The lexicon — lexicon

RENDERED FROM STRUCTURED DATA — this page is not a markdown file. Source: lexicon/terms/pomdp-tensors.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 file for the nine concepts whose domain is 'pomdp-tensors' in lexicon/CONCEPTS.json. Each entry renders ONE concept -- fixed by its definition, never by its English headword -- into Sanskrit, Latin, English, Hindi and Spanish, and records what each rendering gets WRONG alongside what it gets right. NOTHING HERE IS A CLAIM OF PERFECT TRANSLATION. Per rule 18.14 of the operator's charter ('Do not claim perfect translation'), every rendering here is PENDING until an independent back-translation test is RUN, and no test in this file has been run: no translator has been engaged and the reviewer field is NONE for all nine entries. Statuses record what was SEARCHED and FOUND on 2026-07-15 against the named authorities, including the negative results. This file raises no UNI rung, discharges no gate, and states no fact about nature.

authored 2026-07-15 · reviewer: NONE - PENDING EXPERT REVIEW

Falsifiers for this file

• Any register entry marked CITED whose quoted authority does not in fact carry that headword, or does not carry it in the sense claimed. That entry is a pseudopropheta and must be struck to PENDING-CITATION.
• Any COINED entry lacking its morphological derivation.
• Any back_translation_test reported as PASSED while reviewer remains 'NONE - PENDING EXPERT REVIEW'. No test in this file has been run; a PASS appearing without a named tester is a forgery.
• A register renders GENERATIVE_MODEL and GENERATIVE_PROCESS with one term (cross-domain; those concepts live in active-inference-math, but a tensor entry that blurs the model/process line trips the same wire).
• Any entry that lets a Sanskrit doctrinal term quietly assert a metaphysics -- e.g. rendering HIDDEN_STATE with avyakta (Samkhya's unmanifest prakrti) or E_TENSOR with vasana (karmic latency). The ambiguity_notes exist precisely to catch this and are the most load-bearing field in each entry.
• Any entry that treats the C tensor as learned, or the A/B/D/E tensors as set. The asymmetry is OBSERVED in the corpus (01-kitchen-rules M13: learning covers A/B/D/E, never C) and a register that loses it has lost the concept.
• A status field holds anything other than one bare token from {CITED, PENDING-CITATION, COINED, NOT-ATTEMPTED}, or a mixed rendering is graded at the strength of its strongest member rather than its weakest. Run the enum validator in this file's header comment; a compound of cited members is COINED, never CITED.

POMDP —

Store: TRUE · Confidence: LOW for Sanskrit and Latin (both wholesale coinages for a 20th-century formalism; the Latin observabilis trap and the Sanskrit nirṇaya 'completeness' conflict are unrepaired). MEDIUM for Spanish (established calque, unsourced). LOW for Hindi (unsourced). HIGH only for English, which is the concept's own label and not a translation. · Reviewer: NONE - PENDING EXPERT REVIEW

A discrete formalism of hidden states, observations and actions in which states are not directly visible, over which this corpus's engine runs the loop perceive, EFE-plan, act, learn.

register	form	literal gloss	status
sanskrit	अंशतःप्रत्यक्षा मार्कोव-निर्णय-प्रक्रिया (aṃśataḥ-pratyakṣā mārkova-nirṇaya-prakriyā)	a Markov decision-procedure, perceptible in part	COINED
latin	PROCESSUS DECISIONIS MARKOVIANUS PARTIM OBSERVABILIS	a Markovian process of decision, observable in part	COINED
english	partially observable Markov decision process	A discrete formalism of hidden states, observations and actions in which the states are not directly visible, so that an agent must act from a belief over states rather than from the states themselves; the formalism over which this corpus's engine runs perceive, EFE-plan, act, learn.	—
hindi	आंशिक रूप से प्रेक्षणीय मार्कोव निर्णय प्रक्रिया ()	partially observable Markov decision process	PENDING-CITATION
spanish	proceso de decisión de Márkov parcialmente observable	partially observable Markov decision process	PENDING-CITATION

Ambiguity notes

• THE MOST DANGEROUS WORD IN THE LATIN IS observabilis. L&S glosses it 'remarkable, observable' and marks it post-Augustan, and its attested uses do not mean 'capable of being observed' in the technical sense of observability theory. The Vulgate citation L&S gives -- Exod. 12.42, 'nox ista est observabilis Domini' -- means 'this is the night TO BE KEPT/celebrated of the Lord', i.e. a night to be religiously OBSERVED. Latin observo carries both 'to watch/note' and 'to keep/observe (a rite, a law)', and observabilis inherits the second. To a Latinist, PARTIM OBSERVABILIS therefore risks reading 'partly remarkable' or 'partly to-be-ritually-kept'. The word is retained for want of a better deverbative, but this trap is live and is the first thing a reviewer should attack.
• 'Partially observable' is the load-bearing qualifier (CONCEPTS.json) and the reason a recognition density is needed at all. A register that drops 'partim'/'aṃśataḥ' does not name a weaker version of the concept -- it names a DIFFERENT formalism (a plain MDP), in which no belief state and no inference are required. The qualifier is not decoration.
• Sanskrit nirṇaya [MW 555,1] leads with 'COMPLETE ascertainment'. That is precisely what a partially observable process denies: in a POMDP the ascertainment is constitutively incomplete. The compound therefore sets its own qualifier fighting its own head noun. This is recorded rather than repaired because every alternative considered was worse; it is the single weakest joint in the Sanskrit entry.
• nirṇaya further carries legal 'sentence, verdict' and logical 'deduction, inference' senses [MW 555,1]. The decision in a POMDP is an action selection under uncertainty -- neither a juridical verdict nor a deductive conclusion.
• Sanskrit prakriyā [MW 653] carries 'a ceremony, observance, formality' and 'a chapter (esp. the introductory chapter of a work)'. A POMDP is neither a rite nor a book-division. Note the accidental rhyme with the Latin trap above: BOTH registers' words for the process/observing drift toward RITUAL OBSERVANCE. That is a coincidence of two independent lexicons, not evidence of anything, and must not be presented as a finding.
• Latin decisio is literally 'a cutting off' (from decīdo) and its figurative sense II is 'a decision, SETTLEMENT, AGREEMENT' -- Cicero's uses are of legal settlements between parties. A POMDP decision is not negotiated and has no counterparty.
• 'Markov' names a person and imports nothing but the memorylessness property. No register may gloss it descriptively (e.g. as 'memoryless'), because the concept is fixed to the formalism's conventional name, and a descriptive gloss would silently assert which property is meant.

Forbidden mistranslations

register	form	why it is wrong
		Puts Markovianae in agreement with decisionis. It is the process that carries the Markov property, not the decision. A grammatically fluent error that inverts the formalism.
		Yields a Markov decision process -- a different and strictly easier formalism, with no hidden state, no belief, and no need for inference. This is not a translation of the concept but a replacement of it.
		gūḍha [MW 360,3] means 'covered, hidden, concealed, invisible, SECRET, PRIVATE'. It would assert that the process is clandestine. See HIDDEN_STATE.
		'observado' (observed) for 'observable'. The states are not partly observed; they are partly OBSERVABLE -- a modal claim about what could be seen, not a report of what was.

Back-translation test — status: NOT YET RUN -- no translator engaged. Any PASS recorded here without a named tester is a forgery.

Give an independent translator competent in the register ONLY the rendering, with this file's definition withheld. Ask: 'What object does this term denote? Answer in English in one sentence.'

A_TENSOR —

Store: TRUE · Confidence: LOW. Both learned registers are coinages for an object neither lexicon knows, the Spanish rests on a convention its own dictionary does not record, and the Hindi is unsourced. The one thing here held with confidence is a NEGATIVE: verisimilitudo is decisively wrong, and that is evidenced. · Reviewer: NONE - PENDING EXPERT REVIEW

The tensor encoding p(o|s), mapping hidden states to observations, and one of the four tensors this corpus's engine learns by conjugate-Dirichlet counts.

register	form	literal gloss	status
sanskrit	A-प्रदिश (A-pradiśa)	the A-tensor (lit., by MW's attested sense, 'the A-indication/direction' -- which is NOT what is meant)	COINED
latin	TENSOR A	the A-stretcher / the A-tensor	COINED
english	A tensor (likelihood)	The tensor encoding p(o|s) -- the probability of each observation given each hidden state -- which maps hidden states to observations and is one of the four tensors (A, B, D, E) that this corpus's engine learns by conjugate-Dirichlet counts.	—
hindi	A-प्रदिश ()	the A tensor	PENDING-CITATION
spanish	tensor A (verosimilitud)	the A tensor (likelihood)	PENDING-CITATION

Ambiguity notes

• THE CENTRAL TRAP OF THIS ENTRY: a likelihood is NOT a probability of the hidden state. p(o|s) read as a function of s is not a distribution over s and does not normalise over s. Every register word here that means 'how probable the state is' or 'how true it is' names a different object -- the posterior, or worse, a truth-value.
• This is why Latin verisimilitudo is fatal and not merely infelicitous. L&S does not even grant it an independent entry; the headword reads 'vērĭsĭmĭlis, vērĭsĭmĭlĭter, and vērĭsĭmĭlĭtūdo, more correctly written separately, vērī simĭlis, etc., v. under verus and similis' -- i.e. the word IS 'veri similis', LIKE THE TRUTH. It asserts likeness-to-truth, which is exactly the assertion a likelihood does not make.
• SPANISH AND LATIN DIVERGE HERE, AND THE DIVERGENCE MUST NOT BE LAUNDERED. Spanish has conventionalised 'verosimilitud' for statistical likelihood; Latin has not. The Spanish convention is a fait accompli of usage and is admissible in Spanish on those terms. It may NOT be back-borrowed into Latin on the strength of the cognate: the Latin word's attested semantics actively contradict the concept. A translator who reasons 'Spanish says verosimilitud, so Latin says verisimilitudo' has produced a pseudopropheta by cognate.
• Sanskrit pradiś [MW 679,3] carries 'order, command, dominion'. The A tensor commands nothing and has dominion over nothing. These senses are wholly absent from the concept and are live in the attested word.
• tantra [MW 436,1-2] is the tempting etymological parallel: MW gives 'a loom', 'the warp', and 'the leading or principal or essential part, main point, characteristic feature, model, type, system, framework'. The loom/warp sense is a genuine semantic cousin of Latin tendo's 'stretch'. It is nevertheless FORBIDDEN (below), because MW's same entry gives 'a class of works teaching magical and mystical formularies'. Using tantra would import the entire Tantric corpus into a likelihood matrix.
• CONCEPTS.json: 'ambiguity lives in A, controllability lives in B'. A register that renders A and B with interchangeable words for 'matrix/array' loses the division of labour that makes the pair mean anything.
• The parenthetical '(likelihood)' is a gloss of the tensor's ROLE, not part of its name. The name is the letter A (SD-01). A register that translates 'likelihood' and drops the A has named the role and lost the object.

Forbidden mistranslations

register	form	why it is wrong
		L&S analyses the word as vērī simĭlis, 'like the truth'. It asserts probability-of-being-true. A likelihood is not the probability that the state is true; it is the probability of the DATUM given the state. This inverts the conditional.
		TWO independent failures. (1) MW [1179,1] gives 'imagination, supposition, assumption (also as a rhet. figure...)' and does NOT carry 'probability' -- the authority does not license the sense. (2) Even if it did, 'probability' is the wrong object: likelihood is not a probability over states. The modern Hindi सम्भावना = probability is a later development and cannot be retrojected onto MW.
		Etymologically apt (MW 436: 'a loom', 'the warp' -- the stretch-sense that also gives Latin tensor), and doctrinally catastrophic: MW's same entry gives 'a class of works teaching magical and mystical formularies'. Imports the Tantric corpus wholesale. A metaphysics smuggled in on an etymology is the exact failure CONCEPTS.json warns of.
		Violates SD-01. A is an arbitrary index, not an ordinal. A is the likelihood; it is first in nothing. The letters are conventional labels, not abbreviations.
		Collapses likelihood into probability, discarding the conditional direction that is the whole content of p(o|s).

Back-translation test — status: NOT YET RUN -- no translator engaged.

Give an independent translator ONLY the rendering plus the bare phrase 'in a decision formalism', definition withheld. Ask: 'What does this denote, and what is conditioned on what?'

B_TENSOR —

Store: TRUE · Confidence: LOW. The tensor-word is coined in both learned registers; the best attested Sanskrit candidate for the role-gloss carries death and astrology in its own entry; the Hindi cognate has drifted to 'infection'. The action-conditioning -- the one property that matters -- survives in NO register's word and lives only in the definition. · Reviewer: NONE - PENDING EXPERT REVIEW

The tensor encoding p(s'|s,u), mapping a hidden state and an action to the next hidden state, and one of the four tensors learned by conjugate-Dirichlet counts.

register	form	literal gloss	status
sanskrit	B-प्रदिश (B-pradiśa)	the B-tensor; role-gloss 'the transference-tensor'	COINED
latin	TENSOR B	the B-tensor; role-gloss 'the tensor of passing-over'	COINED
english	B tensor (transition)	The tensor encoding p(s'|s,u) -- the probability of the next hidden state given the current hidden state AND the action taken -- and one of the four tensors (A, B, D, E) learned by conjugate-Dirichlet counts.	—
hindi	B-प्रदिश (संक्रमण) ()	the B tensor (transition)	PENDING-CITATION
spanish	tensor B (transición)	the B tensor (transition)	PENDING-CITATION

Ambiguity notes

• THE DISCRIMINATING PROPERTY IS THE ACTION-CONDITIONING, and every register loses it by default. p(s'|s,u) is conditioned on u, the ACTION. A word meaning bare 'change', 'transition', or 'passage' names p(s'|s) -- an uncontrolled Markov chain. CONCEPTS.json: 'ambiguity lives in A, controllability lives in B'. Drop u and B stops being the seat of controllability and the agent stops being an agent.
• Sanskrit saṃkramaṇa [MW 1127] carries 'passage into another world, DECEASE, DEATH'. संक्रमण-प्रदिश is therefore readable as a death-tensor. This is not a remote sense; it is in the entry.
• saṃkramaṇa [MW 1127] also names the sun's passage from one zodiac sign to another -- the Saṃkrānti. This is the sense a Sanskrit reader is MOST likely to reach first, and it imports jyotiṣa (astrology/astronomy). A state transition is not a solar ingress.
• MW's lead sense for saṃkramaṇa is 'going or meeting together, UNION WITH'. Union is not transition: it names arrival at a joint state, not a conditional step from one state to a next. The lead sense is thus the wrong one, and only the later 'transference to' fits.
• Modern Hindi संक्रमण predominantly means INFECTION (संक्रामक = contagious). A Hindi reader in 2026 reads 'संक्रमण-प्रदिश' as an infection-tensor before a transition-tensor. This is a case where the Sanskrit parent and the Hindi daughter have diverged and the register files must not assume the daughter inherits the parent's sense.
• Latin transitus is suspected of carrying an ecclesiastical 'passing away/death' sense (cf. the Transitus Mariae) -- which would mirror the Sanskrit death-drift exactly. THIS IS NOT VERIFIED: no L&S sense-list beyond the short definition 'a going over, passing over, passage' was read this session. Recorded as a suspicion with a falsifier, NOT as a note. Falsifier: read the full L&S entry for transitus and Du Cange for the ecclesiastical sense.
• B and D are both about hidden states and are easily merged. B is the CONDITIONAL step s -> s'; D is the prior at t=0. A register that renders both with a bare 'state-prior' word collapses the dynamics into the initial condition.

Forbidden mistranslations

register	form	why it is wrong
		Names p(s'|s), an uncontrolled Markov chain. Deletes u, and with it the entire controllability of the model. The agent could no longer act.
		REJECTED on doctrinal-freight grounds: in Samkhya, pariṇāma-vāda names the doctrine of the real transformation of prakṛti, and the word would import that metaphysics into a transition matrix. NOTE THE EVIDENCE GRADE: this freight claim was NOT verified against MW or Apte this session. It is a flagged suspicion, not a citation. Falsifier: read MW s.v. pariṇāma before relying on this rejection.
		Reads as 'infection' to a contemporary Hindi speaker. The tensor is not epidemiological.
		Violates SD-01. B is an index, not an ordinal.

Back-translation test — status: NOT YET RUN -- no translator engaged.

Give an independent translator ONLY the rendering plus 'in a decision formalism', definition withheld. Ask: 'What does this denote? What quantities does it depend on?'

C_TENSOR —

Store: TRUE · Confidence: LOW, and lowest of the five tensor entries. The Sanskrit role-gloss sits on a homonym whose alternate reading is 'sacrificed'; the Latin leans on an unlexicalised substantive; and the concept's actual differentia -- SET, not learned; over OBSERVATIONS, not states -- is carried by no word in any register and survives only in the definition. That is a real weakness of the rendering, not a formatting gap. · Reviewer: NONE - PENDING EXPERT REVIEW

The tensor encoding the prior p(o|C) over preferred observations, which -- alone among A, B, C, D and E -- is SET rather than learned by Dirichlet counts, being fixed in this corpus by the Z affect modulator.

register	form	literal gloss	status
sanskrit	C-प्रदिश (इष्टप्रत्यक्ष) (C-pradiśa (iṣṭa-pratyakṣa))	the C-tensor; role-gloss 'preferred-percept'	COINED
latin	TENSOR C	the C-tensor; role-gloss 'tensor of the chosen/preferred things'	COINED
english	C tensor (preferences)	The tensor encoding the prior p(o|C) over PREFERRED OBSERVATIONS -- what the agent expects to see if things go as it prefers -- which alone among A, B, C, D and E is SET rather than learned by Dirichlet counts, being fixed in this corpus by the Z affect modulator.	—
hindi	C-प्रदिश (अभिरुचि) ()	the C tensor (preferences)	PENDING-CITATION
spanish	tensor C (preferencias)	the C tensor (preferences)	PENDING-CITATION

Ambiguity notes

• THE HOMONYM TRAP, AND IT IS SEVERE: MW carries TWO distinct iṣṭa entries. 1. iṣṭa [169,2] from √iṣ = 'sought, wished, desired... cherished'. 2. iṣṭa [169,3] = 'p.p. fr. √yaj' = 'SACRIFICED, WORSHIPPED WITH SACRIFICES'. The two are formally identical in Devanagari and IAST. इष्टप्रत्यक्ष is therefore readable, with full lexical warrant, as 'SACRIFICED OBSERVATION' -- a sacrificial offering of the percept. This is not a strained reading; it is MW's second headword on the facing column. Any use of iṣṭa in a corpus about observations must carry this note.
• THE ASYMMETRY IS THE CONCEPT AND NO REGISTER'S WORD CARRIES IT. C is SET, never learned -- OBSERVED in the corpus (01-kitchen-rules M13: learning covers A/B/D/E, never C; grep-confirmed in this repo at cookbook/01-kitchen-rules.md lines 71-72 and 129-130). CONCEPTS.json states the reason: 'A goal that were learned from its own counts would have no purchase; C is where the goal enters from outside.' A register word implying acquisition, habituation, or learning falsifies C outright. This must be carried by the definition, because no word in any of the five registers carries it.
• C IS A PRIOR OVER PREFERRED OBSERVATIONS, NOT A GOAL-STATE. This is the second-most-common error after the learning one. The preference is over o, not over s. A register that renders C as 'goal', 'end', 'telos', 'target' asserts a preferred STATE OF THE WORLD; C ranges over what the agent expects to SEE. The distinction is the whole reason C multiplies against the A tensor's outputs rather than the B tensor's.
• Latin opto's L&S gloss is 'to choose, SELECT, prefer'. The 'choose/select' senses imply an ACT of choosing between presented alternatives. C is not an act and does not choose; it is a standing distribution, fixed before any alternative is presented. The agent's choosing happens in policy selection, scored against C -- not in C.
• Sanskrit iṣṭa 1 [MW 169,2] glosses include 'cherished' and 'desired'. 'Desired' shades toward a felt state. C asserts nothing about feeling. That this corpus's C is fixed by a modulator NAMED 'affect' makes the trap worse, not better: the name is a label on a mechanism, and no register may let it become a claim about experienced desire. TRUE and HONEST signals are sovereign; C is a TRUE-store object and may not be given a lived-experience word.
• Spanish 'preferencia' and Hindi 'अभिरुचि' both tend to imply a RANKING among options, and the Hindi word additionally shades toward 'taste, inclination, interest'. C is a distribution, not an ordering and not a taste. A distribution induces an ordering; it is not one.

Forbidden mistranslations

register	form	why it is wrong
		MW [169,3] gives iṣṭa from √yaj = 'sacrificed, worshipped with sacrifices'. The compound then means 'sacrificed observation'. The form is identical to the intended iṣṭa 1 [169,2] and the misreading is fully licensed by the lexicon. If the rendering survives review at all it must be accompanied by the disambiguation.
		Asserts a preferred state of the world. C is a prior over preferred OBSERVATIONS. Preferring to see o is not preferring that s obtain -- the two come apart exactly when the A tensor is ambiguous, which is the case the whole formalism exists to handle.
		Contradicts the corpus's OBSERVED asymmetry: learning covers A/B/D/E, never C (M13). C is set. A learned C would have no purchase, per CONCEPTS.json.
		REJECTED on doctrinal-freight grounds: icchā is a quality (guṇa) of the ātman in Vaiśeṣika and icchā-śakti is a technical term in Śaiva metaphysics; the word imports a theory of a desiring self, which C -- a TRUE-store distribution -- must not assert. NOTE THE EVIDENCE GRADE: this freight claim was NOT verified against MW or Apte this session. Flagged suspicion, not citation. Falsifier: read MW s.v. icchā.
		Violates SD-01.

Back-translation test — status: NOT YET RUN -- no translator engaged.

Give an independent translator ONLY the rendering plus 'in a decision formalism', definition withheld. Ask two questions: (1) 'What does this denote?' (2) 'Is this something the system is given, or something it acquires from experience?'

D_TENSOR —

Store: TRUE · Confidence: LOW. Two members of the proposed renderings (Sanskrit pratyaya, Latin initialis) were never looked up and are honestly marked PENDING rather than dressed as citations. The attested members are solid but each carries a live wrong sense -- pūrva 'east', status 'posture' and the sisto-homonym 'fixed'. The last is the worst: it points D at C's differentia. · Reviewer: NONE - PENDING EXPERT REVIEW

The tensor encoding the prior over hidden states at the start of a trial or episode, and one of the four tensors learned by conjugate-Dirichlet counts.

register	form	literal gloss	status
sanskrit	D-प्रदिश (पूर्वावस्थाप्रत्यय) (D-pradiśa (pūrvāvasthā-pratyaya))	the D-tensor; role-gloss 'the prior-state belief'	COINED
latin	TENSOR D	the D-tensor; role-gloss 'tensor of the initial state'	COINED
english	D tensor (initial state prior)	The tensor encoding the prior over hidden states at t=0 -- the start of a trial or episode -- and one of the four tensors (A, B, D, E) learned by conjugate-Dirichlet counts.	—
hindi	D-प्रदिश (प्रारंभिक अवस्था पूर्वप्रायिकता) ()	the D tensor (initial state prior probability)	PENDING-CITATION
spanish	tensor D (distribución a priori del estado inicial)	the D tensor (initial state prior distribution)	PENDING-CITATION

Ambiguity notes

• D IS THE PRIOR AT t=0, NOT THE RUNNING PRIOR. CONCEPTS.json states this explicitly: 'Distinct from PRIOR in general: D is specifically the prior at t=0, not the running prior.' Every register's word for 'prior' is the general word, so every register under-specifies D by default. The word 'initial' is doing all the work and must never be dropped as redundant -- without it the term names a different object that the corpus separately defines.
• SANSKRIT पूर्व READS AS 'EAST' FIRST. MW [643,1] gives 'eastern, to the east of (abl.)' as a distinct sense, and pūrva is the ordinary Sanskrit word for the eastern quarter. A reader meeting पूर्वावस्था without context may well take 'the eastern state'. This is not a fringe sense; it is standard.
• pūrva additionally carries philosophical freight as pūrva-pakṣa (the prima facie position stated in order to be refuted, in the classical debate format). A 'pūrva' view in a śāstra is one that is going to be DEMOLISHED. A prior in Bayesian inference is updated, not refuted, and a register that lets pūrva import the pakṣa sense turns D into a strawman. [Freight claim NOT verified against MW/Apte this session -- flagged suspicion. Falsifier: read MW s.v. pūrva-pakṣa.]
• LATIN 'prior' IS AN ADJECTIVE, NOT A DISTRIBUTION. L&S short def.: 'former, previous, prior, first'. It is a comparative adjective about temporal order. The statistical 'prior' is a substantivised English term of art naming a DISTRIBUTION; Latin prior names no such object and asserts only earlier-ness. Using PRIOR alone in Latin would say 'the earlier one', not 'the prior distribution'. This is why the Latin role-gloss above avoids 'prior' and says STATUS INITIALIS instead.
• Latin prior post-classically also names a monastic superior (the prior of a house). [NOT verified this session -- flagged suspicion; falsifier: check Du Cange.] If confirmed, PRIOR in a Latin register is doubly compromised.
• Latin status's L&S sense I is corporeal -- 'posture, position, attitude, station, carriage', even 'size, height, stature'. The concept needs sense II, 'condition, state'. STATUS INITIALIS is readable as 'the starting posture' (a bodily stance), which is precisely the wrong Markov blanket: WORLD/BODY/MIND are typed separately (M12), and a posture-reading drags a MIND-side object onto the BODY side.
• L&S carries a SECOND status -- the participle of sisto, 'set, fixed, appointed, regular'. This homonym is dangerous here in a specific way: it would make STATUS INITIALIS read 'fixed at the start', i.e. SET. D is LEARNED (by Dirichlet counts); C is the one that is set. The homonym thus points D at exactly C's differentia and would invert the only asymmetry the domain has.
• avasthā [MW 106,1] carries, per MW, 'appearance (in a court of justice)' and 'the female organs of generation' in specific textual contexts. Both are wholly absent from the concept and both are in the entry.

Forbidden mistranslations

register	form	why it is wrong
		L&S: 'former, previous, prior, first' -- a comparative ADJECTIVE of temporal order. Latin prior denotes no distribution. The rendering says 'the earlier tensor', which is false (D is not earlier than A) and empty (it names no object).
		Names the running prior, which the corpus distinguishes from D. D is specifically t=0. Dropping 'initial' silently merges two concepts the ledger keeps apart.
		MW [643,1] carries 'eastern' as a live sense of pūrva. Fully licensed misreading; the rendering cannot stand without disambiguation.
		Yields 'fixed at the start' = SET. D is LEARNED by Dirichlet counts; C is the set one. The misreading inverts the domain's central asymmetry, which makes it the most costly error available in this entry.
		Imports cosmogony. D is a bookkeeping distribution over where an episode might start, re-initialised every trial. It explains nothing and originates nothing.

Back-translation test — status: NOT YET RUN -- no translator engaged.

Give an independent translator ONLY the rendering plus 'in a decision formalism', definition withheld. Ask: (1) 'What does this denote?' (2) 'At what time does it apply?' (3) 'Is it set or learned?'

E_TENSOR —

Store: TRUE · Confidence: LOW. Latin consuetudo is the single best register-fit in this file for a role-gloss (sense I is exactly 'custom, habit, use, usage' and classical), but its sense II is an intimate liaison. Sanskrit is deliberately left NOT-ATTEMPTED: every candidate asserted a felt or karmic disposition that the concept expressly disclaims, and an empty field is the honest result. A future entry filling it with vāsanā would be a regression, not progress. · Reviewer: NONE - PENDING EXPERT REVIEW

The tensor encoding the prior over policies, biasing policy selection before any expected-free-energy score is computed, and one of the four tensors learned by conjugate-Dirichlet counts.

register	form	literal gloss	status
sanskrit	E-प्रदिश (E-pradiśa)	the E-tensor	COINED
latin	TENSOR E	the E-tensor; role-gloss 'tensor of habit/custom'	COINED
english	E tensor (habits / policy prior)	The tensor encoding the prior over policies -- a standing bias toward some courses of action over others, applied BEFORE any expected-free-energy score is computed -- and one of the four tensors (A, B, D, E) learned by conjugate-Dirichlet counts.	—
hindi	E-प्रदिश ()	the E tensor	PENDING-CITATION
spanish	tensor E (hábitos / distribución a priori sobre políticas)	the E tensor (habits / prior over policies)	PENDING-CITATION

Ambiguity notes

• THE CONCEPT DISCLAIMS FEELING AND THE OBVIOUS WORDS ALL ASSERT IT. CONCEPTS.json is explicit: E 'names a prior over policies and asserts nothing about a felt disposition'. Nearly every natural-language word for 'habit' in every register asserts a felt or embodied disposition. This is the entry where the register vocabulary and the concept are most directly at odds, and it is why the Sanskrit role-gloss is left NOT-ATTEMPTED rather than filled with a plausible-looking wrong word.
• 'Habit' is the corpus's own word (the Z modulator sets habits), so English gets it for free and the other registers do not. English's licence here is historical accident, not a warrant for the other four.
• SANSKRIT vāsanā IS THE TRAP THIS ENTRY EXISTS TO REFUSE. MW [947,3] gives 'the impression of anything remaining unconsciously in the mind, the present consciousness of past perceptions, knowledge derived from memory' (citing Śaṃk., Kāv., Kathās.). It is the standard word a translator reaches for, and it fails twice over: it asserts an UNCONSCIOUS MENTAL IMPRESSION (the felt disposition the concept disclaims), and in Advaita, Yoga and Buddhist usage it is bound into karma and rebirth -- vāsanās are the latent traces that propagate across lives. Rendering E as vāsanā would make the corpus assert a doctrine of karmic latency in a matrix of policy counts. FORBIDDEN.
• MW's vāsanā entry [947,3] also lists, among secondary senses, a MATHEMATICAL one: 'proof, demonstration'. A tensor named vāsanā in a mathematical corpus could be read as 'the proof tensor'. Noted for completeness; it is not the main objection.
• LATIN consuetudo's L&S sense II is 'social intercourse, companionship, familiarity, conversation' -- and in classical idiom consuetudo commonly denotes an intimate liaison. TENSOR CONSUETUDINIS is therefore readable as 'the tensor of the love-affair'. The word is retained because sense I ('a being accustomed, custom, habit, use, usage') is exactly right and classical, but the reviewer should know sense II is live.
• E APPLIES BEFORE EFE IS COMPUTED. That is its position in the loop and half its content: E biases policy selection prior to scoring, so it is not a score, not a value, and not a preference over outcomes. A register word meaning 'tendency to succeed' or 'what works' imports the score E precedes.
• E IS A PRIOR OVER POLICIES; C IS A PRIOR OVER OBSERVATIONS. Both are 'what the agent leans toward' in loose English, and loose English is how they get merged. They are different objects over different spaces, and E is learned while C is set.
• Latin habitus was considered for the role-gloss and NOT used: it carries Aristotelian hexis and, in modern usage, Bourdieu's sociological term. [Attestation and freight NOT verified against L&S this session -- flagged suspicion. Falsifier: read L&S s.v. habitus before either using or citing this rejection.]

Forbidden mistranslations

register	form	why it is wrong
		MW [947,3]: 'the impression of anything remaining unconsciously in the mind, the present consciousness of past perceptions, knowledge derived from memory'. Two disqualifications: (1) it asserts an unconscious mental impression -- a felt disposition, which CONCEPTS.json says E does NOT assert; (2) in Advaita/Yoga/Buddhist doctrine vāsanās are karmic latencies carried across rebirths, so the word imports a metaphysics of karma into a policy prior. This is the textbook case of the doctrinal-freight failure CONCEPTS.json warns of.
		Rejected for the same class of reason as vāsanā -- it names formative latent impressions in Yoga/Buddhist psychology (and, separately, a life-cycle sacrament in dharmaśāstra). It would assert both a felt latency and a ritual. NOTE THE EVIDENCE GRADE: NOT verified against MW this session; flagged suspicion, not citation. Falsifier: read MW s.v. saṃskāra.
		अभ्यास means practice/repetition/exercise and additionally carries Yoga-technical freight (abhyāsa as sustained practice, e.g. in the Yoga Sūtra pairing abhyāsa-vairāgya); it names an activity, not a prior. आदत names a personal habit of a person -- a felt disposition, which E disclaims. [Yoga freight NOT verified this session -- flagged suspicion.]
		Imports the expected-free-energy score that E explicitly PRECEDES. E biases selection before any EFE is computed; a success-word makes E a function of the very quantity it is prior to, which is circular.
		E is a prior over POLICIES and is learned; C is a prior over OBSERVATIONS and is set. Different spaces, opposite sides of the domain's one asymmetry.
		Violates SD-01.

Back-translation test — status: NOT YET RUN -- no translator engaged. Sanskrit role-gloss NOT-ATTEMPTED, so untestable by construction.

Give an independent translator ONLY the rendering plus 'in a decision formalism', definition withheld. Ask: (1) 'What does this denote?' (2) 'Does it describe something the system FEELS or is disposed to?' (3) 'Does it range over actions/plans, or over things seen?'

DIRICHLET_LEARNING —

Store: TRUE · Confidence: LOW. The Sanskrit head member gaṇanā is genuinely well-fitted ('reckoning, counting, calculation' is exactly the mechanism) and is the best-attested role-word in this file after parokṣa/pratyakṣa. But the Latin head (numeratio) was never looked up and is honestly PENDING; the Devanagari form of Dirichlet's name is unverified; and 'conjugate' -- the qualifier that carries the argument -- is unsourced in all four non-English registers. · Reviewer: NONE - PENDING EXPERT REVIEW

The exact update of the A, B, D and E tensors by counts + lr * sufficient_stat, which is conjugate and therefore requires no gradient, no autodiff, and no backpropagation.

register	form	literal gloss	status
sanskrit	दिरिक्ले-गणना (dirikle-gaṇanā)	Dirichlet-counting / Dirichlet-reckoning	COINED
latin	NUMERATIO DIRICHLETIANA	Dirichletian counting/numbering	PENDING-CITATION
english	conjugate Dirichlet learning	The exact update of the A, B, D and E tensors by counts + lr * sufficient_stat -- an update that is conjugate, and therefore closed-form, requiring no gradient, no autodiff and no backpropagation.	—
hindi	दिरिक्ले संयुग्मी अधिगम ()	conjugate Dirichlet learning	PENDING-CITATION
spanish	aprendizaje conjugado de Dirichlet	conjugate Dirichlet learning	PENDING-CITATION

Ambiguity notes

• THE WHOLE POINT IS THAT LEARNING HERE IS COUNTING, NOT DESCENDING. CONCEPTS.json: 'Learning here is counting, not descending.' The corpus enforces it mechanically -- an AST guard rejects autodiff/optax/torch/grad/backward in the loop (M13; visible in this repo at cookbook/01-kitchen-rules.md lines 71-72). Every register's ordinary word for 'learning' in an ML context now connotes gradient descent, because that is what the field does. So every register's natural word actively asserts the thing this concept denies. The registers are fighting the concept, not carrying it.
• 'Conjugate' is the load-bearing technical qualifier and is the REASON no gradient is needed: conjugacy makes the posterior update closed-form. A register that renders 'Dirichlet learning' and drops 'conjugate' has removed the explanation and left an unexplained assertion. Worse, 'conjugate' is a heavy homonym in every register -- grammatical conjugation, chemical conjugation, complex conjugates, conjugal relations -- and the Bayesian sense is the rarest of them.
• Sanskrit gaṇanā [MW 343,3] carries 'considering, supposing' and 'regarding, taking notice of' alongside 'reckoning, counting, calculation'. The concept means the arithmetic count ONLY. The 'considering/supposing' senses would reintroduce exactly the deliberative flavour the mechanism lacks: nothing considers anything; counts are incremented.
• gaṇanā also carries 'the being enumerated among' [MW 343,3] -- membership in a class rather than the act of counting. A tensor update is not a membership claim.
• The 'lr' (learning rate) in 'counts + lr * sufficient_stat' is a scaling factor on the count increment, NOT a step size along a gradient. Its name is borrowed from gradient methods and is misleading in every register. No register should render 'lr' with a word implying descent or step-size, and the safest course is to leave the formula untranslated -- the math is the fence.
• Dirichlet is a person (Peter Gustav Lejeune Dirichlet). No register may gloss the name descriptively. The Devanagari transliteration दिरिक्ले proposed here is itself unverified: the name is French-German and its standard Devanagari form, if one exists, was not checked. [Falsifier: check a Hindi mathematics text for the standard form.]

Forbidden mistranslations

register	form	why it is wrong
		Directly contradicts the definition: 'requires no gradient, no autodiff, and no backpropagation'. The corpus enforces this with an AST guard (M13). A register asserting descent describes a machine this corpus refuses to build.
		Conjugacy is WHY the update is exact and gradient-free. Without it the term states a result and deletes its reason, and the no-backprop claim becomes unmotivated assertion.
		Names repetition/practice -- an activity performed over time by a practitioner -- and carries Yoga-technical freight. A conjugate update is a single closed-form arithmetic step, not a practice. [Freight claim NOT verified this session -- flagged suspicion.]
		MW [343,3] carries those senses. They import deliberation into what is strictly increment-a-counter. Licensed misreading; needs disambiguation.
		The Bayesian sense (prior and likelihood in the same family, so the posterior stays in that family) is unrelated to all three. This is the likeliest single error in this entry and the hardest for a non-statistician translator to avoid.

Back-translation test — status: NOT YET RUN -- no translator engaged.

Give an independent translator ONLY the rendering, definition withheld. Ask: (1) 'What does this denote?' (2) 'Does it involve gradients or derivatives?' (3) 'What does conjugate mean here?'

HIDDEN_STATE —

Store: TRUE · Confidence: SPLIT BY REGISTER AFTER THE ADVERSARIAL RE-READ OF 2026-07-15; the previous single rating of MEDIUM ('the highest in this file') is WITHDRAWN as too high for the Sanskrit. LATIN: MEDIUM, and now the strongest rendering in the file outright. STATUS LATENS is well-grounded and the ground was re-checked -- L&S's headword line 'lătĕo, ŭi, 2, v. n.' was read again and the 'v. n.' (intransitive) marking is confirmed verbatim, so the exclusion of an agent is a grammatical fact about the verb rather than a preference. SANSKRIT: LOW-to-MEDIUM, downgraded. Both members remain genuinely cited and MW's own parokṣa/pratyakṣa antonym pairing is confirmed verbatim at [674,2], so the opposition is still lexicographic rather than imposed -- that much survives. But the re-read of the full p. 589 entry found (i) a fourth denied sense, 'past, completed', which the previous rating did not know about and which points at the live/current state s in exactly the wrong direction, and (ii) adverbial forms glossing 'secretly', 'behind one's back', 'without the knowledge of' -- the agentive-concealment implicature for which this very entry rejects gūḍha. The entry was rating its own word against a standard it had not applied to it. The two registers are NOT equally grounded and the previous rating implied they were: Latin excludes the concealer by grammar, Sanskrit only by sense-selection. The Advaita parokṣa-jñāna freight remains flagged and unverified, so it contributes nothing to either rating. · Reviewer: NONE - PENDING EXPERT REVIEW

The quantity s that an agent cannot observe directly and must infer, and about which the recognition density is a belief.

register	form	literal gloss	status
sanskrit	परोक्षावस्था (parokṣāvasthā)	a state beyond the range of sight	COINED
latin	STATUS LATENS	a state that lies hidden	COINED
english	hidden state	The quantity s that an agent cannot observe directly and must therefore infer, and about which the recognition density is a belief.	—
hindi	प्रच्छन्न अवस्था ()	hidden state	PENDING-CITATION
spanish	estado oculto	hidden state	PENDING-CITATION

Ambiguity notes

• THE CONCEPT'S OWN NOTE IS THE FENCE: CONCEPTS.json states "'Hidden' means not directly observable, NOT concealed, secret, or withheld by anyone." There is no agent doing the hiding. Every register has an obvious word for 'hidden' that smuggles a concealer in, and choosing between them is the entire work of this entry.
• WHY LATIN latens AND NOT occultus: lateō is marked 'v. n.' in L&S -- verb NEUTER, i.e. intransitive. It means 'to lie hid' as something the subject does, or rather undergoes, with no agent. occultus and absconditus are participles of occulō and abscondō, transitive verbs of actively hiding, and both therefore imply someone who hid the thing. The grammar of the Latin verb, not a stylistic preference, is what selects latens.
• parokṣa's OWN GLOSS CONTAINS FOUR SENSES THE CONCEPT DENIES. MW [589]: 'beyond the range of sight, invisible, ABSENT, UNKNOWN, unintelligible', and -- found only on the adversarial re-read of 2026-07-15 -- 'PAST, COMPLETED'. The hidden state is NOT absent -- it is present and doing causal work, which is why it can be inferred at all. It is not unknown in principle -- it is precisely what the recognition density is a belief ABOUT. 'Unintelligible' is worse still: an unintelligible state could not be inferred. And 'past, completed' is worse than all three: s is the CURRENT hidden state, live at time t and causing o right now; a word that reads 'the completed state' names something finished and gone, which is the one thing s must not be. The count in this note previously read 'two' in its heading and 'three' in its body, and both were wrong: the best-attested Sanskrit word carries FOUR wrong senses alongside its one right one.
• THE SECRECY OBJECTION THIS ENTRY RAISES AGAINST गूढ APPLIES TO ITS OWN CHOSEN WORD, AND THAT WAS MISSED UNTIL THE ADVERSARIAL RE-READ. This entry rejects gūḍha because MW [360,3] gives 'SECRET, PRIVATE' -- i.e. it posits a concealer, which the concept forbids. But MW's parokṣa entry [589] carries, in its own adverbial forms: parokṣeṇa 'out of sight, SECRETLY, mysteriously'; parokṣāt 'SECRETLY, without the knowledge of'; parokṣam 'out of sight, BEHIND ONE'S BACK, in the absence or WITHOUT THE KNOWLEDGE OF'; parokṣe 'BEHIND THE BACK OF'. 'Behind one's back' and 'without the knowledge of' are not merely secrecy-adjacent -- they are agentive: they presuppose a party FROM WHOM something is kept. That is the precise implicature the concept's own fence forbids ('NOT concealed, secret, or withheld by anyone') and the precise ground on which this entry rejects gūḍha. The defence available is narrow and must not be overstated: the ADJECTIVE parokṣa in its lead sense is agent-neutral ('beyond the range of sight'), and it is the adjective, not the adverbs, that enters parokṣāvasthā. But the adverbial senses are in the same entry, they are the forms a reader meets most often in texts, and they establish that Sanskrit parokṣa is NOT the clean agent-free word this entry represented it as. The Latin rendering has a real structural defence here that the Sanskrit does not: lateō's intransitivity is a GRAMMATICAL fact about the verb that excludes an agent, whereas parokṣa's agent-neutrality is only a fact about which sense one selects. STATUS LATENS is therefore better grounded than parokṣāvasthā, and this entry previously implied they were equally strong.
• parokṣa carries Advaita epistemological freight: parokṣa-jñāna (mediate/indirect knowledge) is opposed to aparokṣa-jñāna (immediate knowledge), and the pair is technical in Vedanta. Using parokṣa may import a theory of mediate cognition. [Freight claim NOT verified against MW/Apte this session -- flagged suspicion, not citation. Falsifier: read MW s.v. parokṣa-jñāna and aparokṣa.]
• avasthā [MW 106,1] carries 'appearance (in a court of justice)' and, per MW, 'the female organs of generation' in specific textual contexts. Neither is in the concept; both are in the entry.
• LATIN status's SENSE I IS BODILY. L&S I.A: 'mode or way of standing, of holding one's body (at rest), posture, position, attitude, station, carriage'; I.C 'size, height, stature'. The concept needs sense II ('condition, state'). A posture-reading is not merely odd here -- it lands the term on the wrong side of the WORLD/BODY/MIND typing (M12), where hidden states are MIND-side inferential objects and postures are BODY-side interoceptive ones.
• L&S carries a second, participial status ('set, fixed, appointed, regular', from sisto). STATUS LATENS is largely protected from this reading by latens, but the homonym is live in the bare word and matters for D_TENSOR.
• SPANISH AND LATIN DIVERGE, AS THEY DID FOR verosimilitud. Spanish 'estado oculto' is the settled term of art despite 'oculto' descending from occulō and carrying exactly the concealment implicature Latin occultus is rejected for. Usage has bleached the Spanish word; it has not bleached the Latin one. A translator may NOT reason from 'estado oculto' back to STATUS OCCULTUS -- that is the pseudopropheta-by-cognate move, the same error as verisimilitudo, and this file rejects it in both directions.

Forbidden mistranslations

register	form	why it is wrong
		MW [360,3]: 'gUQa mfn. covered, hidden, concealed, invisible, SECRET, PRIVATE, RV. &c.', with 'disguised' in a sub-entry (Manu, MBh.). Asserts secrecy and disguise -- someone has hidden something. The concept's note forbids exactly this: not concealed, secret, or withheld by anyone.
		Both are participles of TRANSITIVE verbs of hiding (occulō, abscondō) and so posit a concealer. The concept has none. Contrast lateō, marked 'v. n.' in L&S -- intransitive, no agent. [The L&S entries for occultus and absconditus were NOT read this session; the claim rests on the transitivity of their base verbs. Falsifier: read both entries.]
		REJECTED on doctrinal-freight grounds: in Samkhya, avyakta names the unmanifest prakṛti/pradhāna -- primordial undifferentiated matter -- so the word would make the hidden state an item of Samkhya cosmology. NOTE THE EVIDENCE GRADE: NOT verified against MW this session; flagged suspicion, not citation. Falsifier: read MW s.v. avyakta.
		Means secret/mystery -- and in many contexts an esoteric doctrine. Asserts both secrecy and initiatory withholding. [NOT verified against MW this session -- flagged suspicion.]
		Asserts secrecy outright. Note this is NOT an objection to the established 'estado oculto', which usage has bleached; 'secreto' has not been bleached and reintroduces the concealer.
		The state is inferable -- that is the entire premise. parokṣa's 'unknown, unintelligible' senses [MW 589] drift this way and must be fenced off. A state that could not be known could not be the object of a recognition density.

Back-translation test — status: NOT YET RUN -- no translator engaged.

Give an independent translator ONLY the rendering, definition withheld. Ask: (1) 'What does this denote?' (2) 'Is anyone hiding it?' (3) 'Can it be known?' (4) 'Is the thing it denotes happening now, or is it finished and past?'

OBSERVATION —

Store: TRUE · Confidence: LOW-to-MEDIUM. The comparative that previously stood here -- 'lower than HIDDEN_STATE' -- is WITHDRAWN: the adversarial re-read of 2026-07-15 downgraded HIDDEN_STATE's Sanskrit to LOW-to-MEDIUM as well, so the two Sanskrit renderings now sit at the same grade and this entry is no longer the lower of the pair. Attestation is not admissibility: pratyakṣa is attested exactly as claimed [MW 674,2] -- re-confirmed verbatim under audit, parenthetical included, and it remains the only fully CITED headword-in-sense in this file -- and it may still be inadmissible, because MW's own noun gloss 'ocular EVIDENCE' asserts probative force that o does not have, before the unverified Nyāya-pramāṇa objection is even reached. That gap -- CITED, re-confirmed, yet possibly forbidden -- is the most important thing in this entry, and the audit widened it rather than closing it: the evidentiary objection got STRONGER (it is now lexical, not merely doctrinal) at the same time as the citation got firmer. Attestation and admissibility moved in opposite directions on the same re-read, which is precisely why the file keeps them as two tests. The reason the domain's two best-attested Sanskrit words may both have to be re-cut together is unchanged and now better evidenced on both sides. · Reviewer: NONE - PENDING EXPERT REVIEW

The quantity o that actually arrived at the sensor, which is what a belief is scored against and what a receipt records.

register	form	literal gloss	status
sanskrit	प्रत्यक्ष (pratyakṣa)	present before the eyes; that which is present before the eyes	CITED
latin	OBSERVATUM	the thing observed; a thing that has been noted	COINED
english	observation	The quantity o that actually arrived at the sensor -- what a belief is scored against, and what a receipt records.	—
hindi	प्रेक्षण ()	observation	PENDING-CITATION
spanish	observación	observation	PENDING-CITATION

Ambiguity notes

• THE ACT/DATUM SPLIT IS THE FIRST TRAP AND IT CATCHES THREE REGISTERS. The concept is the QUANTITY o that arrived -- a datum, a thing. It is not the act of observing. Latin observatio, Spanish observación and Hindi प्रेक्षण all name the act primarily (Spanish softens this with 'accion Y EFECTO de'). This is why the Latin here is OBSERVATUM (the thing observed) and not observatio, and it is the one place in this file where the coined form is clearly better than the attested alternative.
• THE PRAMĀṆA PROBLEM, AND IT MAY BE FATAL TO THE SANSKRIT. In Nyāya, pratyakṣa is the first of the pramāṇas -- the MEANS OF VALID KNOWLEDGE. The word does not merely name a percept; in its technical use it certifies the percept as veridical. But o is 'the quantity that ACTUALLY ARRIVED at the sensor' -- arrival, not validity. A sensor can deliver a corrupt, noisy, or miscalibrated reading and it is still o; that is precisely why the A tensor is needed and why ambiguity lives in A. Rendering OBSERVATION as pratyakṣa therefore risks asserting that every observation is valid knowledge -- a metaphysical claim the corpus must not make, and one that would quietly dissolve the whole point of a likelihood. This is the exact hazard CONCEPTS.json names ('Sanskrit philosophical vocabulary carries DOCTRINAL FREIGHT... no translation choice may quietly assert a metaphysics') and the operator's charter names ('Do not assume Sanskrit automatically prevents ambiguity'). The Nyāya-pramāṇa claim is NOT verified against MW or a Nyāya source this session -- it is flagged, not cited. Falsifier: read MW s.v. pramāṇa and a Nyāya-sūtra reference on pratyakṣa as pramāṇa. If confirmed, the CITED status above stands for the word but the rendering may still have to be struck on freight grounds -- attestation and admissibility are different tests, and passing the first does not pass the second. UPDATE, ADVERSARIAL RE-READ 2026-07-15: the objection is now PARTLY GROUNDED IN MW ITSELF and no longer rests solely on the unverified Nyāya claim. The re-read of [674,2] found noun senses the entry had not recorded, among them 'OCULAR EVIDENCE, direct perception, apprehension by the senses'. 'Evidence' is an evidentiary/probative word: it says the percept ESTABLISHES something. o establishes nothing -- it is what arrived, veridical or not, and a corrupt sensor reading is still o. So the validity implicature this note feared is visible in MW's plain lexical glosses, without any appeal to Nyāya. State the evidence grades separately and do not merge them: the LEXICAL evidentiary sense is now CITED [MW 674,2]; the NYĀYA-pramāṇa doctrinal claim remains UNVERIFIED and flagged. The first is sufficient on its own to keep the rendering under examination. The re-read also found 'superintendence of, care for' -- a supervisory sense wholly absent from the concept; o supervises nothing.
• pratyakṣa's LEXICAL SENSE IS OCULAR ('present before the EYES', MW 674,2; the word is literally prati + akṣa, 'before the eye'). o is whatever arrived at ANY sensor -- a thermometer reading, a byte count, an interoceptive hardware signal (M12). The register word is narrower than the concept in modality and would have to be stretched.
• THE BRIDGE NOTE, WHICH IS THIS CONCEPT'S REASON FOR EXISTING: CONCEPTS.json says o in an equation and a receipt in a ledger are 'the same kind of object, which is why a confabulated observation corrupts the model rather than merely embarrassing the author'. Any register that renders o as a mere 'impression', 'appearance', or 'sense-datum' -- something mental and private -- severs that bridge. The word must be able to denote a LOGGED, EXTERNAL, CHECKABLE thing, because a receipt is one.
• Latin observō carries 'to watch FOR, seek to catch' [L&S] -- purposive, anticipatory watching. o is passive arrival; nothing is sought. The verb also carries the 'keep/observe a rite or law' sense that makes observabilis so treacherous in POMDP.
• OBSERVATION is the counterpart of HIDDEN_STATE and the lexicon already knows it: MW defines praty-akṣa expressly as 'opp. to parokṣa'. Whatever this entry finally settles on, it must remain the antonym of whatever HIDDEN_STATE settles on. If pratyakṣa is struck on pramāṇa grounds, parokṣāvasthā loses its lexicographic anchor and BOTH entries must be re-cut together.

Forbidden mistranslations

register	form	why it is wrong
		L&S gives 'an observing, watching' -- the ACT. The concept is the DATUM that arrived. Rendering the act for the thing makes o a process rather than a quantity, and a process cannot be what a belief is scored against.
		Names a MEANS OF VALID KNOWLEDGE. It asserts validity, which o does not have and does not need -- o is what arrived, veridical or not. This is the sharpest form of the pramāṇa problem and the reason it must be fenced even for pratyakṣa.
		Names sight/the faculty of seeing (and, in Buddhist usage, a VIEW in the doctrinal sense -- micchā-dṛṣṭi, 'wrong view'). Names the faculty or a doctrine, not the datum. [NOT verified against MW this session -- flagged suspicion.]
		Makes o mental and private. o must be able to be a LOGGED external quantity -- a receipt -- because CONCEPTS.json makes the equation's o and the ledger's receipt the same kind of object. A private impression cannot be a receipt.
		o is what ARRIVED, not what is the case. A corrupt sensor reading is still o. Any validity-word collapses the distinction between the observation and the hidden state it is evidence about -- which is the distinction the entire domain is built on.

Back-translation test — status: NOT YET RUN -- no translator engaged.

Give an independent translator ONLY the rendering, definition withheld. Ask: (1) 'Does this denote an act or a thing?' (2) 'Does it imply the thing is true or valid?' (3) 'Could this word denote a number written into a log by a machine?'