optimisation, not a syllogism. The freight is real and is not neutralised by the compound.
• IN THE UNI PROGRAM this is the framing LENS and not a running mechanism: no AIF loop exists in the Rust crate and G is not computed as a scheduler. A rendering in any register that implies UNI runs active inference is drift, regardless of how good the philology is.

Forbidden mistranslations

register	form	why it is wrong
sanskrit	kriyānumāna read as 'inference about action'	Reverses the claim. Action is not the OBJECT of the inference; action is one of the two routes by which the same quantity is minimized. A genitive/objective reading destroys the concept.
sanskrit	any rendering built on buddhi	MW gives buddhi as 'intelligence, reason, intellect, mind, discernment, judgement' and it is the Sāṃkhya Intellect -- a tattva, an evolute of prakṛti. It imports a whole metaphysics and an experiencer. The formalism admits partitions for a Newton's cradle.
latin	ratiocinatio activa	Ratiocinatio names discursive human reasoning. The scheme is not a style of reasoning and applies to systems that do not reason at all.
english	'active learning'	A different, established machine-learning term (querying an oracle for labels). Collision would silently reroute the reader to a different literature.

Back-translation test — status: NOT RUN: no reader has been asked

Give a Sanskrit-literate reader who has NOT seen the source term only the coined form 'sakriyānumāna' and its literal gloss, and ask three questions: (a) does this name a scheme in which the same quantity is minimized both by revising a belief and by changing the world? (b) does it name inference whose SUBJECT MATTER is action? (c) does it presuppose someone who reasons?

VARIATIONAL_FREE_ENERGY — variational free energy

Store: TRUE · Domain: active-inference-math · Confidence: LOW for the coinages; the Latin head noun mensura is CITED and the Sanskrit measure-root is CITED only at one remove (inside pramāṇa and mātrā, not as a standalone māna headword). The Hindi/Spanish calques are recorded as usage, unverified, and are believed to CONFLICT with the concept. · Reviewer: NONE - PENDING EXPERT REVIEW

A scalar functional F[q,o] = E_q(s)[ln q(s) - ln p(o,s)], in nats, that scores a belief q(s) against observations o and upper-bounds surprisal, so minimizing it over q both drives q toward the true posterior and tightens the bound.

register	form	literal gloss	status
sanskrit	वैकल्पिकमुक्तमान (vaikalpika-mukta-māna)	variational free MEASURE -- deliberately 'measure', not 'energy'	COINED
latin	mensura libera variationalis	variational free measure -- again 'measure', not 'energy'	COINED
english	variational free energy	A scalar functional F[q,o] = E_q(s)[ln q(s) - ln p(o,s)], measured in nats, that scores a belief q(s) against observations o and upper-bounds surprisal; minimizing it over q both drives q toward the true posterior and tightens the bound.	—
hindi	विचरणात्मक मुक्त ऊर्जा ()	variational free energy (calque, INCLUDING the energy root -- see ambiguity_notes)	PENDING-CITATION
spanish	energía libre variacional	variational free energy (calque, INCLUDING the energy root -- see ambiguity_notes)	PENDING-CITATION

Ambiguity notes

• THE CENTRAL TRAP OF THIS ENTRY: 'free energy' is a physics LOAN and is NOT thermodynamic free energy. Any register that renders it through a heat/energy root imports a false physical claim. The Sanskrit and Latin entries above therefore deliberately DECLINE the calque and use a measure-root instead; the Hindi and Spanish entries record that the standing usage in those languages DOES take the energy root, which is a live conflict between attested usage and the concept, and is recorded rather than resolved by fiat.
• This is a two-front problem and both fronts must travel: rendering F with an energy word is false to the physics, but refusing the standing calque makes the term unrecognisable to anyone who reads the literature. This file does not have the authority to settle that; an expert reviewer does.
• F scores BELIEFS, never actions. Conflating F with G (expected free energy) is named in CONCEPTS.json as the commonest error in applying the framework. No register may render F and G with one term.
• 'Variational' is not 'variable', 'varying', or 'variant'. It refers to the calculus of variations: F is a functional over a FAMILY of candidate distributions q, and is optimised over that family.
• F is a functional (it eats a function q and returns a scalar), not a function of a number. Registers with a strong noun/verb distinction should not render it as an act of measuring; it is the measure yielded.

Forbidden mistranslations

register	form	why it is wrong
sanskrit	mukta-ūrjā / muktorjā (मुक्तोर्जा), 'free energy'	ūrjā is energy/vigour. This is the single forbidden rendering for this concept: it asserts a thermodynamic quantity that F is not. It would make the corpus's own honesty rail unsayable in Sanskrit.
sanskrit	any rendering on tejas, śakti, or bala	All are power/energy/vigour roots, and śakti additionally carries Tantric and Sāṃkhya doctrinal freight. Same false physical claim, plus a metaphysics.
latin	energia libera	Imports the thermodynamic reading directly. Additionally, 'energia' in Latin is a Greek loan whose classical use is a rhetorical quality of style, not a physical quantity -- so it is wrong twice over.
any	rendering F and G with one term, or with a shared head plus only a modifier distinction that a reader may drop	F scores beliefs, G scores policies. CONCEPTS.json registers their conflation as the commonest error in the field.
english	'Helmholtz free energy' / 'Gibbs free energy'	Real thermodynamic quantities. The name-collision is the whole reason the fence exists.

Back-translation test — status: NOT RUN: no reader has been asked

Hand a physicist who does not know active inference ONLY the target-register term and its literal gloss, and ask: (a) is this a quantity of heat, work, or thermodynamics? (b) does it score a belief or an action? (c) is it a number, or a rule for producing a number from a candidate distribution?

EXPECTED_FREE_ENERGY — expected free energy

Store: TRUE · Domain: active-inference-math · Confidence: VERY LOW for Sanskrit -- the proposed coinage is built on a morpheme (pratyāśā) that the author BELIEVES carries a forbidden affective sense and could not verify in this pass. It is offered flagged, not endorsed. This is the weakest entry in the file and is the first one an expert should overturn. · Reviewer: NONE - PENDING EXPERT REVIEW

A scalar G(pi), in nats, that scores a candidate policy rather than a belief, equal to risk plus ambiguity and equivalently to the negative sum of epistemic and pragmatic value, so that the lowest-G policy is the chosen one.

register	form	literal gloss	status
sanskrit	प्रत्याशितमुक्तमान (pratyāśita-mukta-māna)	expected free MEASURE (of a policy) -- 'measure', not 'energy'	COINED
latin	mensura libera exspectata	expected free measure	COINED
english	expected free energy	A scalar G(pi), in nats, that scores a candidate policy rather than a belief, equal to risk plus ambiguity and equivalently to the negative sum of epistemic and pragmatic value; the lowest-G policy is the chosen one.	—
hindi	प्रत्याशित मुक्त ऊर्जा ()	expected free energy (calque, carrying the forbidden energy root)	PENDING-CITATION
spanish	energía libre esperada	expected free energy (calque, carrying the forbidden energy root)	PENDING-CITATION

Ambiguity notes

• 'Expected' is the MATHEMATICAL EXPECTATION OPERATOR E[.], a probability-weighted average. It is not hope, not anticipation, not waiting, and not what one supposes will happen. This is the sharpest translation trap in the entry, and Sanskrit apekṣā and pratyāśā, Hindi pratyāśā, Spanish 'esperada' (from esperar, which ALSO means 'to hope') all carry the affective sense. Spanish is the worst offender: 'esperada' is ambiguous between 'expected' and 'hoped-for' in ordinary usage.
• G scores POLICIES; F scores BELIEFS. They are different objects and must never share a term. The 'free energy' element is the same physics loan as in F and is forbidden through a heat/energy root for the same reason.
• NAME COLLISION ON RECORD, carried from CONCEPTS.json: Douady & Couder's phyllotaxis control parameter G_DC = v0*T/r0 is a completely unrelated quantity that also goes by G. A register must not let the letter do the work.
• The lowest G is chosen, not the highest. A register that renders G with a value/merit/worth root (implying more is better) inverts the selection rule. This is why a 'value' head was declined here even though 'epistemic value' and 'pragmatic value' use one -- those are the NEGATED components.
• G is 'expected' in a second sense too, which compounds the trap: it is expected under the agent's OWN predictive distribution q(o|pi), so it is an expectation taken with respect to a belief that may be wrong. It is not what will happen.

Forbidden mistranslations

register	form	why it is wrong
sanskrit	apekṣā-based rendering, e.g. apekṣita-	MW s.v. apekṣā (retrieved 2026-07-15) gives 'looking for, expectation, HOPE, need, requirement' and 'dependence on'. The hope sense asserts an experiencer who wants an outcome; the dependence sense asserts a causal relation. The mathematics contains a weighted average and nothing else.
sanskrit	sambhāvanā-based rendering	MW s.v. sambhāvanā (retrieved 2026-07-15) gives 'imagination, supposition, assumption; fitness, adequacy; competency, ability' and also 'worship, honour, respect'. MW does NOT carry a mathematical probability sense. Using it for the expectation operator would assert a dictionary sense the dictionary does not have -- the pseudopropheta failure mode exactly.
spanish	'energía libre esperanzada' or reading 'esperada' as 'hoped-for'	esperar is 'to hope' as well as 'to expect'. The mathematics has no hope in it. If Spanish cannot disambiguate, the ambiguity must be recorded, not translated away.
any	any rendering implying the highest-G policy is chosen	Inverts the selection rule. G is minimized.
any	any heat/energy root	Same false physical claim as in VARIATIONAL_FREE_ENERGY.

Back-translation test — status: NOT RUN: no reader has been asked

Hand a bilingual reader with undergraduate probability but no active inference ONLY the target-register term and its gloss. Ask: (a) does 'expected' here mean a probability-weighted average, or does it mean hoped-for/anticipated? (b) does this score a plan or a belief? (c) is the best plan the one with the highest or the lowest value of this quantity?

SURPRISAL — surprisal

Store: TRUE · Domain: active-inference-math · Confidence: LOW-MEDIUM. Unusually, the FORBIDDEN entries here are the best-evidenced part: vismaya and admiratio are both CITED from their authorities in this pass and both are confirmed affect words, so the fence is on solid ground even though the proposed coinages are not. · Reviewer: NONE - PENDING EXPERT REVIEW

The self-information -ln p(o) of an observation under a model, measuring how unexpected that observation was, in nats, and being the quantity that variational free energy upper-bounds.

register	form	literal gloss	status
sanskrit	अप्रत्याशितमात्रा (apratyāśita-mātrā)	unexpectedness-quantity; the measure of how far an observation fell outside what the model gave weight to	COINED
latin	improvisum (mensura improvisi)	the unforeseen (thing); the measure of the unforeseen	COINED
english	surprisal	The self-information -ln p(o) of an observation under a model, in nats, measuring how unexpected that observation was under that model; it is the quantity that variational free energy upper-bounds.	—
hindi	आश्चर्य-मात्रा ()	surprise-quantity -- PROPOSED WITH A WARNING: āścarya is an affect word and may fail this entry's own fence	PENDING-CITATION
spanish	sorpresa (información propia / autoinformación)	surprisal; the standing Spanish information-theory term is 'sorpresa' or 'autoinformación', the latter avoiding the affect root	PENDING-CITATION

Ambiguity notes

• THE DEFINING TRAP: surprisal is NOT the felt emotion of surprise. A register that renders it with an affect root asserts an experiencer, and CONCEPTS.json names that as a lane-crossing -- the cardinal sin of merging the TRUE and HONEST stores. -ln p(o) is defined for a thermostat and for a burning fuse, neither of which is surprised by anything.
• The English word 'surprisal' is a deliberate technical coinage precisely to hold this fence: information theorists did not say 'surprise'. Any register that reaches for its ordinary surprise-word has undone the work the English coinage did, and this is why the Sanskrit above takes a mātrā (quantity) head rather than a vismaya (wonder) head.
• Distinct from ENTROPY, which is EXPECTED surprisal -- the average of this quantity over the distribution. Surprisal attaches to ONE observation; entropy attaches to a distribution. No register may merge them.
• Surprisal is relative to a MODEL, never absolute. -ln p(o) presupposes the p. The same observation carries different surprisal under different models, so a register implying an intrinsic property of the event is wrong.
• The Hindi and Spanish entries above are recorded as PROBABLY FAILING this entry's own fence, because their standing usage is affect-rooted. That conflict is printed rather than resolved.

Forbidden mistranslations

register	form	why it is wrong
sanskrit	vismaya (विस्मय)	MW s.v. vismaya (retrieved 2026-07-15) gives m. 'wonder, surprise, amazement, bewilderment, perplexity' -- and additionally 'pride, arrogance' and 'doubt, uncertainty'. It is an affect/rasa word (vismaya is the sthāyibhāva of the adbhuta rasa). Using it asserts an experiencer AND drags in aesthetic-theory freight AND collides with the doubt sense used for AMBIGUITY. Wrong three ways.
sanskrit	āścarya	Wonder/marvel. Same affect problem, plus a connotation of the marvellous that -ln p(o) certainly lacks: a highly surprising observation is not a marvel, it is a low-probability one.
latin	admiratio	Lewis & Short s.v. admiratio (retrieved 2026-07-15) gives I 'An admiring; admiration' and II 'Wonder, surprise, astonishment'. Both senses are affective and the first is positively evaluative. A thermostat's -ln p(o) admires nothing.
english	'surprise'	The very confusion the coinage 'surprisal' was minted to prevent. Reintroducing it in English defeats the fence at the source language.
any	rendering surprisal and entropy alike	Entropy is EXPECTED surprisal. One is a property of a single observation, the other of a whole distribution.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader of the target register ONLY the term and its gloss and ask: (a) does this word describe something a person or animal FEELS? (b) could this quantity be computed for a thermostat, with no one feeling anything? (c) does it attach to one observation, or to a whole distribution of possible observations?

GENERATIVE_MODEL — generative model

Store: TRUE · Domain: active-inference-math · Confidence: LOW. The janaka modifier is genuinely CITED in the required adjectival sense, which is the strongest Sanskrit result in this file; but MW's own 'chiefly ifc.' note is a live grammatical objection to the compound as written, and it is recorded above rather than suppressed. · Reviewer: NONE - PENDING EXPERT REVIEW

The joint distribution p(o,s) that an agent HOLDS over observations and hidden states, specifying how states are believed to cause observations and how states are believed to evolve, and which is a choice the agent makes rather than a measurement of the world.

register	form	literal gloss	status
sanskrit	जनकप्रतिरूप (janaka-pratirūpa)	generating counterpart/image -- the HELD image that generates expected observations	COINED
latin	exemplar generativum	generating model/pattern	COINED
english	generative model	The joint distribution p(o,s) that an agent holds over observations and hidden states, specifying how states are believed to cause observations and how states are believed to evolve; it is a choice the agent makes, not a measurement of the world.	—
hindi	जनक मॉडल ()	generative model	PENDING-CITATION
spanish	modelo generativo	generative model	PENDING-CITATION

Ambiguity notes

• HALF OF THE FIELD'S MOST-CONFUSED PAIR. This entry is meaningless unless read against GENERATIVE_PROCESS, and CONCEPTS.json registers it as an outright FALSIFIER of this lexicon if any register renders the two with one term. The model is HELD by the agent; the process is what actually happens. Every register below must keep two words.
• 'Generative' does NOT mean creative, productive, fertile, or generous. It means: it specifies a joint distribution from which observations could be generated. A register reaching for a fertility/begetting root -- which the Sanskrit janaka literally is, MW giving 'begetting' and m. 'a progenitor, father' -- must fence that sense off explicitly.
• CRITICAL AND EASILY MISSED, carried from NA-04: the generative model is NOT a model of the external states themselves. It is a model of how external states are expected to SHOW UP IN THE BLANKET. A register that renders it as 'a model of the world' asserts exactly the reach that the Markov blanket denies.
• 'Model' in the ordinary senses -- an exemplar to be imitated, a person who poses, a scale replica -- are all live homonyms in every one of these five registers. The Latin 'exemplar' is especially exposed: its ordinary sense is a pattern to be COPIED, which reverses the direction of fit.
• The definition's sting is in its tail: it is A CHOICE THE AGENT MAKES, not a measurement. Minimizing F never tests whether this model was the right one (NA-02, honest bound 1). A register implying the model is read off the world destroys the BOUND_GAP concept downstream.

Forbidden mistranslations

register	form	why it is wrong
any	ANY term also used for GENERATIVE_PROCESS	Registered in CONCEPTS.json as a falsifier of the lexicon itself. If model and process share a word, 'a low F is not a correctness certificate' becomes unsayable, because F is computed against the model while the process is what actually generated o.
sanskrit	reading janaka as m. 'progenitor, father'	MW gives exactly that as a substantive sense. A joint distribution is not a father. The adjectival sense 'generative, generating, producing, causing' is the intended one and the substantive must be fenced off.
sanskrit	prakṛti-based rendering	prakṛti is the Sāṃkhya primordial nature -- one of the two ultimate principles. It imports an entire metaphysics AND it names the world-side generator, which is GENERATIVE_PROCESS's territory, not the agent's held model. Wrong twice.
latin	'exemplar' read in its ordinary sense of a pattern to be imitated	Reverses the direction of fit: the model is fitted to observations, the observations do not copy the model.
any	'model of the world' / 'world-model'	The model is of how the world shows up in the blanket, not of the world. The distinction is the content of NA-04.

Back-translation test — status: NOT RUN: no reader has been asked

Present a reader with the target-register terms for GENERATIVE_MODEL and GENERATIVE_PROCESS SIDE BY SIDE, and ask: (a) are these two different things? (b) which one is inside the agent and which is outside it? (c) if the two disagree, which one is wrong? (d) does either word suggest begetting, fatherhood, or fertility?

GENERATIVE_PROCESS — generative process

Store: TRUE · Domain: active-inference-math · Confidence: LOW for the coinage; HIGH only for the negative claim that no attested Sanskrit or Latin term names this object, which follows from the concept being a 20th-21st century distinction. · Reviewer: NONE - PENDING EXPERT REVIEW

The actual dynamics of the world that really produce observations, which exist independently of any agent's beliefs and need not share the generative model's form, variables, or factorization.

register	form	literal gloss	status
sanskrit	जनकव्यापार (janaka-vyāpāra)	the generating ACTIVITY -- what actually operates, as opposed to the image that is held	COINED
latin	processus generans	the process that is generating	COINED
english	generative process	The actual dynamics of the world that really produce observations; they exist independently of any agent's beliefs and need not share the generative model's form, variables, or factorization.	—
hindi	जनक प्रक्रिया ()	generative process	PENDING-CITATION
spanish	proceso generativo	generative process	PENDING-CITATION

Ambiguity notes

• ATTESTATION HONESTY, carried verbatim in kind from CONCEPTS.json: the string 'generative process' appears in 0 files of the corpus (grep, 2026-07-15). It is introduced by the lexicon backbone, not quoted from any chapter. Its CONTENT is already carried by NA-02 ('the whole construction is conditional on p(o,s) -- a choice, not a measurement'). This entry therefore renders a term the corpus itself does not use, and that is recorded rather than glossed over.
• THE OTHER HALF OF THE MOST-CONFUSED PAIR. Keeping model and process distinct is what makes 'a low F is not a correctness certificate' sayable at all: F is computed against the MODEL, while the PROCESS is what actually generated o. If a register merges them, the corpus loses the ability to state its own central honesty bound.
• The process 'need not share the model's form, variables, or factorization'. This is stronger than saying the model might be inaccurate: the world need not even be carved into the same THINGS the model posits. A register implying the process is a more accurate version of the model has understated the gap.
• The process is not a 'true model'. It is not a model at all -- nobody holds it. A register that renders it with any belief-, view-, or representation-root has re-imported the very confusion the pair exists to prevent.
• Sanskrit vyāpāra carries Mīmāṃsā and Nyāya freight (it is a term of art for operation/instrumentality in causal analysis) and Vedāntic usage besides. The freight is real; it is chosen here anyway because its ACTIVITY sense is exactly right and because it is maximally far from pratirūpa (image), which is the property that matters most.

Forbidden mistranslations

register	form	why it is wrong
any	ANY term also used for GENERATIVE_MODEL	Registered in CONCEPTS.json as a falsifier of the lexicon itself. This is the single hardest fence in this file.
sanskrit	prakṛti	Sāṃkhya primordial nature. It is seductive here because it genuinely names the world-side generator -- and that is exactly the danger: it would import Sāṃkhya's dualist metaphysics (prakṛti/puruṣa) as though the mathematics endorsed it. The mathematics endorses no metaphysics. CONCEPTS.json's rule is that no translation choice may quietly assert one.
any	'the true model' / 'the real model'	Two errors at once: it makes the process a kind of model (nobody holds it), and it recruits 'true', colliding with both TRUE_POSTERIOR's technical sense and the TRUE store.
any	'ground truth'	A machine-learning term for labelled data, which is an OBSERVATION, not the world's dynamics. It also smuggles in the idea that the process is accessible for comparison, which is precisely what makes the bound gap unobservable.

Back-translation test — status: NOT RUN: no reader has been asked

Show a reader the target-register term for GENERATIVE_PROCESS alone and ask: (a) is this something an agent holds, believes, or chooses? (b) does it exist if there is no agent? (c) must it be built out of the same variables the agent's model uses?

RECOGNITION_DENSITY — recognition density (approximate posterior q)

Store: TRUE · Domain: active-inference-math · Confidence: VERY LOW. Both Sanskrit components are unverified and the head is suspected of importing a Śaiva doctrine. This entry and EXPECTED_FREE_ENERGY are the two an expert should attack first. · Reviewer: NONE - PENDING EXPERT REVIEW

The distribution q(s) that is an agent's honest and improvable belief over hidden states, being the object optimized during perception and generally NOT equal to the true posterior.

register	form	literal gloss	status
sanskrit	प्रत्यभिज्ञासान्द्रता (pratyabhijñā-sāndratā)	recognition-density	COINED
latin	densitas agnoscens	recognising density	COINED
english	recognition density (approximate posterior q)	The distribution q(s) that is an agent's honest and improvable belief over hidden states; it is the object optimized during perception and is generally not equal to the true posterior.	—
hindi	अभिज्ञान घनत्व ()	recognition density	PENDING-CITATION
spanish	densidad de reconocimiento	recognition density	PENDING-CITATION

Ambiguity notes

• THE q-VS-p DISTINCTION IS LOAD-BEARING: q is what you HOLD and OPTIMIZE; p(s|o) is what you CANNOT COMPUTE. CONCEPTS.json states that a register rendering both as one word destroys the bound-gap concept. This entry and TRUE_POSTERIOR must therefore keep two clearly distinct terms in every register.
• SEVERE DOCTRINAL FREIGHT AND A PROBABLE DEFECT IN THE PROPOSED SANSKRIT: pratyabhijñā is the name of the Pratyabhijñā school of Kashmir Śaivism, where it means the recognition of one's own identity with Śiva. If that is what a Sanskrit reader hears, the coinage imports a specific non-dual metaphysics into a term for an approximate posterior -- which would be the exact failure CONCEPTS.json forbids. The author could not verify this in this pass and flags the coinage as probably unusable pending expert review.
• 'Recognition' is a historical accident of the variational-inference literature (the 'recognition model' of Helmholtz machines). It does not mean recognising a face, recognising an acquaintance, or formal recognition/acknowledgement. All three senses are live in all five registers.
• 'Density' is the probability-density sense, not physical density, thickness, or crowding. Classical Latin densitas and Hindi ghanatva are both physical-thickness words, so both coinages above are stretching a physical root to a measure-theoretic one -- which is what the European languages historically did too, but it is a stretch and is recorded as one.
• q is 'honest and improvable' in the definition. Note the sovereign hazard: 'honest' here is the ordinary English word, NOT the HONEST store. q is a TRUE-store object. The two senses must not be rendered alike, or the sentence 'q is an honest belief' would read as a claim about lived experience.

Forbidden mistranslations

register	form	why it is wrong
any	any term also used for TRUE_POSTERIOR	Destroys the bound-gap concept, which is the receipt for the INADMISSIBLE claim 'a low F means the model is correct'. Named in CONCEPTS.json.
sanskrit	pratyabhijñā used bare	Names a school of Kashmir Śaiva philosophy and its central soteriological doctrine. Imports a metaphysics; asserts a self that recognises itself.
sanskrit	any rendering on jñāna	MW s.v. jñāna (retrieved 2026-07-15) gives 'knowing... knowledge, (esp.) the higher knowledge (derived from meditation on the one Universal Spirit)'. The 'esp.' sense is explicitly the Vedāntic one. q is an approximate posterior that is usually wrong; rendering it with a word whose flagged primary sense is liberating knowledge of the Universal Spirit is the worst available direction of error.
any	'belief' rendered with a faith/trust root (śraddhā, fides, fe, viśvās)	q is a probability distribution. Faith roots assert a believing subject and a credence-without-evidence connotation; fides additionally carries loyalty and good-faith senses.
english	'the posterior'	Drops 'approximate', which is the entire content. q is generally NOT the posterior.

Back-translation test — status: NOT RUN: no reader has been asked

Show a reader the target-register terms for RECOGNITION_DENSITY and TRUE_POSTERIOR together and ask: (a) are these two different objects? (b) which one can the agent actually compute? (c) which one is generally wrong? (d) [Sanskrit only] does the term evoke a school of philosophy or a religious doctrine?

TRUE_POSTERIOR — true posterior

Store: TRUE · Domain: active-inference-math · Confidence: LOW. The modifier yathārtha is CITED but is suspected of being wrong precisely because its attested sense is correspondence-to-reality, which the concept forbids. A CITED headword is not thereby an admissible rendering -- this entry is the clearest case in the file of that distinction. · Reviewer: NONE - PENDING EXPERT REVIEW

The exact conditional p(s|o) implied by the generative model, generally intractable because it requires the marginal p(o) = SUM_s p(o,s), and which q approximates without necessarily ever equalling.

register	form	literal gloss	status
sanskrit	यथार्थपश्चात्सम्भाव्यता (yathārtha-paścāt-sambhāvyatā)	reality-conformable after-probability	COINED
latin	posterius verum (distributio posterior vera)	the true posterior (distribution)	COINED
english	true posterior	The exact conditional p(s|o) implied by the generative model, generally intractable because it requires the marginal p(o) = SUM_s p(o,s); q approximates it without necessarily ever equalling it.	—
hindi	यथार्थ पश्च प्रायिकता ()	true posterior probability	PENDING-CITATION
spanish	posterior verdadera (distribución a posteriori verdadera)	true posterior distribution	PENDING-CITATION

Ambiguity notes

• THE SOVEREIGN HOMONYM AND THE REASON THIS ENTRY IS DANGEROUS: 'true' here is a technical term of Bayesian inference meaning 'exact, as implied by the model'. It does NOT mean the TRUE store, and it does NOT mean a UNI ledger status. CONCEPTS.json states the homonym must not be rendered with the store's term. If a register uses one word for both, then 'a True signal can be wrong' and 'the true posterior is the exact one' start contaminating each other, and the sovereign vocabulary leaks.
• SECOND-ORDER TRAP, worse than the first: the true posterior is 'true' RELATIVE TO THE MODEL, not relative to the world. It is the exact conditional implied by p(o,s) -- and p(o,s) is a choice the agent made. So a confidently wrong model has a perfectly well-defined 'true' posterior that is true to nothing. A register whose word for 'true' means 'corresponding to reality' asserts something flatly false here.
• The Sanskrit yathārtha above is exposed to exactly that second trap: MW gives it as 'accordant with reality, conformable to truth', which is the CORRESPONDENCE sense. It is the best-attested candidate available and it is arguably wrong for that very reason. This conflict is recorded, not resolved.
• 'Posterior' is not the anatomical posterior, not 'later in time', and not 'subsequent'. It means 'conditioned on the observation'. The Latin and Spanish forms are especially exposed because their words genuinely mean 'later/behind' in ordinary usage -- Lewis & Short's entry for the correlate 'prior' confirms the classical sense is temporal/ordinal.
• 'Intractable' means computationally unavailable, not unknowable in principle and not forbidden. The obstruction is the marginal p(o) = SUM_s p(o,s), a sum that is generally too large to do.

Forbidden mistranslations

register	form	why it is wrong
any	rendering the technical 'true' with the term used for the TRUE store	Explicitly named in CONCEPTS.json. It would make 'a True signal can be wrong' unsayable, and would let a Bayesian technical adjective read as a UNI ledger status.
any	rendering 'true posterior' as 'the correct belief' or 'the right answer'	It is exact with respect to a chosen model, which may be wrong about the world. Correctness is precisely what it does not certify -- that is the content of BOUND_GAP.
any	any term also used for RECOGNITION_DENSITY	Destroys the bound gap. Named in CONCEPTS.json.
latin	'posterior' read in its classical temporal sense	Lewis & Short's correlate entry 'prior' confirms the classical sense is 'former, previous... freq. to be translated first' -- i.e. ordinal/temporal. The Bayesian sense is not classical and the word is being reassigned; readers of classical Latin will hear 'the later one'.
sanskrit	satya-based rendering	satya is truth/reality in the strongest metaphysical register (satya as Brahman, satya as the real). It would assert correspondence to reality, and additionally collide with the TRUE store's vocabulary.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and ask: (a) does 'true' here mean 'corresponds to how the world actually is'? (b) or does it mean 'exactly implied by the model we chose'? (c) can the agent compute this object? (d) is this the same word your register uses for a signal that is falsifiable and reproducible (the TRUE store)?

BOUND_GAP — bound gap

Store: TRUE · Domain: active-inference-math · Confidence: VERY LOW for the Sanskrit form specifically, which the author believes collides with sīmanta and flags for replacement. The ambiguity notes are the load-bearing part of this entry and do not depend on the coinage surviving. · Reviewer: NONE - PENDING EXPERT REVIEW

The non-negative slack D_KL[q(s)||p(s|o)] between variational free energy and surprisal, which is zero exactly when q equals the true posterior and is unobservable without the very posterior that was intractable to begin with.

register	form	literal gloss	status
sanskrit	सीमान्तर (sīmāntara)	bound-interval; the gap between the bound and what it bounds	COINED
latin	intervallum termini	the interval of the bound	COINED
english	bound gap	The non-negative slack D_KL[q(s)||p(s|o)] between variational free energy and surprisal; it is zero exactly when q equals the true posterior, and is unobservable without the very posterior that was intractable to begin with.	—
hindi	परिबंध अंतराल ()	bound gap	PENDING-CITATION
spanish	brecha de la cota	the gap of the bound	PENDING-CITATION

Ambiguity notes

• THIS ENTRY IS A RECEIPT, and that is its whole job: it is the receipt for the INADMISSIBLE claim 'a low F means the model is correct'. It explains why a CONFIDENTLY WRONG model can sit at low F. A register that renders it as a mere technical slack term, without the epistemic sting, has translated the algebra and lost the point.
• THE SELF-SWALLOWING PROPERTY, which is the hardest thing to carry into any register: the gap is UNOBSERVABLE without the very posterior that was intractable to begin with. You cannot measure how wrong you are, because measuring it requires the thing you could not compute. A register implying it could be evaluated, monitored, or bounded in practice asserts the opposite of the concept.
• 'Gap' is not a hole, a breach, a lack, a deficiency, or a failure. It is non-negative slack in an inequality. In particular it is not a defect to be fixed: it is zero only in the limiting case q = p(s|o).
• 'Bound' here is a mathematical upper bound, NOT a tie, a fetter, a limit-in-the-sense-of-restriction, or an obligation. Sanskrit bandha and Hindi 'bandh' both carry the fetter sense strongly (bandha is bondage in Yoga/Vedānta soteriology -- heavy doctrinal freight), and the Hindi rendering above is exposed to it.
• The gap IS a KL divergence, D_KL[q||p(s|o)], so this entry inherits every ambiguity note from KL_DIVERGENCE -- in particular its asymmetry. A register that renders it as a symmetric distance asserts a false mathematical property here too.

Forbidden mistranslations

register	form	why it is wrong
sanskrit	sīmānta / sīmanta	sīmanta is 'the parting of the hair' and the name of a saṃskāra (the sīmantonnayana rite). The proposed sīmāntara is uncomfortably close and a reader may hear it. This is a real defect in the proposed coinage and is recorded rather than hidden.
sanskrit	any rendering on bandha	bandha is bondage/fetter and in Yoga and Vedānta it is the soteriological opposite of mokṣa. It would import a liberation metaphysics into an inequality's slack term.
any	rendering 'gap' as a flaw, deficiency, or error	It is non-negative slack, not a mistake. Naming it an error suggests it should be driven to zero as a matter of hygiene; it cannot even be evaluated.
any	any rendering implying the gap can be measured or monitored	Inverts the concept. Its unobservability is the point and is what makes 'confidently wrong at low F' possible.
any	a symmetric 'distance' root	Inherited from KL_DIVERGENCE: D_KL[q||p] is not symmetric and is not a metric.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader with undergraduate probability the target-register term and gloss and ask: (a) can an agent evaluate this quantity for itself in practice? (b) is a large value of it a sign the agent knows it is wrong? (c) is 'bound' here an upper bound in an inequality, or a restriction/fetter?

ACCURACY — accuracy

Store: TRUE · Domain: active-inference-math · Confidence: LOW for the coinages. MEDIUM-HIGH for the negative finding, which is the valuable part of this entry: the attested Latin 'accuratio' is CITED, retrieved this pass, and is judged INADMISSIBLE on sense-mismatch. That judgement rests on quoted dictionary text plus the concept's own note, not on the author's taste. · Reviewer: NONE - PENDING EXPERT REVIEW

The term E_q(s)[ln p(o|s)] in F's complexity-minus-accuracy decomposition, measuring how well a belief explains the observations actually received.

register	form	literal gloss	status
sanskrit	उपपत्तिमात्रा (upapatti-mātrā)	explanation-measure; the measure of how well the belief accounts for what arrived	COINED
latin	mensura explicandi	the measure of explaining	COINED
english	accuracy	The term E_q(s)[ln p(o|s)] in F's complexity-minus-accuracy decomposition, measuring how well a belief explains the observations actually received.	—
hindi	व्याख्या-मात्रा ()	explanation-measure (NOT 'śuddhatā'/'yathārthatā', which would assert correctness -- see ambiguity_notes)	PENDING-CITATION
spanish	precisión (término de ajuste a los datos)	the fit-to-data term; RECORDED WITH A WARNING -- Spanish 'precisión' collides head-on with the separate concept PRECISION, and 'exactitud' would assert correctness	PENDING-CITATION

Ambiguity notes

• THE TRAP, AND IT IS SUBTLE AND SEVERE: this is the TECHNICAL sense, not the ordinary 'correctness' sense. CONCEPTS.json makes the sting explicit -- a CONFABULATED OBSERVATION RAISES APPARENT ACCURACY while corrupting the model, which is exactly what the complexity term prices. So 'accuracy' here measures fit to the observations you actually got, INCLUDING fabricated ones. It is not a virtue and it is not correctness.
• THIS IS WHY THE ATTESTED LATIN WORD WAS DECLINED. Lewis & Short s.v. 'accuratio' (retrieved 2026-07-15) gives 'accuracy, exactness, carefulness (very rare)' -- from accuro, 'to do with care'. That is the ordinary sense of careful workmanship: a virtue of the person doing the work. F's accuracy term is a property of a fit, and a careless liar can have a high one. A CITED headword that carries the WRONG SENSE is not a licence; it is a trap with a citation attached, and it is the most dangerous kind.
• Not 'accuracy' in the measurement-science sense either, where accuracy (closeness to the true value) is contrasted with precision (repeatability). That contrast is a THIRD distinct sense, and it collides with the separate concept PRECISION in this same domain. Three senses of this word pair are live at once.
• The decomposition is COMPLEXITY MINUS ACCURACY, and F is MINIMIZED -- so accuracy enters NEGATIVELY and minimizing F pushes accuracy UP. A register that renders accuracy as a cost or penalty inverts the sign. It is the term you want large.
• Accuracy is an expectation under q, E_q(s)[ln p(o|s)] -- it is averaged over the belief, not evaluated at a point estimate. A register implying a single best-guess state is being scored has dropped the expectation.

Forbidden mistranslations

register	form	why it is wrong
latin	accuratio	ATTESTED BUT WRONG, which is why it heads this list. Lewis & Short: 'accuracy, exactness, carefulness (very rare)', from accuro 'to do with care'. It names the carefulness of the person doing the work. F's accuracy names the fit of a belief to received data and is raised by a confabulated observation. Citing a real dictionary entry does not make a rendering admissible -- the sense must match too, and here it does not.
sanskrit	yathārthatā	MW s.v. yathārtha (retrieved 2026-07-15): 'accordant with reality, conformable to truth'. It asserts correspondence to REALITY. F's accuracy is correspondence to the OBSERVATIONS RECEIVED, which may be false. Using it would make the corpus's own warning -- that a confabulated observation raises apparent accuracy -- self-contradictory in Sanskrit.
sanskrit	śuddhatā / samyaktva	Purity/correctness roots, and samyak- carries heavy Buddhist freight (the samyak- terms of the Eightfold Path). Both assert rightness.
spanish	'precisión'	Collides head-on with the separately registered concept PRECISION in this same domain, which is inverse variance. Two different quantities would share one word in a single lexicon.
any	any rendering making accuracy a cost, penalty, or error term	Inverts the sign. F = complexity - accuracy, and F is minimized, so accuracy is pushed up.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and gloss and ask: (a) if I invent an observation that never happened and my belief fits my invention perfectly, does this quantity go up? (b) does this word describe a virtue of the person doing the modelling? (c) does it mean 'close to the true value'? (d) is a high value good or bad for the objective being minimized?

COMPLEXITY — complexity

Store: TRUE · Domain: active-inference-math · Confidence: MEDIUM for the Sanskrit, which is the highest in this file: gaurava is CITED from MW in this pass, MW's own lāghava entry independently corroborates the opposed pair, and -- on the corrected 2026-07-15 quote -- MW independently carries gaurava as 'heaviness in argumentation, cumbrousness, needless multiplication of causes', so the PARSIMONY SENSE is CITED rather than asserted. UPGRADED from the original text, which said that sense was PENDING-CITATION; that statement rested on an incomplete quote of MW and was wrong in the file's own favour-refusing direction (it under-claimed). What is still PENDING-CITATION, and is now actively SUSPECT, is the NAVYA-NYĀYA SCHOOL ATTRIBUTION: MW's loci are Sarvad. and KapS., neither of which is Navya-Nyāya. Do not read the CITED status as covering the school claim. Confidence is held at MEDIUM rather than raised, because the honorific senses remain live and no expert has read this entry. · Reviewer: NONE - PENDING EXPERT REVIEW

The term D_KL[q(s)||p(s)] in F, measuring how far a belief has moved off its prior and penalizing that movement so that minimizing F selects the simplest sufficient explanation rather than the best fit.

register	form	literal gloss	status
sanskrit	गौरव (gaurava)	weight, heaviness; 'heaviness in argumentation, cumbrousness, needless multiplication of causes' (MW, CITED) -- the cost of a needlessly heavy explanation. NOTE the school attribution has been REMOVED from this gloss: MW attests the argumentation sense but its loci are Sarvad. (Sarvadarśanasaṃgraha) and KapS. (Kapila-Sūtra = Sāṃkhya-sūtra) Sch., NOT a Navya-Nyāya text. See morphological_derivation.	CITED
latin	complexitas	complexity	COINED
english	complexity	The term D_KL[q(s)||p(s)] in F, measuring how far a belief has moved off its prior and penalizing that movement, so that minimizing F selects the simplest sufficient explanation rather than the best fit.	—
hindi	जटिलता ()	complexity (the ordinary sense: intricacy, tangledness -- which is NOT this quantity; see ambiguity_notes)	PENDING-CITATION
spanish	complejidad	complexity (ordinary sense: intricacy -- same mismatch as Hindi)	PENDING-CITATION

Ambiguity notes

• THE PRIMARY TRAP: this is NOT complexity in the ordinary sense of intricacy, complicatedness, tangledness, or having many parts. It is D_KL[q(s)||p(s)] -- how far the belief MOVED OFF ITS PRIOR. A belief can be intricate and cheap (if the prior was already intricate), or simple and expensive (if the prior was elsewhere). The Hindi and Spanish standing renderings both carry the intricacy sense and both therefore mis-describe the quantity; that is recorded above.
• SOVEREIGN HOMONYM, named in CONCEPTS.json: this is NOT the scale-ladder sense of 'complexity', which NA-10 records as NOT-MEASURED (no scalar orders rungs by it). A register must not render both with one term. If it does, a NOT-MEASURED quantity in the NATURA ledger and a computed term of F share a word, and a reader could take the second as evidence for the first -- which would be a nature/UNI lane crossing.
• The Sanskrit gaurava is the strongest attested candidate in this file, and its aptness is not an accident: MW records gaurava ('weight, heaviness') and lāghava ('brevity, conciseness') as an explicitly OPPOSED PAIR, and MW ALSO carries gaurava in the sense 'heaviness in argumentation, cumbrousness, needless multiplication of causes' -- i.e. the dictionary itself attests the word already doing duty as the name of an ARGUMENTATIVE FAULT, which is exactly the valence F's complexity term needs. (That sense was missing from this file's original quote and is restored 2026-07-15; see the citation.) The freight recorded below is real but is now WEAKER THAN IT LOOKED: gaurava's honorific senses ('importance, high value or estimation', 'respect shown to a person') do sit on the same word, so a Sanskrit reader may still hear a compliment where the mathematics intends a penalty -- but the pejorative technical sense is attested alongside them and a reader with the Sarvad./KapS. usage in view has the right valence available. The honorific risk is a genuine ambiguity, not a mistranslation, and the back-translation test below still tests for it.
• COMPLEXITY IS A COST AND ACCURACY IS A BENEFIT: F = complexity - accuracy, minimized. So this term is the one you want SMALL. A register that renders it with a merit/importance root (as gaurava's honorific senses would) inverts the valence.
• 'Simplest sufficient explanation' is the payoff, and it is why the term exists: without it, minimizing F would just be best-fit, and best-fit is how a model confabulates. The complexity term is what prices a raised apparent accuracy. Read with ACCURACY -- neither entry is intelligible alone.

Forbidden mistranslations

register	form	why it is wrong
latin	complexio	ATTESTED BUT WRONG. Lewis & Short (retrieved 2026-07-15) gives 'A combination or connection', 'comprehension or comprisal of a matter', a rhetorical period, 'the conclusion of a syllogism or a dilemma', and grammatical contraction. Not one of those senses is a quantity, let alone a divergence from a prior. A translator reaching for the obvious cognate lands on a syllogism's conclusion.
any	any term also used for the scale-ladder sense of complexity	CONCEPTS.json registers these as sovereign homonyms. The ladder sense is NOT-MEASURED in the NATURE-LEDGER; this sense is a computed term of F. Merging them would let a computed UNI-side quantity appear to discharge a NATURA-side absence.
sanskrit	gaurava read in its honorific senses	MW gives 'importance, high value or estimation', 'gravity, respectability, venerableness', 'respect shown to a person'. In F this term is a PENALTY. The honorific reading inverts the valence exactly.
any	any rendering as intricacy, complicatedness, or number-of-parts	Describes the wrong property. The quantity is distance from the prior, not internal structure.
any	'model complexity' in the parameter-counting sense (e.g. AIC/BIC penalties)	A related but distinct idea. D_KL[q||p] is a divergence between distributions, not a count of parameters, and it is belief-specific rather than model-structural.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and gloss and ask: (a) if my belief is very intricate but is exactly what I already believed before seeing any data, is this quantity large or small? (b) is a large value good or bad? (c) does this word suggest honour, importance, or respect?

KL_DIVERGENCE — Kullback-Leibler divergence

Store: TRUE · Domain: active-inference-math · Confidence: LOW-MEDIUM. Both Sanskrit morphemes are CITED, but the compound is a coinage and bheda's attested senses are a poor fit whose freight is fenced rather than removed. The strongest content here is the mathematical fence, which does not depend on any register. · Reviewer: NONE - PENDING EXPERT REVIEW

A non-negative measure of the difference between two distributions, zero if and only if they are equal (Gibbs' inequality), whose non-negativity is precisely what makes F upper-bound surprisal.

register	form	literal gloss	status
sanskrit	कुल्बाक्-लाइब्लर-भेदमात्रा (kulbāk-lāibler-bheda-mātrā)	Kullback-Leibler difference-measure	COINED
latin	divergentia Kullback-Leibleriana	Kullback-Leibler divergence	COINED
english	Kullback-Leibler divergence	A non-negative measure of the difference between two distributions, zero if and only if they are equal (Gibbs' inequality); its non-negativity is precisely what makes F upper-bound surprisal.	—
hindi	कुल्बैक-लाइब्लर अपसरण ()	Kullback-Leibler divergence	PENDING-CITATION
spanish	divergencia de Kullback-Leibler	Kullback-Leibler divergence	PENDING-CITATION

Ambiguity notes

• THE MATHEMATICAL FENCE, named in CONCEPTS.json: D_KL is ASYMMETRIC. D_KL[q||p] does not equal D_KL[p||q]. A register rendering it with a symmetric 'distance' root ASSERTS A FALSE MATHEMATICAL PROPERTY. This is not a nuance -- it is a wrong claim about mathematics, and under the twin admission tests it fails the math test outright, whatever the philology says.
• It is not a metric: besides asymmetry it violates the triangle inequality. Words like 'distance', 'gap', 'separation', 'antara', 'distantia', 'distancia', 'dūrī' all carry metric intuitions and all are exposed.
• 'Divergence' is a term of art here and is NOT the vector-calculus divergence (div F, the scalar field), NOT divergence in the sense of two things drifting apart over time, and NOT divergence in the sense of a series failing to converge. All three are live collisions in mathematical registers specifically -- i.e. the readers most likely to encounter this term are the ones most likely to hit the wrong sense.
• Sanskrit bheda's attested senses run to 'breaking, splitting, cleaving, rending, tearing, piercing' and to 'disunion, schism' and to philosophical DUALISM (bheda vs abheda is a central axis of Vedāntic dispute). A divergence between distributions tears nothing and takes no side on non-dualism. The freight is real and is fenced below.
• The non-negativity (Gibbs' inequality) is not incidental decoration: it is PRECISELY what makes F an upper bound on surprisal. A register that renders D_KL as though it could be signed, or as a mere 'difference', drops the property the whole construction rests on. Read with BOUND_GAP, which IS a KL divergence.

Forbidden mistranslations

register	form	why it is wrong
any	any symmetric 'distance' root (Latin distantia, Spanish distancia, Hindi dūrī, Sanskrit dūratā)	Asserts symmetry and metricity, both mathematically false. CONCEPTS.json names this explicitly: 'A register rendering it with a symmetric distance root asserts a false mathematical property.'
sanskrit	bheda used bare, in its dualism sense	MW gives bheda's philosophical sense as dualism; bheda/abheda is a live Vedāntic controversy. A KL divergence is not a metaphysical position.
sanskrit	bheda read as 'breaking, rending, schism'	MW's first glosses are violent ('breaking, splitting, cleaving, rending, tearing, piercing') and social ('disunion, schism'). Neither is a measure between distributions.
any	dropping the eponym and rendering only 'divergence' or 'difference-measure'	Under-specifies: there are many divergences (Rényi, Jensen-Shannon, f-divergences). The name picks one.
any	'relative entropy' used as if interchangeable in every context	It IS a standard synonym, but it re-imports the entropy word and thereby the thermodynamic trap fenced at ENTROPY. Acceptable only where that fence is separately held.

Back-translation test — status: NOT RUN: no reader has been asked

Give a mathematically literate reader the target-register term and gloss and ask: (a) is D(A,B) the same number as D(B,A)? (b) is this a distance? (c) can it be negative? (d) does it refer to the divergence operator of vector calculus?

ENTROPY — entropy

Store: TRUE · Domain: active-inference-math · Confidence: LOW. The coinage is weak and self-flagged for a doctrinal collision. The negative claim -- that no classical Sanskrit or Latin term names Shannon entropy, a 1948 concept -- is HIGH confidence and follows from the date. · Reviewer: NONE - PENDING EXPERT REVIEW

The expected surprisal H of a distribution, measuring the average uncertainty of outcomes drawn from it, in nats.

register	form	literal gloss	status
sanskrit	माध्यमिकाप्रत्याशितमात्रा (mādhyamika-apratyāśita-mātrā)	average unexpectedness-measure	COINED
latin	entropia	entropy	COINED
english	entropy	The expected surprisal H of a distribution, measuring the average uncertainty of outcomes drawn from it, in nats.	—
hindi	एन्ट्रॉपी ()	entropy (loanword)	PENDING-CITATION
spanish	entropía	entropy	PENDING-CITATION

Ambiguity notes

• THE DEFINING TRAP, named in CONCEPTS.json: this is SHANNON entropy, NOT thermodynamic entropy. The shared English word is a loan, and a register that reaches for a heat/disorder root imports a false physical claim. This is the same class of error as rendering 'free energy' thermodynamically, and the two traps compound: a reader who takes both loans literally will conclude the free-energy principle is a statement of thermodynamics. It is not.
• 'Entropy as disorder' is a further, popular-science mistranslation that is wrong even WITHIN thermodynamics and is catastrophic here. H is the average uncertainty of outcomes under a distribution. A uniform distribution over two tidy outcomes has higher entropy than a peaked distribution over a thousand messy ones. Disorder is not the variable.
• Entropy is EXPECTED SURPRISAL: it is the average of SURPRISAL over the distribution. Surprisal attaches to ONE observation; entropy attaches to a DISTRIBUTION. CONCEPTS.json marks these distinct and no register may merge them. The Sanskrit coinage above is deliberately built out of the surprisal coinage to make the relation visible while keeping the terms separate.
• Entropy is not information, and not 'amount of information' in the loose sense. It is also not negentropy, not information CONTENT of a message, and not compressibility -- though all are relatives. It is a property of a distribution, not of a string or an object.
• In this domain entropy appears chiefly INSIDE other quantities -- most importantly as AMBIGUITY, which is the expected entropy of the likelihood, E_q(s|pi)[H[p(o|s)]]. A register's entropy word must survive being nested inside that expression.

Forbidden mistranslations

register	form	why it is wrong
any	any heat, disorder, decay, or chaos root	Imports a false physical claim. CONCEPTS.json: 'Shannon entropy, NOT thermodynamic entropy; the shared English word is a loan and a register that reaches for a heat/disorder root imports a false physical claim.'
hindi	any rendering on ūṣmā (heat)	Thermodynamic. Same false claim, and it would additionally make the free-energy loan look consistent, compounding two errors into an apparent physics.
sanskrit	any rendering on tamas	tamas is the Sāṃkhya guṇa of darkness/inertia/disorder. It is doctrinal freight AND the disorder mistranslation in one word -- the worst available option.
sanskrit	mādhyamika read as the Madhyamaka school	Nāgārjuna's school. The author's own proposed coinage is exposed to this and is flagged for probable replacement.
any	rendering entropy and surprisal alike	Entropy is expected surprisal. One is a property of a distribution, the other of a single observation.

Back-translation test — status: NOT RUN: no reader has been asked

Give a physicist and a Sanskrit/target-register reader (separately) the term and gloss and ask: (a) is this a quantity of thermodynamics? (b) does it measure disorder or messiness? (c) does it attach to one outcome or to a whole distribution? (d) [Sanskrit] does it name a school of philosophy?

EVIDENCE — evidence (marginal likelihood)

Store: TRUE · Domain: active-inference-math · Confidence: LOW for the coinages; HIGH for the two refutations, which are the substance of this entry: 'evidentia' and 'pramāṇa' are both CITED from their authorities this pass and both are judged INADMISSIBLE on quoted sense-mismatch. The forbidden list here is better evidenced than the proposals. · Reviewer: NONE - PENDING EXPERT REVIEW

The marginal p(o) = SUM_s p(o,s) of observations under a model, whose negative logarithm is surprisal and whose logarithm F lower-bounds in negated form.

register	form	literal gloss	status
sanskrit	प्रान्तिकसम्भाव्यता (prāntika-sambhāvyatā)	marginal probability	COINED
latin	probabilitas marginalis	marginal probability	COINED
english	evidence (marginal likelihood)	The marginal p(o) = SUM_s p(o,s) of observations under a model; its negative logarithm is surprisal, and F lower-bounds its logarithm in negated form.	—
hindi	उपांत प्रायिकता ()	marginal probability (deliberately NOT 'sākṣya'/'pramāṇ', which mean evidence-as-proof)	PENDING-CITATION
spanish	verosimilitud marginal	marginal likelihood (deliberately NOT 'evidencia')	PENDING-CITATION

Ambiguity notes

• SOVEREIGN HOMONYM, AND CONCEPTS.json CALLS IT A SERIOUS ONE: this is NOT 'evidence' in the method sense (EVIDENCE_CLASS, receipts). One is a NORMALIZING CONSTANT OF A MODEL; the other is WHAT A CLAIM IS KNOWN BY. Rendering them alike would let a model's marginal read as a UNI gate -- i.e. it would let an arithmetic quantity look like it discharges an evidence requirement. That is the failure this whole corpus is built to prevent, appearing inside a single word.
• THE ENGLISH IS ALREADY BAD AND THE OTHER REGISTERS INHERIT IT: Bayesian statistics chose 'evidence' for p(o), and that choice is the source of the collision. The English register cannot fix this (the term is standard), so the parenthetical '(marginal likelihood)' is load-bearing and must not be dropped. Where a register CAN dodge -- as the Spanish 'verosimilitud marginal' does -- it should.
• p(o) is a property OF A MODEL, not of the world and not of a claim. Two models assign different evidence to the same observation. A register implying evidence is something the observation HAS, independently, has asserted the opposite.
• 'Marginal' means summed/integrated over the hidden states, SUM_s p(o,s). It does not mean slight, borderline, peripheral, minor, or barely-adequate -- all of which are live in every register. Spanish 'marginal' and Hindi 'upānt' both carry the peripheral sense.
• High evidence is not a certificate of a good model: p(o) can be raised by a model that is flexible enough to expect anything. Read with BOUND_GAP and with ACCURACY -- the same anti-confabulation logic applies.

Forbidden mistranslations

register	form	why it is wrong
latin	evidentia	ATTESTED, AND WRONG IN BOTH ITS SENSES. Lewis & Short (retrieved 2026-07-15): the classical sense is 'clearness, distinctness', Cicero's rendering of ἐνάργεια -- a quality of style, not a quantity. Its only 'proof' sense is Vulgate (biblical, late). So a translator reaching for the cognate gets either rhetorical perspicuity or scriptural proof, and p(o) is neither. Doubly disqualified.
sanskrit	pramāṇa	THE MOST TEMPTING AND MOST DANGEROUS OPTION. MW s.v. pramāṇa (retrieved 2026-07-15): 'measure, scale, standard' and a MEANS OF ACQUIRING CERTAIN KNOWLEDGE (the six pramāṇas of Vedānta). It is an epistemic INSTRUMENT and a warrant. p(o) is a normalizing constant. Using pramāṇa would make a model's arithmetic read as a means of valid knowledge -- exactly the homonym CONCEPTS.json says must not be rendered alike, and it would import Nyāya-Mīmāṃsā epistemology besides.
sanskrit	pratyakṣa or sākṣya	MW s.v. pratyakṣa (retrieved 2026-07-15) gives n. 'ocular evidence, direct perception, apprehension by the senses (in Nyāya one of the 4 Pramāṇas or modes of proof)'. That is evidence-as-warrant and perception besides. p(o) is neither perceived nor a warrant.
any	any term also used for EVIDENCE_CLASS (method domain)	Named in CONCEPTS.json as a sovereign homonym. It would let a marginal likelihood read as a receipt.
any	reading 'marginal' as slight, minor, or borderline	It means summed over hidden states. The peripheral sense is live in Spanish, Hindi and English alike.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader BOTH the target-register term for this concept AND the target-register term for EVIDENCE_CLASS (from the method domain) and ask: (a) are these the same word? (b) is the thing named here something that could justify a claim in an argument? (c) is it a property of the observation, or of a model? (d) does 'marginal' here suggest 'barely enough'?

NAT — nat

Store: TRUE · Domain: active-inference-math · Confidence: MEDIUM. The transliteration route is low-risk and conventional, and the mātrā gloss rests on a CITED MW entry whose attested prosodic 'mora' sense independently demonstrates the exact register required (a named unit in a formal system). This is among the safest entries in the file. · Reviewer: NONE - PENDING EXPERT REVIEW

The unit of information under natural logarithms, equal to log2(e) or about 1.442695 bits, in which F, G, entropy and surprisal are all measured.

register	form	literal gloss	status
sanskrit	नैट् (प्राकृतिकमात्रा) (naiṭ (prākṛtika-mātrā))	nat; 'natural-logarithm unit-of-measure'	COINED
latin	nat (unitas naturalis)	nat; natural unit	COINED
english	nat	The unit of information under natural logarithms, equal to log2(e) or approximately 1.442695 bits; F, G, entropy and surprisal are all measured in nats.	—
hindi	नैट ()	nat (unit of information under natural logarithms)	PENDING-CITATION
spanish	nat	nat (unidad de información en logaritmos naturales)	PENDING-CITATION

Ambiguity notes

• LOAD-BEARING UNDER THE HONESTY RAIL, and this is why the entry matters more than its size suggests: CONCEPTS.json states that an F or G REPORTED WITHOUT ITS LOG BASE IS NOT A NUMBER (NA-02). Every value carries value + units + scope. So the nat is not a footnote to the math -- it is the thing that makes an F value admissible at all. A register that treats the unit as optional decoration has broken the rail, not a convention.
• A nat is NOT a bit. 1 nat = log2(e) ~ 1.442695 bits. Reporting F in nats and reading it as bits inflates every quantity by ~44%. The conversion is exact and the confusion is common.
• 'Nat' is a clipping of 'natural' (natural logarithm). It has nothing to do with nature in the NATURA sense, nothing to do with natural numbers, and nothing to do with anything being 'natural' in the ordinary sense. The Sanskrit and Latin glosses above use nature-roots (prākṛtika, naturalis) to explain the etymology and thereby walk straight into that collision -- flagged below.
• SOVEREIGN HAZARD SPECIFIC TO THIS FILE: rendering 'nat' via prākṛtika would put the Sāṃkhya word prakṛti inside a unit name. It is only an etymological gloss, but a Sanskrit reader hearing 'prakṛti-measure' may reasonably ask which metaphysics the unit belongs to. Transliteration ('naiṭ') is therefore the primary proposal and the gloss is secondary.
• Unit names are among the FEW items in this whole lexicon where transliteration is not a failure but the correct answer. A translated unit name is usually a defect: it invites the reader to think the quantity is different in their language.

Forbidden mistranslations

register	form	why it is wrong
any	translating 'nat' as though it meant 'natural thing' or 'unit of nature'	It is a clipping of 'natural logarithm'. Any nature-semantics is an etymological accident and, in this corpus specifically, collides with the sovereign NATURA vocabulary.
sanskrit	prākṛtika-mātrā used as the primary term	Puts prakṛti (Sāṃkhya primordial nature) inside a unit of information. Offered above only as an etymological gloss; the transliteration is the primary form.
any	'bit'	A different unit. 1 nat = log2(e) ~ 1.442695 bits. Substituting one for the other silently rescales every F, G, H and surprisal value by ~1.44.
any	omitting the unit when reporting F or G	Not a translation error but the error the entry exists to prevent: per NA-02, a free-energy value without its log base is not a number.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and ask: (a) is this a unit of measurement? (b) does it have anything to do with nature, the natural world, or natural numbers? (c) is it the same as a bit? (d) if I tell you F = 3 in this unit, do you know what F is -- or do you need something else?

RISK — risk

Store: TRUE · Domain: active-inference-math · Confidence: LOW for the coinages; HIGH for the refutations of periculum and saṃśaya, both CITED from their authorities in this pass with the disqualifying senses quoted. The saṃśaya finding is the sharper one: it would have collapsed both terms of G into one word. · Reviewer: NONE - PENDING EXPERT REVIEW

The divergence D_KL[q(o|pi)||p(o|C)] between a policy's predicted outcomes and its preferred outcomes, forming one of the two terms of expected free energy.

register	form	literal gloss	status
sanskrit	इष्टविचलनमात्रा (iṣṭa-vicalana-mātrā)	measure of deviation from the desired	COINED
latin	divergentia ab optatis	divergence from the things wished for	COINED
english	risk	The divergence D_KL[q(o|pi)||p(o|C)] between a policy's predicted outcomes and its preferred outcomes; one of the two terms of expected free energy.	—
hindi	अभीष्ट-विचलन ()	deviation from the desired (deliberately NOT 'jokhim', which means hazard)	PENDING-CITATION
spanish	divergencia respecto a lo preferido	divergence from the preferred (deliberately NOT 'riesgo')	PENDING-CITATION

Ambiguity notes

• THE TRAP IS THE WORD ITSELF, and CONCEPTS.json states it flatly: technical sense ONLY. This is NOT danger, NOT hazard, NOT the ordinary sense of risk, and NOT operational risk in an ops or SLO context. A register reaching for a peril root ASSERTS A THREAT THAT THE MATH DOES NOT CONTAIN. Risk here is a KL divergence between two distributions over outcomes. Nothing is at stake in it; nothing can go wrong in it.
• THIS COLLISION IS ESPECIALLY HOT IN THIS CORPUS, which also carries an ops/SLO vocabulary, an approval plane, and a do-no-harm gate. If 'risk' in the EFE sense and 'risk' in the operational sense share a word, a sentence like 'the lowest-risk policy' becomes ambiguous between 'the one whose predicted outcomes best match preferences' and 'the one least likely to break production'. Those can be opposite policies.
• Risk is one of the TWO terms of G, the other being AMBIGUITY: G = risk + ambiguity. It is also equivalently the negative of pragmatic value. A register must let risk sit inside that sum without the sum reading as 'danger plus vagueness'.
• Risk is a divergence, so it inherits KL_DIVERGENCE's asymmetry fence: D_KL[q(o|pi)||p(o|C)] is not symmetric, and it is not a distance between predicted and preferred outcomes in any metric sense.
• The preferences p(o|C) are a PRIOR over preferred observations, not a reward and not a utility. So risk is not expected loss, not regret, and not variance of return -- all of which are what 'risk' means in finance and decision theory, and any of which a quantitatively literate reader may assume. The finance sense (risk as variance) is arguably the most dangerous false friend here because it is also mathematical and therefore looks right.

Forbidden mistranslations

register	form	why it is wrong
latin	periculum	ATTESTED AND WRONG. Lewis & Short (retrieved 2026-07-15): II.B 'Risk, hazard, danger, peril', marked 'the common signif. of the word'. The math contains no danger. (Its literal sense I, 'a trial, experiment, attempt, proof', is also wrong but interestingly so -- it is closer to the method domain's vocabulary than to this quantity.)
sanskrit	saṃśaya	MW s.v. saṃśaya (retrieved 2026-07-15) gives 'uncertainty, irresolution, hesitation, doubt' AND 'difficulty, danger, RISK of or in or to'. It looks like a direct hit and is wrong twice: its danger sense imports the threat the math lacks, and its doubt sense belongs to AMBIGUITY, the OTHER term of G. One word for both terms of a two-term sum collapses the sum by construction.
sanskrit	bhaya, āpad, or any peril/fear root	bhaya is fear -- an affect, asserting an experiencer. āpad is calamity. Neither is a divergence.
hindi	jokhim (जोखिम)	Hazard/peril in ordinary Hindi. Same false threat-claim.
spanish	riesgo	Peril in ordinary Spanish, and additionally the finance sense (riesgo as variance of return). Both wrong; the second wrong in a way that looks mathematical.
any	rendering risk as expected loss, regret, or variance	The decision-theory and finance senses. Risk here is D_KL[q(o|pi)||p(o|C)] -- a divergence from a preference prior, not a moment of a return distribution.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and gloss and ask: (a) does this word imply something bad might happen? (b) does it imply anyone could be harmed? (c) is it about not knowing where you are, or about outcomes not matching what was wanted? (d) is it the variance of a return?

AMBIGUITY — ambiguity

Store: TRUE · Domain: active-inference-math · Confidence: LOW for the coinages; HIGH for the refutation of ambiguitas, which is CITED with its disqualifying sense quoted. Note the reflexive hazard recorded above: this file's own schema uses 'ambiguity' in the forbidden sense, and that could not be fixed from inside this entry. · Reviewer: NONE - PENDING EXPERT REVIEW

The expected entropy of the likelihood, E_q(s|pi)[H[p(o|s)]], measuring how uninformative observations will be even if the policy reaches the states it intends.

register	form	literal gloss	status
sanskrit	अवबोधाक्षमता (avabodhākṣamatā)	informing-incapacity; the degree to which observations will fail to tell you which state you are in	COINED
latin	inopia indicii	poverty of indication; the want of anything the observation could tell you	COINED
english	ambiguity	The expected entropy of the likelihood, E_q(s|pi)[H[p(o|s)]], measuring how uninformative observations will be even if the policy reaches the states it intends.	—
hindi	अनिश्चायकता ()	non-determinativeness: the property of observations that they will not settle which state you are in (deliberately NOT 'dvi-arthaktā' = double meaning)	PENDING-CITATION
spanish	ambigüedad (entropía esperada de la verosimilitud)	ambiguity in the technical sense; the parenthetical is load-bearing because Spanish 'ambigüedad' otherwise reads as double meaning	PENDING-CITATION

Ambiguity notes

• A LIVE TRAP FOR THIS VERY FILE, and CONCEPTS.json says so explicitly: this is NOT linguistic or semantic ambiguity -- and THIS LEXICON'S OWN ambiguity_notes FIELDS USE THE ORDINARY SENSE. So the word 'ambiguity' appears in this file with two incompatible meanings: the field name means 'senses a word wrongly carries', and this concept means E_q(s|pi)[H[p(o|s)]]. That collision is inside the schema itself and cannot be renamed away here; it is recorded so no reader mistakes one for the other.
• THE DEFINING PROPERTY: high ambiguity means YOU ACT AND STILL DO NOT KNOW WHERE YOU ARE. It is not that the words are unclear -- it is that the WORLD's mapping from states to observations is uninformative. You can reach exactly the state you intended and learn nothing by arriving.
• It is a property of the LIKELIHOOD p(o|s) -- of the A tensor, of the sensor -- and NOT a mental state. Doubt, hesitation, indecision, and confusion are all states of a mind; ambiguity is a fact about a channel. A blurry camera has high ambiguity and no doubts. Every doubt-root in every register fails on exactly this point, which is why they are all forbidden below.
• Ambiguity is one of the TWO terms of G, with RISK the other: G = risk + ambiguity. It is also (up to sign) the epistemic term. Minimizing G therefore includes REDUCING ambiguity, which is what makes an agent probe, disambiguate, and look before it leaps. A register that renders it as mere vagueness will not carry that.
• 'Expected' here is again the mathematical expectation operator, taken under q(s|pi) -- so all the traps at EXPECTED_FREE_ENERGY apply inside this term too. And 'entropy' here is Shannon entropy, so ENTROPY's heat/disorder fence applies as well. This one short concept inherits two other entries' fences.

Forbidden mistranslations

register	form	why it is wrong
latin	ambiguitas	ATTESTED AND WRONG -- the cognate is the trap. Lewis & Short (retrieved 2026-07-15): 'equivocalness, double sense, ambiguity, uncertainty', with rhetorical and legal citations. That is words having two meanings. E_q(s|pi)[H[p(o|s)]] is a property of a sensory channel. The Latin cognate delivers precisely the wrong sense with full classical authority behind it.
sanskrit	saṃśaya	MW (retrieved 2026-07-15): 'uncertainty, irresolution, hesitation, doubt in or of'; '(in Nyāya) doubt about the point to be discussed (one of the 16 categories)'. It is a MENTAL STATE and a Nyāya category of dialectic. Ambiguity is a property of a channel, not a hesitation in a mind. It additionally carries 'danger, risk', which collides with RISK, the other term of G.
sanskrit	dvi-arthatā or any 'two-meanings' compound; also śleṣa	The linguistic sense exactly; śleṣa is moreover a figure of Sanskrit poetics (deliberate double meaning), which is the ordinary-sense trap plus aesthetic-theory freight.
any	any doubt, hesitation, indecision, or confusion root	All name mental states. Ambiguity is a property of p(o|s) and is well-defined for a sensor with no mind attached.
any	any vagueness or unclarity root	Vagueness is a property of a description; ambiguity here is expected entropy of a likelihood. Also collides with Latin 'evidentia' ('clearness'), fenced at EVIDENCE.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and gloss and ask: (a) is this about a word or statement having more than one meaning? (b) is it a feeling of doubt in someone's mind? (c) could a camera with a dirty lens have a lot of it, with nobody present? (d) is it about not reaching your goal, or about not learning where you are?

EPISTEMIC_VALUE — epistemic value

Store: TRUE · Domain: active-inference-math · Confidence: LOW. jñāna is CITED but its MW-flagged primary sense makes it a liability rather than an asset, which is why the Hindi register deliberately routes around it via sūcanā. Another case where a CITED headword is not an admissible rendering. · Reviewer: NONE - PENDING EXPERT REVIEW

The expected information gain of a policy, being the expected divergence between the belief that would be held after observing and the belief held now, which minimizing G maximizes.

register	form	literal gloss	status
sanskrit	ज्ञानलाभप्रत्याशा (jñāna-lābha-pratyāśā)	expected knowledge-gain	COINED
latin	lucrum cognoscendi exspectatum	the expected gain of coming-to-know	COINED
english	epistemic value	The expected information gain of a policy: the expected divergence between the belief that would be held after observing and the belief held now. Minimizing G maximizes it.	—
hindi	प्रत्याशित सूचना-लाभ ()	expected information gain	PENDING-CITATION
spanish	valor epistémico (ganancia de información esperada)	epistemic value; the parenthetical carries the actual content	PENDING-CITATION

Ambiguity notes

• ITS READING AS INFORMATION GAIN IS ITSELF APPROXIMATE, and CONCEPTS.json is careful about this in a way most of the literature is not: the epistemic term additionally treats q(s|o,pi) as standing in for p(s|o), so WHERE q IS A POOR POSTERIOR THE READING DEGRADES (NA-02). So 'expected information gain' is what the term is usually called and what it approximately is -- not what it exactly is. A register that renders it as simply 'information gain' has quietly upgraded an approximation into an identity.
• 'Value' does NOT mean worth, merit, price, or moral value. It is a term in a sum that gets minimized. Every value-root in every register carries an evaluative connotation that the mathematics does not have -- and the danger is specific: 'epistemic value' sounds like a claim that knowledge is GOOD, which is a philosophical position, not a term of G.
• SIGN TRAP, and it is genuinely confusing even in English: G = risk + ambiguity, and G is MINIMIZED; but epistemic value and pragmatic value enter G NEGATIVELY (G = -(epistemic + pragmatic)). So minimizing G MAXIMIZES epistemic value. A register must not let 'minimize' and 'value' collide into 'less knowledge is better'.
• 'Epistemic' is not 'epistemological'. The former concerns knowledge itself; the latter concerns the theory of knowledge. A register that renders it via a philosophy-of-knowledge root has named a discipline instead of a quantity.
• DROPPING THIS TERM YIELDS THE CONFIDENT WRONG ACTOR -- an agent that pursues its goal and never probes, never disambiguates, and is therefore certain and mistaken. That is the practical content and any register must preserve it. Its partner failure is at PRAGMATIC_VALUE: drop THAT and you get the explorer who arrives nowhere.

Forbidden mistranslations

register	form	why it is wrong
sanskrit	jñāna used bare, or any jñāna compound read in MW's 'esp.' sense	MW flags the salient sense as 'the higher knowledge (derived from meditation on the one Universal Spirit)'. That is Vedāntic liberating knowledge. Epistemic value is an expected KL divergence, computed by a thermostat-like system with no spirit in sight. The direction of error is the flattering one -- which the honesty rail says is the direction to distrust most.
sanskrit	prajñā or vijñāna	prajñā is wisdom with heavy Buddhist freight (the pāramitā); vijñāna is discriminative consciousness, a skandha in Buddhism and a tattva elsewhere. Both import a metaphysics and both assert a knower.
any	any rendering implying knowledge is good, valuable, or meritorious	Asserts a philosophical position. The term is an addend, not an endorsement.
any	'curiosity' or any drive/motivation root	Asserts an experiencer with a desire to know. The quantity is defined without one. This is a popular gloss in the literature ('curiosity-driven exploration') and it is a lane-crossing here.
any	'information gain' presented as an exact identity	It is an approximation that degrades when q is a poor posterior (NA-02). Rendering it as an identity drops a stated honesty bound.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and gloss and ask: (a) does this say that knowing things is good or worthwhile? (b) does it presuppose someone who wants to know? (c) is a policy with more of this preferred, and does that mean G goes up or down? (d) [Sanskrit] does the term suggest spiritual or liberating knowledge?

PRAGMATIC_VALUE — pragmatic value

Store: TRUE · Domain: active-inference-math · Confidence: VERY LOW. No component of either the Sanskrit or the Latin was verified in this pass, and the proposed Hindi is self-flagged for yogic freight. The ambiguity notes are the only part of this entry the author would defend. · Reviewer: NONE - PENDING EXPERT REVIEW

The expected log-preference E_q(o|pi)[ln p(o|C)] of a policy, measuring how far its predicted outcomes move the agent toward its preferred observations.

register	form	literal gloss	status
sanskrit	प्रयोजनप्रत्याशा (prayojana-pratyāśā)	expected purpose-attainment	COINED
latin	profectus ad optata exspectatus	expected advance toward the things wished for	COINED
english	pragmatic value	The expected log-preference E_q(o|pi)[ln p(o|C)] of a policy, measuring how far its predicted outcomes move the agent toward its preferred observations.	—
hindi	प्रत्याशित प्रयोजन-सिद्धि ()	expected attainment of the aim	PENDING-CITATION
spanish	valor pragmático	pragmatic value	PENDING-CITATION

Ambiguity notes

• NOT 'PRAGMATIC' IN THE ORDINARY OR PHILOSOPHICAL-SCHOOL SENSE, and CONCEPTS.json names both. Ordinary 'pragmatic' means practical, hard-headed, unsentimental. Philosophical 'Pragmatism' is Peirce, James and Dewey -- a whole theory of truth. This is neither: it is E_q(o|pi)[ln p(o|C)], an expected log-preference. Spanish 'pragmático' and the cognates in every European register carry the school sense and are exposed.
• 'Value' again does not mean worth or merit -- see EPISTEMIC_VALUE. And the same SIGN TRAP applies: pragmatic value enters G negatively, so minimizing G maximizes it.
• The preference p(o|C) is a PRIOR OVER PREFERRED OBSERVATIONS, not a reward, not a utility, and not a goal-state. So pragmatic value is not expected reward and not expected utility -- and a reader with a reinforcement-learning background will assume it is exactly that. This is the sharpest false friend of the entry and it is a technically-literate reader's error, not a naive one.
• DROPPING THIS TERM YIELDS THE EXPLORER WHO RESOLVES UNCERTAINTY FOREVER AND ARRIVES NOWHERE. That is the practical content, and it is the mirror of EPISTEMIC_VALUE's failure mode. Neither entry is intelligible without the other: EFE holds both in one quantity ON PURPOSE, and goal-only and explore-only are both malformed.
• 'Log-preference' means the LOGARITHM of the preference prior. A register that drops the log renders a different quantity. The log is not decoration: it is what makes the term additive with the others and what puts it in nats.

Forbidden mistranslations

register	form	why it is wrong
any	any rendering via Pragmatism the philosophical school	Peirce/James/Dewey. It would assert a theory of truth as a term in a sum. The cognate delivers this in most European registers automatically.
any	'expected reward' / 'expected utility'	The preference is a PRIOR OVER OBSERVATIONS, not a reward function or a utility. This is the whole point of encoding goals as distributions rather than rewards -- see PREFERENCE. An RL-literate reader will make this error by default.
any	any rendering implying practicality, hard-headedness, or common sense	The ordinary sense of 'pragmatic'. Describes a temperament, not a quantity.
sanskrit	artha used bare	artha is object/purpose/meaning/wealth and is one of the four puruṣārthas (the aims of life). It imports a whole ethical framework and is besides ambiguous between purpose and meaning.
hindi	any rendering on siddhi	Yogic attainment/supernatural power. The author's own proposed Hindi is exposed to this and it is flagged above.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and gloss and ask: (a) does this refer to the philosophical school of Pragmatism? (b) does it mean being practical or realistic? (c) is this the same as expected reward? (d) if a policy has more of this, does G go up or down?

POLICY — policy

Store: TRUE · Domain: active-inference-math · Confidence: LOW for the coinages; HIGH for the refutations of naya and consilium, both CITED this pass with their disqualifying senses quoted. naya is the single best example in this file of an attested, plausible, authoritative rendering that must nevertheless be refused. · Reviewer: NONE - PENDING EXPERT REVIEW

A candidate action sequence pi over which expected free energy is scored and among which one is selected.

register	form	literal gloss	status
sanskrit	कर्मक्रम (karma-krama)	action-sequence -- exactly what the definition says, and nothing more	COINED
latin	series actionum	a sequence of actions	COINED
english	policy	A candidate action sequence pi over which expected free energy is scored and among which one is selected.	—
hindi	क्रिया-अनुक्रम ()	action sequence (deliberately NOT 'nīti', which means policy in the governance/ethics sense)	PENDING-CITATION
spanish	secuencia de acciones (política, en el sentido técnico)	action sequence; 'política' is the standing calque and is recorded with a warning	PENDING-CITATION

Ambiguity notes

• A SEVERE TRAP IN A CORPUS THAT ALSO CARRIES GOVERNANCE RULES, and CONCEPTS.json says exactly that: this is NOT policy in the governance, institutional, or document sense. This corpus contains an approval plane, a sign-off gate, doctrine documents, and operating rails -- all of which are 'policy' in the ordinary sense. If one word serves both, 'the policy was selected' becomes ambiguous between 'the agent picked an action sequence' and 'the institution adopted a rule'.
• IT IS A SEQUENCE OF ACTIONS, NOTHING MORE. Not a strategy, not a plan in the sense of an intention, not a rule, not a decision procedure, not a disposition. In particular it is NOT the reinforcement-learning sense of policy (a state-to-action MAPPING, pi(a|s)) -- which is a different object, and which an RL-literate reader will assume by default. In active inference pi indexes a candidate SEQUENCE; in RL pi is a function from states to actions. This false friend is technical, is extremely common, and is not flagged in most of the literature.
• A policy is a CANDIDATE. It is one of the things being scored, and most of them are not chosen. A register whose word implies adoption, commitment, or authority has described the outcome of selection rather than the object of it.
• Policies are not deliberated. G is computed over them and the lowest is selected; no one considers, weighs, or debates. The Latin consilium is disqualified on precisely this point despite being the most natural classical word.
• The corpus's own rail states that G(pi) is FRAMING VOCABULARY, NOT A SCHEDULER -- scheduling goes by PENDING-burndown, inadmissible-event catch, MigrationGate and read-agency. So in the UNI program a 'policy' in this sense is not a thing that gets run. Any register implying UNI selects policies by G is drift.

Forbidden mistranslations

register	form	why it is wrong
sanskrit	naya	ATTESTED, PLAUSIBLE, AND WRONG. MW s.v. naya (retrieved 2026-07-15) gives m. (√1. nī) 'leading (of an army)'; 'conduct, behaviour, (esp.) prudent conduct, good management, POLITY, CIVIL AND MILITARY GOVERNMENT'; 'wisdom, prudence, reason'; 'plan, design'; 'leading thought, maxim, principle, system, method, doctrine'. It is the governance sense with full authority -- the exact sense CONCEPTS.json forbids -- plus a prudence/wisdom evaluation the math does not contain. Its 'plan, design' sense is the near-miss that makes it dangerous.
sanskrit	nīti	Policy, ethics, statecraft (as in Nītiśāstra). Governance and morality in one word.
sanskrit	karman read in its doctrinal sense	Action-with-moral-consequence across lives. The author's own proposed compound is exposed to this; if the fence cannot hold, rebuild on kriyā or ceṣṭā.
latin	consilium	Lewis & Short (retrieved 2026-07-15): 'deliberation, consultation, a considering together, counsel'; 'a determination, resolution, measure, plan, purpose, intention'. It asserts DELIBERATION and INTENTION. Policies are enumerated and scored, not deliberated, and a candidate policy is nobody's intention.
spanish	'política'	Means politics as well as policy. Worst collision available in a corpus with a governance vocabulary.
any	the reinforcement-learning sense (a state-to-action mapping pi(a|s))	A different mathematical object. Here pi indexes a candidate action SEQUENCE. Technically literate readers make this error by default.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and gloss and ask: (a) is this a rule an organisation adopts? (b) does it involve anyone deliberating or deciding wisely? (c) is it a sequence of actions, or a rule for choosing an action in each situation? (d) has it been adopted, or is it just a candidate?

POLICY_PRECISION — policy precision (gamma)

Store: TRUE · Domain: active-inference-math · Confidence: LOW for the coinages. The units claim (nats^-1) and the NOT-MEASURED claim are carried from CONCEPTS.json/NA-02 and are the parts of this entry that matter. · Reviewer: NONE - PENDING EXPERT REVIEW

The inverse-temperature parameter gamma > 0, carrying units of inverse nats, that sharpens the softmax over policies, with gamma approaching 0 giving indifference and gamma approaching infinity giving deterministic argmin G.

register	form	literal gloss	status
sanskrit	कर्मक्रमतैक्ष्ण्य (karma-krama-taikṣṇya)	sharpness over action-sequences	COINED
latin	acumen seriei actionum	sharpness of the action-sequence (distribution)	COINED
english	policy precision (gamma)	The inverse-temperature parameter gamma > 0, in units of inverse nats, that sharpens the softmax over policies; gamma approaching 0 gives indifference and gamma approaching infinity gives deterministic argmin G.	—
hindi	क्रिया-अनुक्रम तीक्ष्णता ()	sharpness over action sequences	PENDING-CITATION
spanish	precisión de la política (gamma)	policy precision; inherits the 'política' collision recorded at POLICY and the 'precisión' collision recorded at ACCURACY	PENDING-CITATION

Ambiguity notes

• UNITS ARE ROUTINELY DROPPED IN THE LITERATURE AND CONCEPTS.json REFUSES TO DROP THEM: the softmax argument must be DIMENSIONLESS and G is in nats, so gamma is in NATS^-1 (inverse nats). A gamma reported without units is not a number, by the same rail that governs NAT. Most published usage omits this.
• NOT-MEASURED AS A NATURAL CONSTANT, and this is a stated fact about the literature, not a hedge: gamma is FIT PER MODEL OR PER SUBJECT, and NO REPLICATED CROSS-SPECIES VALUE EXISTS (NA-02). Any register or text that presents gamma as a constant of nature is asserting something nobody has measured.
• REGISTERED SEPARATELY FROM PRECISION ON PURPOSE, and CONCEPTS.json is explicit: the corpus uses 'precision' in BOTH senses -- inverse variance in NA-04's H_mu-eta blocks, and inverse temperature over policies in NA-02. A register collapsing them MERGES TWO DIFFERENT QUANTITIES. This entry and PRECISION must keep different terms in every register. The Spanish register above currently fails this and the failure is recorded.
• 'Inverse temperature' is a statistical-mechanics loan (beta = 1/kT). It is NOT a temperature, NOT thermal, and NOT physical. Same class of loan as 'free energy' and 'entropy', and the same fence applies: a heat root imports a false physical claim.
• gamma does not change WHICH policy is best -- argmin G is invariant to gamma. It changes only how DECISIVELY the best one is selected. A register implying gamma affects the ranking has misdescribed it: at gamma -> 0 the agent is indifferent among policies it can still rank perfectly well.

Forbidden mistranslations

register	form	why it is wrong
any	any term also used for PRECISION	CONCEPTS.json registers them separately precisely because the corpus uses both senses. Merging them merges inverse variance with inverse temperature -- two different quantities with different units.
latin	praecisio	Lewis & Short (retrieved 2026-07-15): 'A cutting off'; in rhetoric 'a breaking off abruptly, = ἀποσιώπησις'; 'An overreaching (late Lat.)'. It never means precision-as-sharpness or precision-as-exactness. See PRECISION.
any	any heat/temperature root for 'inverse temperature'	A statistical-mechanics loan. gamma is not thermal and has units of inverse nats.
any	any rendering implying gamma is a measured constant	NOT-MEASURED as a natural constant; fit per model or subject, with no replicated cross-species value (NA-02).
any	any rendering implying gamma changes which policy is best	argmin G is invariant to gamma. It sets decisiveness, not ranking.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and gloss and ask: (a) what are its units? (b) is it a constant of nature or something fitted? (c) does raising it change which policy is best, or only how sharply that one is chosen? (d) is it the same quantity as the precision that means inverse variance?

PRECISION — precision

Store: TRUE · Domain: active-inference-math · Confidence: VERY LOW for the Sanskrit, which the author explicitly flags as the weakest coinage in the file and expects to be replaced. HIGH for the praecisio refutation, which is CITED with all four senses quoted. This entry's value is entirely in what it forbids. · Reviewer: NONE - PENDING EXPERT REVIEW

The confidence weighting on a distribution or signal, formally the inverse variance or the inverse-covariance (precision) matrix, which sets how strongly a term is allowed to drive inference.

register	form	literal gloss	status
sanskrit	व्यत्ययव्युत्क्रम (vyatyaya-vyutkrama)	inverse-of-variance	COINED
latin	pondus fiduciae (inversa variantia)	the weight of confidence; the inverse of variance	COINED
english	precision	The confidence weighting on a distribution or signal, formally the inverse variance or the inverse-covariance (precision) matrix; it sets how strongly a term is allowed to drive inference.	—
hindi	प्रसरण-व्युत्क्रम ()	inverse of variance (deliberately NOT 'yathārthatā' or 'śuddhatā')	PENDING-CITATION
spanish	precisión (inversa de la varianza)	precision; the parenthetical is mandatory because Spanish 'precisión' collides with ACCURACY's natural rendering and with POLICY_PRECISION	PENDING-CITATION

Ambiguity notes

• THREE SENSES ARE LIVE AT ONCE AND CONCEPTS.json NAMES ALL THREE. (1) This concept: inverse variance / inverse-covariance matrix, as in NA-04's H_mu-eta blocks. (2) POLICY_PRECISION: inverse temperature over policies, as in NA-02 -- registered separately, and a register collapsing them merges two different quantities. (3) The ORDINARY sense of exactness, or measurement precision (repeatability, as contrasted with accuracy). CONCEPTS.json calls the third 'a third trap'. Neither technical sense is the ordinary one.
• AND THE ORDINARY SENSE IS ACTIVELY MISLEADING: in measurement science, precision (repeatability) is CONTRASTED with accuracy (closeness to truth). Importing that contrast here would pair this concept against ACCURACY as though they were opposites in the same scheme. They are not: ACCURACY is a term of F, and precision is a confidence weighting. They belong to different parts of the machinery and are not a pair.
• THE LATIN COGNATE IS A TRAP WITH A CITATION, and it is the cleanest example in this file: 'praecisio' is a real classical word, it looks exactly like 'precision', and Lewis & Short shows it means 'a cutting off' -- with the primary attested instance being a cutting off of genitals (Apuleius), and the rhetorical sense being aposiopesis, breaking off mid-sentence. A translator who trusts the cognate produces nonsense with impeccable classical credentials.
• Precision SETS HOW STRONGLY A TERM IS ALLOWED TO DRIVE INFERENCE. That is its functional content and it is what a register must carry: it is a gain, a weighting, a volume knob on a signal -- not a property of the signal's correctness. A precise signal is one that is allowed to matter, not one that is right. A high-precision wrong signal drives inference hard in the wrong direction, which is the whole reason precision is a modelling choice with consequences.
• 'Confidence' in the definition is the weighting, NOT a felt confidence and NOT a confidence interval. Both are live homonyms; the second is especially hot in this corpus, where M2's verdict rule is stated in terms of confidence-interval bounds.

Forbidden mistranslations

register	form	why it is wrong
latin	praecisio	THE EXEMPLARY FALSE COGNATE. Lewis & Short (retrieved 2026-07-15): 'A cutting off (post-Aug.)', with 'genitalium' (Apuleius) as the attested instance; transf. 'the piece cut off'; in rhetoric 'a breaking off abruptly, = ἀποσιώπησις'; 'An overreaching (late Lat.)'. It has never meant exactness or inverse variance. The cognate's resemblance to English 'precision' is the entire trap.
any	any term also used for POLICY_PRECISION	Merges inverse variance with inverse temperature. Registered separately in CONCEPTS.json for exactly this reason; the corpus uses both senses in different chapters.
any	any exactness, accuracy, correctness, or repeatability root	The ordinary and metrological senses. Precision here is a weighting, not a property of correctness. Importing the metrological sense also falsely pairs it against ACCURACY.
any	any term also used for ACCURACY	Two different objects in the same domain. Spanish currently fails this and the failure is recorded rather than hidden.
any	rendering 'confidence' with a felt-confidence or a confidence-interval root	It is an inverse variance used as a weight. Neither an affect nor an interval estimate.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and gloss and ask: (a) does a high value mean the signal is correct? (b) does it mean the signal is allowed to influence the result strongly? (c) is this the same word your register uses for the sharpness of policy selection? (d) is it the same word you would use for a careful, exact measurement?

PRIOR — prior

Store: TRUE · Domain: active-inference-math · Confidence: LOW for the renderings; the valuable content is the double collision in Latin, which rests on quoted Lewis & Short text retrieved this pass and which the author did not anticipate. · Reviewer: NONE - PENDING EXPERT REVIEW

The distribution p(s) held over hidden states before the current observation is taken into account.

register	form	literal gloss	status
sanskrit	पूर्वसम्भाव्यता (pūrva-sambhāvyatā)	prior probability	COINED
latin	distributio prior	the prior distribution	COINED
english	prior	The distribution p(s) held over hidden states before the current observation is taken into account.	—
hindi	पूर्व प्रायिकता ()	prior probability	PENDING-CITATION
spanish	distribución a priori	prior distribution	PENDING-CITATION

Ambiguity notes

• THE TEMPORAL HOMONYM IS NAMED IN CONCEPTS.json AND IS UNAVOIDABLE IN LATIN: 'prior' means 'before the current observation is taken into account', which is a LOGICAL/EPISTEMIC priority, not a temporal one. A prior can be formed after the observation arrives, so long as the observation is not conditioned on. Lewis & Short's quoted 'former, previous, prior, freq. to be translated first' is precisely the temporal sense to avoid -- so the Latin register's most natural word delivers the trap with classical authority.
• A SECOND AND LESS OBVIOUS COLLISION, found by reading the actual dictionary entry rather than assuming it: Lewis & Short's tropological sense of 'prior' is 'better, superior, PREFERABLE, more excellent or important'. So in Latin the word for PRIOR also means PREFERABLE -- and PREFERENCE is a separately registered concept in this same domain. A Latin reader encountering 'distributio prior' may hear 'the preferable distribution'. That would merge p(s) with p(o|C), which are different objects doing opposite jobs (one is what you believe before looking, the other is what you want to see). This collision was not anticipated before the entry was retrieved and is recorded as a finding.
• 'A priori' in the Kantian sense (knowledge independent of experience) is a third live sense, and it is hot in Spanish and in philosophical registers generally. The Bayesian prior is not a priori in that sense at all: it is usually built from experience.
• READ WITH CAVITY_PRINCIPLE (M22): a hierarchy level must NEVER treat an upstream prior as fresh evidence, but must divide it out. Double-counting your own prior is how a hierarchy manufactures confidence from nothing. A register whose prior-word suggests 'established background knowledge' invites exactly that error.
• The prior is p(s) over HIDDEN STATES. Distinct from D_TENSOR, which is specifically the prior at t=0 rather than the running prior, and from E_TENSOR, which is a prior over POLICIES. Three priors in the machinery; a register needs to keep them apart even though this file only covers this one.

Forbidden mistranslations

register	form	why it is wrong
latin	'prior' read in its tropological sense	Lewis & Short: 'better, superior, preferable, more excellent or important'. It would merge PRIOR with PREFERENCE -- p(s) with p(o|C) -- which are different objects with opposite roles. Found by reading the entry, not predicted.
latin	'prior' read in its literal temporal sense	Lewis & Short: 'former, previous, prior, freq. to be translated first'. The Bayesian sense is logical, not temporal: a prior may be written after the data arrive.
any	the Kantian 'a priori'	Knowledge independent of experience. A Bayesian prior is typically derived FROM experience. Live in Spanish especially, where 'a priori' is ordinary usage.
any	any rendering as 'background knowledge', 'established belief', or 'what we already know'	Invites the cavity-principle violation (M22): treating an upstream prior as fresh evidence. It is a distribution, not a store of knowledge.
any	any term also used for D_TENSOR or E_TENSOR	D is the prior at t=0; E is a prior over policies; this is the running prior over hidden states. Three distinct objects.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and gloss and ask: (a) does 'prior' here mean earlier in time? (b) does it mean better or preferable? (c) could a prior be written down after the observation arrives? (d) does it mean knowledge that does not come from experience?

PREFERENCE — preference (C)

Store: TRUE · Domain: active-inference-math · Confidence: LOW. iṣṭa is CITED and is the closest attested Sanskrit candidate, but MW's own glosses show it carries affective and honorific senses the concept forbids, and a ritual homonym sits on the same form. The Latin collision with PRIOR is the notable finding. · Reviewer: NONE - PENDING EXPERT REVIEW

A prior p(o|C) over PREFERRED observations that encodes a goal as a distribution rather than as a reward signal, and which breaks the tie between revising beliefs and acting on the world.

register	form	literal gloss	status
sanskrit	इष्टवितरण (iṣṭa-vitaraṇa)	distribution over the desired (observations)	COINED
latin	distributio optatorum	distribution of the things wished for	COINED
english	preference (C)	A prior p(o|C) over preferred observations that encodes a goal as a distribution rather than as a reward signal; it breaks the tie between revising beliefs and acting on the world.	—
hindi	अभीष्ट वितरण ()	distribution over the desired (deliberately NOT 'pasand', which is liking/taste)	PENDING-CITATION
spanish	distribución de preferencias (C)	preference distribution; 'preferencia' carries the liking sense and is recorded with a warning	PENDING-CITATION

Ambiguity notes

• NOT A LIKING, A TASTE, OR A FELT DESIRE, and CONCEPTS.json states that an affect rendering ASSERTS AN EXPERIENCER. This is the entry's hardest fence and the one every register's natural word fails: English 'preference', Spanish 'preferencia', Hindi 'pasand', and MW's own glosses of iṣṭa ('liked, beloved, cherished') are ALL affective. p(o|C) is a probability distribution. A thermostat has one.
• THE THERMOREGULATION CASE FIXES WHY IT IS NEEDED and is worth carrying into every register: a cold mammal can REVISE ITS PREDICTION of its own temperature, or it can SHIVER -- and only one is survivable (NA-04). Both minimize the same quantity. The preference is what BREAKS THE TIE between changing the belief and changing the world. Without C, 'just believe you are warm' is a perfectly good solution to the free-energy problem, and the animal dies comfortable.
• IT ENCODES A GOAL AS A DISTRIBUTION RATHER THAN AS A REWARD SIGNAL. This is a deliberate architectural choice, not a notational one, and it is what separates this framework from reinforcement learning. A register that renders C as reward, utility, objective, or goal-state has silently re-imported RL and made the framework's distinguishing claim invisible.
• 'C' is a conventional label, like A/B/D/E. It is not an abbreviation of anything and must be transliterated or indexed, never translated as a word.
• SOVEREIGN NOTE, carried from C_TENSOR: C is -- alone among A, B, C, D and E -- SET rather than learned by Dirichlet counts. A goal learned from its own counts would have no purchase; C is where the goal enters from outside. A register implying preferences are learned or acquired contradicts the corpus's own observed asymmetry.

Forbidden mistranslations

register	form	why it is wrong
any	any liking, taste, desire, or wanting root	Asserts an experiencer. CONCEPTS.json: 'NOT a liking, taste, or felt desire -- an affect rendering asserts an experiencer.' A thermostat has a p(o|C).
sanskrit	iṣṭa read via MW's second entry (√yaj)	MW's second iṣṭa (retrieved 2026-07-15) means 'sacrificed, worshipped with sacrifices'; n. 'sacrifice'. A ritual homonym sitting directly on the chosen word. The compound must be read with √iṣ, not √yaj.
sanskrit	iṣṭa read via its honorific glosses	MW also gives 'beloved, cherished, reverenced'. p(o|C) reveres nothing.
sanskrit	icchā or kāma	Desire in the strongest affective sense; kāma is besides one of the puruṣārthas and a deity. Asserts a wanting subject and imports an ethical framework.
latin	using the adjective 'prior' (trop. 'preferable') to render 'preferred'	Lewis & Short gives 'preferable' as a tropological sense of 'prior' -- the same word this file uses for the separately registered concept PRIOR. It would merge p(s) with p(o|C).
any	'reward', 'utility', 'objective', or 'goal state'	C is a prior over observations, not a reward signal. Rendering it as reward erases the framework's distinguishing architectural claim and re-imports RL.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and gloss and ask: (a) does this word imply someone who WANTS something? (b) could a thermostat have one? (c) is it a reward the system receives, or a distribution it holds? (d) is it learned from experience or set from outside?

LIKELIHOOD — likelihood

Store: TRUE · Domain: active-inference-math · Confidence: LOW. No component verified; the obvious Latin candidate was sought and not resolved, and is fenced as PENDING rather than reported absent. The sambhāvanā refutation is the one CITED result and it is a negative one. · Reviewer: NONE - PENDING EXPERT REVIEW

The mapping p(o|s) from hidden states to observations, realized as the A tensor in a discrete POMDP.

register	form	literal gloss	status
sanskrit	अवस्थादर्शनसम्बन्ध (avasthā-darśana-sambandha)	the state-to-observation relation	COINED
latin	relatio status ad observata	the relation of state to observed things	COINED
english	likelihood	The mapping p(o|s) from hidden states to observations, realized as the A tensor in a discrete POMDP.	—
hindi	सम्भाविता फलन ()	likelihood function	PENDING-CITATION
spanish	verosimilitud	likelihood (the standing statistical term)	PENDING-CITATION

Ambiguity notes

• THE DEFINING TRAP, named in CONCEPTS.json: this is NOT probability in the ordinary sense of 'how likely something is to happen'. It is THE STATE-TO-OBSERVATION MAPPING. The English word 'likelihood' is one of the worst terms of art in statistics precisely because its ordinary meaning is 'probability', and every register's natural rendering (Spanish verosimilitud, Hindi sambhāvitā, Latin verisimilitudo) inherits that ordinary meaning wholesale.
• THE TECHNICAL POINT THAT THE ORDINARY SENSE DESTROYS: as a likelihood, p(o|s) is read as a function OF s FOR FIXED o -- and in that reading it does NOT integrate to 1 over s and IS NOT A PROBABILITY DISTRIBUTION over s. A register that renders it with a probability word asserts a normalization that does not hold. This is a mathematical error, not a stylistic one.
• It is a MAPPING (equivalently, in the discrete case, a tensor -- the A tensor). A register that renders it as a quantity or a number has named the wrong kind of object: a likelihood is the whole map, not one of its values.
• ITS EXPECTED ENTROPY IS AMBIGUITY. That is the link that makes AMBIGUITY a property of a channel rather than a mental state, and a register's likelihood-word must survive being nested inside E_q(s|pi)[H[p(o|s)]].
• 'Likelihood' vs 'marginal likelihood' (EVIDENCE) is a further collision inside this same domain: p(o|s) is the likelihood, p(o) is the marginal likelihood, and they are different objects. The Spanish register above uses 'verosimilitud' for this concept and 'verosimilitud marginal' for EVIDENCE, which keeps them related and distinct -- but only if the modifier is never dropped.

Forbidden mistranslations

register	form	why it is wrong
any	any ordinary 'probability' or 'chance' root	CONCEPTS.json: 'NOT probability in the ordinary sense of how likely something is to happen.' Additionally asserts a normalization: p(o|s) as a function of s does not integrate to 1 over s.
any	'plausibility', 'believability', 'verisimilitude'	The ordinary senses of the Romance cognates. They describe how convincing something seems, which is a property of a story, not a map from states to observations.
sanskrit	sambhāvanā	MW (retrieved 2026-07-15) gives 'imagination, supposition, assumption; fitness, adequacy; competency, ability' -- and no mathematical probability sense at all. Using it would assert a dictionary sense the dictionary lacks.
sanskrit	darśana read as a philosophical system	The ṣaḍdarśana are the six orthodox systems. The author's own coinage is exposed to this and it is flagged.
any	any rendering as a single number or quantity	It is a mapping / tensor, not one of its values.

Back-translation test — status: NOT RUN: no reader has been asked

Give a statistically literate reader the target-register term and gloss and ask: (a) is this a probability distribution over the hidden state? (b) does it sum to 1 over the states? (c) is it a number, or a map? (d) is it the same as the marginal likelihood p(o)?

TRANSITION — transition

Store: TRUE · Domain: active-inference-math · Confidence: LOW. No component verified in either classical register; the obvious Latin candidate was sought and not resolved, and is fenced rather than reported absent. · Reviewer: NONE - PENDING EXPERT REVIEW

The mapping p(s'|s,u) from a current hidden state and an action to the next hidden state, realized as the B tensor in a discrete POMDP.

register	form	literal gloss	status
sanskrit	अवस्थान्तरगमन (avasthāntara-gamana)	passage to another state	COINED
latin	transitio inter status	a passing between states	COINED
english	transition	The mapping p(s'|s,u) from a current hidden state and an action to the next hidden state, realized as the B tensor in a discrete POMDP.	—
hindi	अवस्था-संक्रमण ()	state transition	PENDING-CITATION
spanish	transición	transition	PENDING-CITATION

Ambiguity notes

• SOVEREIGN HOMONYM AND CONCEPTS.json COUNTS THREE SENSES IN THIS CORPUS: (1) this concept, p(s'|s,u), the B tensor; (2) a MAJOR_TRANSITION in the evolutionary sense (Maynard Smith & Szathmary) -- which lives in the NATURA vocabulary and whose citation can never be a UNI gate; (3) a DD-TDD state transition, which is a build-process object. Three different senses, and a register that renders them alike lets an evolutionary claim, a build-process step and a tensor share a word.
• IT IS A MAPPING, NOT AN EVENT. The commonest error: 'transition' in ordinary usage names an occurrence (the system transitioned), whereas p(s'|s,u) is the whole conditional map, including for transitions that never occur. The B tensor is not a history.
• IT IS CONDITIONED ON AN ACTION u, and that is what distinguishes it from a plain Markov chain. CONTROLLABILITY LIVES IN B (as ambiguity lives in A). A register that drops the action-dependence has rendered a different, uncontrolled object and has removed the agent's purchase on the world.
• Not a transition in the sense of a change of regime, a phase transition, or a political/organisational transition. The regime sense is especially live in this corpus, which carries REGIME_BOUNDARY in the NATURA wing -- and that is a fourth collision.
• s' (the next state) is not a different variable from s; the prime is a time index. A register that renders s and s' as two different KINDS of state has misread the notation.

Forbidden mistranslations

register	form	why it is wrong
any	any term also used for MAJOR_TRANSITION (evolutionary) or for a DD-TDD state transition	CONCEPTS.json registers three senses in this corpus. Merging them would let a NATURA-side evolutionary claim or a build-process step share a word with a tensor -- and a nature citation is never a UNI gate.
any	any rendering as an event, occurrence, or change that happened	It is a conditional mapping, defined over transitions that never occur. Not a history.
any	any rendering that drops the action-dependence	p(s'|s,u) is conditioned on u. Without it, this is a plain Markov chain and the agent has no purchase on the world. Controllability lives in B.
any	'phase transition' or any regime-change root	Collides with REGIME_BOUNDARY in the NATURA wing, a different concept in a sovereign vocabulary.
hindi	saṅkramaṇ used without disambiguation	Also means infection/contagion in modern Hindi usage. The author's own proposal is exposed and it is flagged.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and gloss and ask: (a) does this name something that happened, or a rule saying what follows what? (b) does the rule depend on what the agent did? (c) is this the same word your register would use for evolution's major transitions, or for a phase transition?

ELBO — evidence lower bound (ELBO)

Store: TRUE · Domain: active-inference-math · Confidence: VERY LOW for the coinages, which are unwieldy and built on an unlicensed probability head. The identity claim (ELBO = -F) and the inherited evidence homonym are the load-bearing content. · Reviewer: NONE - PENDING EXPERT REVIEW

The negative of variational free energy, being a lower bound on the log evidence ln p(o), such that maximizing it is identical to minimizing F.

register	form	literal gloss	status
sanskrit	प्रान्तिकसम्भाव्यताधःसीमा (prāntika-sambhāvyatā-adhaḥ-sīmā)	lower bound on the marginal probability (logarithm)	COINED
latin	terminus inferior probabilitatis marginalis	lower bound of the marginal probability	COINED
english	evidence lower bound (ELBO)	The negative of variational free energy; a lower bound on the log evidence ln p(o), such that maximizing it is identical to minimizing F.	—
hindi	साक्ष्य निम्न परिबंध ()	evidence lower bound -- WARNING: 'sākṣya' means evidence-as-testimony and probably fails this entry's own fence	PENDING-CITATION
spanish	cota inferior de la evidencia (ELBO)	evidence lower bound; inherits the 'evidencia' collision recorded at EVIDENCE	PENDING-CITATION

Ambiguity notes

• SAME QUANTITY AS F WITH THE SIGN FLIPPED, and CONCEPTS.json is emphatic: A REGISTER MUST NOT PRESENT THEM AS TWO MECHANISMS. The machine-learning literature MAXIMIZES the ELBO where the active-inference literature MINIMIZES F. Two communities, two sign conventions, ONE quantity. A reader who meets both and concludes there are two things has been misled by vocabulary alone.
• THE ENTRY INHERITS EVIDENCE'S SOVEREIGN HOMONYM AND MAKES IT WORSE: 'evidence' here is p(o), the marginal likelihood -- NOT evidence in the method sense (receipts, EVIDENCE_CLASS). 'Evidence lower bound' therefore reads, to anyone in this corpus, as though it bounded how well-evidenced a claim is. It does not. It bounds a model's normalizing constant. The Hindi calque above uses 'sākṣya' (testimony) and is flagged as probably failing on exactly this.
• 'Lower bound' means it sits BELOW ln p(o) and rises toward it. Contrast BOUND_GAP, where F sits ABOVE surprisal. Same slack, opposite direction, because of the sign flip -- and a register that renders 'bound' identically in both places without the direction will confuse them.
• 'Bound' is a mathematical bound, not a fetter, tie, or obligation. Same fence as BOUND_GAP; Hindi 'paribandh' and Sanskrit bandha both carry the fetter sense.
• ELBO is an acronym in English and is frequently left untranslated in every language, including in Hindi and Spanish technical writing. That is a defensible treatment -- like 'nat', it is a name -- but unlike 'nat' it is an acronym whose expansion carries the trap, so leaving it opaque HIDES the evidence homonym rather than resolving it.

Forbidden mistranslations

register	form	why it is wrong
any	any rendering presenting the ELBO and F as two different mechanisms or quantities	CONCEPTS.json: 'Same quantity as F with the sign flipped... A register must not present them as two mechanisms.' Maximizing ELBO IS minimizing F.
any	rendering 'evidence' here with the method-sense evidence word (receipts, testimony, proof)	Inherits EVIDENCE's sovereign homonym. It would make the ELBO read as a bound on how well-supported a claim is, which would let an arithmetic bound look like a UNI gate.
hindi	sākṣya	Testimony/courtroom evidence. The author's own proposed Hindi is exposed to this and it is flagged rather than hidden.
any	any fetter/bondage root for 'bound'	Same fence as BOUND_GAP. It is an inequality's bound.
any	rendering 'lower bound' without preserving the direction	ELBO bounds ln p(o) from BELOW; F bounds surprisal from ABOVE. The sign flip is the whole relation between them.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader who knows variational inference the target-register terms for ELBO and for VARIATIONAL_FREE_ENERGY and ask: (a) are these two different quantities or one quantity with two sign conventions? (b) does 'evidence' here mean support for a claim? (c) does maximizing the first do the same work as minimizing the second?

FREE_ENERGY_PRINCIPLE — free energy principle

Store: TRUE · Domain: active-inference-math · Confidence: LOW for the coinages, and the author flags siddhānta as probably overclaiming. The two-directions-travel structure is the load-bearing content and is carried directly from CONCEPTS.json. · Reviewer: NONE - PENDING EXPERT REVIEW

The proposal that self-organizing systems at non-equilibrium steady state minimize a variational free-energy bound on surprisal, which designates a MODEL STRUCTURE rather than an empirical finding.

register	form	literal gloss	status
sanskrit	मुक्तमानसिद्धान्त (mukta-māna-siddhānta)	the free-measure doctrine	COINED
latin	principium mensurae liberae	the principle of the free measure	COINED
english	free energy principle	The proposal that self-organizing systems at non-equilibrium steady state minimize a variational free-energy bound on surprisal; it designates a model structure rather than an empirical finding.	—
hindi	मुक्त ऊर्जा सिद्धांत ()	free energy principle (calque, carrying the forbidden energy root)	PENDING-CITATION
spanish	principio de energía libre	free energy principle (calque, carrying the forbidden energy root)	PENDING-CITATION

Ambiguity notes

• ITS STATUS IS GENUINELY CONTESTED AND BOTH DIRECTIONS MUST TRAVEL. CONCEPTS.json is unusually careful here and this entry must not simplify it: (1) demanding the FEP itself be falsifiable is a CATEGORY ERROR (Andrews 2021), because it designates a model structure; AND (2) a model structure IS NOT AN EMPIRICAL FINDING. Specific process-theory models built on it are falsifiable; the structure is not. NEITHER MAY BORROW THE OTHER'S CREDIT. A register that lets the FEP read as an established empirical law has taken a side; so has one that lets it read as unfalsifiable pseudoscience. Both are wrong.
• 'PRINCIPLE' IS THE PROBLEM WORD IN EVERY REGISTER. In English it can mean a law of nature (the uncertainty principle), a rule of conduct (a person of principle), a foundational assumption (first principles), or an organising idea. The FEP is closest to the third. Sanskrit siddhānta -- the author's own proposal -- means an ESTABLISHED CONCLUSION of a school, which leans toward the first and second and probably OVERCLAIMS. It is flagged for that reason.
• IT IS A PROPOSAL, and the definition says so. Not a discovery, not a law, not a finding, not a theorem. A register whose word for 'principle' implies discovery has upgraded a model structure into an empirical result, which is precisely the move both CONCEPTS.json and the honesty rail forbid.
• 'Free energy' carries the same physics-loan fence as everywhere else in this file: a heat/energy root imports a false physical claim. The Hindi and Spanish calques take the energy root and the conflict is recorded, as at VARIATIONAL_FREE_ENERGY.
• 'Non-equilibrium steady state' is a technical condition from statistical physics and is not decoration: the claim is scoped to systems at NESS. A register that drops the qualifier renders a much stronger and much less defensible claim -- that everything minimizes free energy.

Forbidden mistranslations

register	form	why it is wrong
any	any rendering as a natural LAW or an established empirical finding	It designates a MODEL STRUCTURE, not an empirical finding. Rendering it as a law lets the structure borrow the credit of the falsifiable process-theory models built on it -- which CONCEPTS.json forbids in that exact direction.
any	any rendering implying it is unfalsifiable and therefore empty	The opposite error and equally forbidden: demanding the FEP itself be falsifiable is a category error (Andrews 2021). Both directions travel; a register may take neither side.
sanskrit	siddhānta read as a school's settled doctrine	Overclaims. siddhānta is an ESTABLISHED conclusion; the FEP is a proposal whose status is contested. The author's own coinage is exposed to this and is flagged.
any	any heat/energy root	Same false physical claim as at VARIATIONAL_FREE_ENERGY.
any	dropping 'at non-equilibrium steady state'	Renders an unscoped and much stronger claim. The condition is part of the proposal.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and gloss and ask: (a) is this a law of nature that has been established by experiment? (b) is it an empty or unfalsifiable claim? (c) is it a way of setting up models, or a finding about the world? (d) does it apply to all systems or to a scoped class?

ACTION_AS_INFERENCE — action as inference

Store: TRUE · Domain: active-inference-math · Confidence: LOW-MEDIUM. anumāna is CITED in the required sense and the sentence-rather-than-compound decision is well-motivated by the objective-genitive risk. But karman's doctrinal freight is unresolved and no Latin component was verified. · Reviewer: NONE - PENDING EXPERT REVIEW

The consequence of the blanket partition that a prediction error can be resolved either by changing internal states to fit the world or by changing active states until the world fits the model, both minimizing the same quantity.

register	form	literal gloss	status
sanskrit	कर्मैवानुमानम् (karmaivānumānam)	action IS precisely inference	COINED
latin	actio ipsa illatio est	action itself IS inference	COINED
english	action as inference	The consequence of the blanket partition that a prediction error can be resolved either by changing internal states to fit the world, or by changing active states until the world fits the model; both minimize the same quantity.	—
hindi	क्रिया ही अनुमान है ()	action IS inference (rendered as a sentence, with the emphatic 'hī', for the same reason as the Sanskrit)	PENDING-CITATION
spanish	la acción como inferencia	action as inference; Spanish 'como' is ambiguous between 'in the manner of' and 'in the capacity of' and the second is meant	PENDING-CITATION

Ambiguity notes

• 'AS' IS THE WHOLE PROBLEM AND IT IS INVISIBLE IN ENGLISH. The English 'action as inference' can be read three ways: (1) action resembling inference (simile -- wrong); (2) action considered in its capacity as inference (identity -- CORRECT); (3) action in the role of inference (functional substitute -- wrong, it is not a stand-in). Spanish 'como' has the same three-way ambiguity, which is why the Sanskrit and Hindi above force the identity reading with an emphatic particle instead.
• ACTION IS NOT A FACULTY BOLTED ONTO INFERENCE; it is inference run through the other half of the blanket. CONCEPTS.json puts it exactly that way. A register that renders this as a comparison, an analogy, or a metaphor has turned a structural consequence into a figure of speech.
• THE PREFERENCE C IS WHAT BREAKS THE TIE between the two routes, and the entry is incomplete without it. Both routes minimize the same quantity, so the formalism ALONE does not say whether to shiver or to believe you are warm. Read with PREFERENCE; the thermoregulation case is what makes the tie-break non-optional.
• 'Changing active states until the world fits the model' is not wishful thinking and not self-deception -- it is ACTING. The phrasing invites a misreading in which the agent simply decides the world is different. The active states are the real output channel; the world really changes.
• 'Prediction error' is a term this corpus CUT from the concept list as redundant (it appears in the cut_as_redundant list of CONCEPTS.json), yet the definition of THIS concept uses the phrase. That is a small internal tension in the source and is recorded here rather than silently smoothed: this entry's definition depends on a term the backbone deliberately does not register.

Forbidden mistranslations

register	form	why it is wrong
any	any rendering as simile, analogy, or 'action is LIKE inference'	It is an identity, not a comparison. The blanket partition makes action a route of inference, not a thing resembling one.
sanskrit	the compound karma-anumāna	Reads as 'inference about action' (objective genitive) or 'action-type inference' (descriptive) -- both reverse or dilute the claim. This is why the entry is a sentence with eva rather than a compound.
spanish	'la acción como inferencia' read with 'como' = 'in the manner of'	Delivers the simile reading. 'la acción ES inferencia' may be the safer form and is noted in the Spanish citation.
any	any rendering suggesting the agent changes its beliefs INSTEAD of acting	Inverts the branch. Changing active states means acting on the world; the belief-revision route is the OTHER branch, and choosing it in the thermoregulation case is fatal.
sanskrit	karman read doctrinally	Action-with-moral-consequence across lives. Same freight fenced at POLICY; if it cannot be held, rebuild on kriyā.

Back-translation test — status: NOT RUN: no reader has been asked

Give a reader the target-register term and gloss and ask: (a) does this say action RESEMBLES inference, or that it IS inference? (b) does it mean the agent changes its mind instead of acting? (c) what decides which of the two routes is taken?