E0 - How to read this work (the Evidence Constitution)
How to read this page
Three ways to read this page. Precise is the document itself, exactly as it is written in the repository. Plain and Clear were written for this website to help you meet that document — they are about it. They are not it, and they are not evidence.
The Encyclopedia is the UNI method written out as a reference work: 39 pages, arranged in wings, setting out what the programme is attempting and why it is built the way it is. This is where the ideas are explained in order and in prose, rather than as code, as runbooks, or as dated receipts.
Every chapter is authored against two ledgers and never ahead of them. One records what UNI has built, and the evidence class of each claim. The other records nature's own regularities, kept separate on purpose. That way a fact about biology is never quietly reused as a fact about the software. Where a chapter and a ledger disagree, the chapter is the thing that is wrong. Every chapter closes with an invitation to falsify it, and a recorded negative is published beside the result it qualifies rather than after it.
Read "How to read this work" first. It is the evidence constitution: the classes, the four ledger states, and the rule that a finished chapter is not the same as a working system. Then the calibration ledger, which carries the figures every other chapter is required to use.
What it is not: a description of a person or of a mind. The programme calls itself a developmental active-inference simulation, a bounded peek into a toy world, and its own index prints how much of the developmental ladder has actually been earned — roughly two rungs out of eleven or more. It is also not a report of what is running today. For what ran, and when, go to the evidence record.
Your browser cannot switch reading levels, so the document itself is shown.
Precise — the source document
This is the document. Rendered from the repository at the commit above, with nothing rewritten for the web. A gate re-renders it on every deploy and fails the build if a single byte differs.
This is the onboarding chapter, and it is the one chapter that asserts no capability at all. It describes how the UNI Encyclopedia is governed: how every claim in the work is graded, what would prove each claim false, and which rules the authors are forbidden from breaking. Read it as a contract, not as a result. The UNI program itself is a developmental active-inference SIMULATION: a no-backprop, nested-Markov-blanket toy that grows from a modeled zygote toward a modeled speaking three-year-old. It is never a person, never a mind, never alive. The honest position of the whole program, printed here so it is the first number you see and not the last, is this: roughly 2 of 11+ developmental rungs are earned. Everything else is parked, negative, pending, or fenced.
A constitution is not evidence. Having a disciplined ledger does not mean the science is proven. It means that when something IS proven, you will be able to tell, because the proof will sit at a stated evidence class, next to a stated falsifier, beside the negatives that travel with it. The rules below are the reusable asset; the capability claims live in the other wings, and none of them is raised by anything written here.
The six rules of the constitution (method M1)
The Evidence Constitution (ledger row M1, class method) is built from six rules, carried verbatim from the source ledger:
- A falsifier per claim. No claim ships without a stated condition that would prove it false. A claim with no falsifier is not a claim; it is marketing, and it is forbidden.
- The append-only ledger is the single source of truth. There are four states only: PASS, FAIL, NEGATIVE, PENDING. Corrections are forward-only (a row is superseded with lineage, never silently edited). Prose cites the ledger; where prose and ledger disagree, the ledger wins and the prose is wrong.
- Calibration only moves DOWN. Wording is calibrated down to the measured value, never up, including under urgency. The fence gets louder under pressure, not wider.
- The verdict is the CI bound that excludes the threshold, never the point estimate (ledger row M2).
- DONE means test-covered (Class E/D), not feature-working (Class A). A passing test never satisfies a criterion that demands a runtime observation.
- Negatives are content. A partial, a negative, or a "most pieces do not help" decomposition is a measurement that the design is incomplete, not an exit and not a failure to hide. The negatives are the credibility.
The falsifier for M1 itself is operable and is the load-bearing test of this whole work: a claim asserted ahead of its ledger, wording calibrated UP rather than down, or a verdict edited rather than superseded. Any of those is a constitutional breach, not a stylistic lapse.
The evidence classes (method M30, taxonomy A through U)
Every claim is tagged with an evidence class, and the class is a ceiling: a chapter may card a claim at or below its class, never above. The A-through-F provenance subset (ledger row M30, class method) is the working core: A = live / observed at runtime; C = code or static inspection or, in the fuller A-U taxonomy, a pre-registered held-out dev-gate; E = test-passes; F = doc or prior-claim, inheritable but which MUST be re-verified before it is leaned on. The full taxonomy adds B (mechanism plus operator observation), U (claimed-but-unproven), and method (a governance pattern, not an empirical claim). "Class U - not claimed" is itself a standing fence: Class-U content is described as not-yet-built or parked, never as capability.
Two worked examples make the ceiling concrete. A Class-E example: a passing test suite documents that code behaves as written, but it cannot stand in for a Class-A runtime observation; in a sibling delivery system an audit found a handoff route that was unit-tested (Class E) yet, on querying the running system, had fired zero times at runtime (the Class-A check), so the route "existed" without ever having "happened." A Class-A example: an anchor observed at runtime is the strongest class, and still its interpretation is fenced, because an anchor is not a capability. The exactness honesty rule (ledger row M10) is the standing trap here: the JAX core runs float32, so its anchors hold only to about 6e-8 (about 1e-6 for a single-step filter), and the genuine sub-1e-10 tier lives only in the NumPy / Rust-f64 path. A genome docstring that claimed sub-1e-10 was caught as an overclaim and corrected. Never card a float32 anchor at the f64 tier.
Class authority, when sources disagree (method M9)
When two sources conflict, the constitution resolves them by class, not by confidence (ledger row M9, class method). Class-B (tool state) overrides Class-G (the program's own narrative); Class-A (observed at runtime) overrides Class-E (a test passes). A test passing does NOT satisfy a criterion that demanded a Class-A deployment-time check. The discipline is to mark [CONFLICT:unresolved] and resolve it with a direct read of the live system, never to let the more flattering source win. The falsifier is a criterion marked satisfied by a Class-E test where it demanded a Class-A observation, or the narrative trusted over the tool.
Bars before build, held once (method M2 and M3)
The empirical PASS tier is governed by two more method rows. Bars-before-build, held-once (ledger row M2): pre-register the bar (the margin against the threshold) and a named ablation before measuring; touch the held set exactly ONCE, behind an atomic seal-before-scoring and a once-only sentinel; and read the verdict off the CI bound, not the point estimate. Its falsifier: a held set touched more than once, a verdict read off the point estimate, or a build started before its bar was registered. Validator-derived reproduced:true (ledger row M3): a reproduction flag must be derived by the validator from at least 5 distinct seeds and a real, non-degenerate confidence interval that contains the value, never written in as a hardcoded literal. Its falsifier: a reproduced:true emitted as a literal rather than derived. This rule has teeth precisely because it was violated once and caught: an audit found reproduced:true had been a literal, and the fix (the audit's central correction) was to make every such flag validator-derived.
On the negatives count (a provenance fence)
The constitution treats negatives as the credibility of the work, but it also fences the headline number against careless re-use. The recorded ledger snapshot is 882 rows = 350 PASS / 0 FAIL / 183 NEGATIVE / 349 PENDING (Constitution §0). That figure is a recorded snapshot, not a count you can reconstruct from this published work: the ledger derives from a deduplicated merge of 615 extracted claims, the snapshot enumerates 882 rows, and the carded body shows only about 140 distinct rows. A skeptic counting the published rows cannot independently derive 183. So the rule is to cite the snapshot AS a snapshot, provenance-flagged, and to headline only the negatives the ledger body actually enumerates, never the bare "183 published negatives" as a credibility number until that count is reconstructable. That is the calibration-down rule applied to the constitution's own favorite statistic.
What is NOT claimed in E0 (the Evidence Constitution)
- Ceiling: that "we have a rigorous constitution" is NOT the same as "we have proven the science," and nothing here shows any capability. The single strongest thing a careless reader might infer, that a disciplined ledger is itself evidence UNI works, is NOT shown. The most we claim is the exact, in-class statement: M1, M2, M3, M9, and M30 are reusable governance patterns at class
method(proven and reusable as disciplines, never raisable into a capability claim), and the honest program position is roughly 2 of 11+ developmental rungs earned. - Fences engaged: red line 7 (never raise a claim above its source evidence class); the standing "Class U - not claimed" fence; the developmental-SIMULATION framing (red line under §0); the no-PII / no-patent-math rule (red line 10); the exactness-tier honesty fence M10 (never card a float32 anchor at the f64 tier); and the provenance fence on the 183-negative snapshot. By construction this is a method chapter and engages no capability red line, because it makes no capability claim.
- Negatives that travel with this claim (cite alongside, never strip): the provenance flag on the 882/350/0/183/349 snapshot (183 is not re-derivable from the carded rows); the recorded fact that
reproduced:truewas once a hardcoded literal and had to be corrected (the live evidence that M3 is needed, not decorative); and the float32-overclaim correction that grounds M10. These are not asides; they are why the constitution exists. - Parked / owed: nothing is parked by this chapter and no Class-A observation is owed by it, because it asserts no empirical result. It does, however, carry the program's standing owed items by reference: the parked frontiers and the owed sign-to-park belong to the capability chapters (Wing S), not here.
- One-line honest summary a skeptic could not dispute: this chapter proves only that a falsifiable, class-graded, append-only, calibrate-down discipline exists and is documented; it proves nothing about whether UNI works.
Falsify this
The lead, operable falsifier for this chapter is the constitution's own first rule turned on the encyclopedia: find one published claim in this work that is stated above its ledger evidence class, that carries no falsifier, that has been calibrated UP rather than down, or whose verdict was silently edited rather than superseded with lineage, or find one reproduced:true that is a hardcoded literal rather than validator-derived from at least 5 seeds and a real non-degenerate CI. Any single instance falsifies the claim that this work is governed by its own constitution.
Sources
Ledger: CLAIM-LEDGER.md Section 0 (Constitution and Standing Fences) and method rows M1, M2, M3, M9, M30 (with M10 and the §0 snapshot cited inline). Authoring spec: MASTER-PLAN.md PART I (FM-1 Evidence Constitution, FM-2 the A-U rubric, FM-3 standing fences and red lines, FM-4 the not-claimed template) and the E0 entry in Wing E. Narrative grounding (PII-redacted): curated/uni-gpt-digest.md (the constitution as the reusable asset; the validator-derived reproduced:true fix; the float32 exactness correction), curated/ideation-explorer-digest.md (the DD-TDD evidence contract; the A-F class taxonomy with per-criterion verdicts; the Class-E-vs-Class-A handoff-never-fired finding), reconciled via curated/00-INDEX.md. Archive pointers (local-only, never published): ...-UNI-GPT, ...-SolutionWright-IdeationExplorer.
sha256 0adaee9f4a050900 — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
Everything else in the encyclopedia is governed by the rules set out here, and this is the one chapter claiming no capability at all. It describes how each claim is graded, what would show it false, and which rules the authors may not break. Read it as a contract, not a result. The program behind it is a developmental simulation that grows from a modelled zygote toward a modelled speaking child, and it is never a person and never a mind. The chapter is blunt that a disciplined ledger, a list of claims added to and never edited, is not the same thing as a correct science. What a good ledger buys you is narrower. When something is shown, you will be able to tell, because it will sit at a stated evidence class beside a stated result that would show it wrong, and beside the negatives that travel with it.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 0adaee9f4a050900
Clear — written for this website, not the source document
The evidence constitution is set out here, and the chapter then insists on its own limits. It asserts no capability. What sits underneath it is a developmental simulation, so nothing in the constitution is offered as a scientific result.
Six rules carry the constitution. Every claim ships with a stated condition that would show it false; a claim without one is marketing and is forbidden. One ledger is the source of truth: rows are added and never edited, there are four states only, and a correction supersedes a row with its lineage attached. Calibration only moves wording down toward the measured value, including under pressure. A verdict is the interval bound that excludes the threshold, not the point estimate. Done means test-covered, which is not the same as working. And negative results are content, not something to hide.
The chapter then explains the class ladder. Classes run from a live runtime observation, through code inspection or a gate registered in advance over data set aside and scored once, then through a passing test suite. Below those sits an inherited document that must be re-checked before it is leaned on, and outside them sit a class meaning claimed but unearned and a class for governance patterns. The class is a ceiling: a chapter may record a claim at or below it, never above. Two worked examples make the ceiling concrete, including a case where a route was covered by tests yet had never fired in the running system.
Further sections cover how conflicting sources are resolved by class rather than by confidence. They also cover the discipline of registering a bar before building, touching a set held back for scoring exactly once, and deriving a reproduction flag from multiple seeds rather than writing it in by hand. That rule exists because it was once violated and caught.
A short section puts a limit around the corpus's own favourite statistic. The recorded count of published negatives is a snapshot, not a figure a reader can rebuild from the rows on show, so it must be cited as a snapshot.
The chapter closes with what it does not claim, and with a test a reader can run. Find one published claim stated above its ledger class, or carrying no stated result that would show it wrong, or calibrated upward, or with a verdict quietly edited. If you can, the chapter's own claim fails.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 0adaee9f4a050900