History / Library Of Compound Cures

Two entries filed, and the word SEAL taken back off a source digest A six-reader audit of all three Libraries against every sealed artefact on the wiki, then a filing decision per artefact, then an adversarial pass told to REFUTE each gap. 37 pages read, 22 sealed-eligible. It found real gaps and it also produced a false result of its own, and both are here. FILED — two artefacts that were carried by no entry and named by no slot: library/compounds/crispr-clinical-guide-atlas.md seal 61ef3254, HELD library/materials/fatigue-yield-vs-regulator.md seal d5fab652, ADMITTED The guide atlas is HELD in a clean clone because its program refuses without the 3 GB assembly — but its 3,022-line sealed transcript is SHIPPED in this repository, digest-pinned beside a deliberately truncated control that REFUSES, so its seal can be checked without the download. That is recorded in the entry. THE WORD SEAL, TAKEN BACK. A-Drug-An-AI-Designed published a block headed "Markers and seal, recomputed by the programs on every run" whose last value was labelled SEAL. Measured 2026-09-09: it is the sha256 of five SOURCE FILES concatenated, and none of rentosertib-structure-lock-exact, tnik-network-degree-exact or modality-register-exact accumulates a transcript or digests its own results — zero transcript sites, zero digest sites in all three. A seal in this programme is over the MEASUREMENT, never over the instrument. The block is relabelled SOURCE DIGESTS and SOURCE-SET DIGEST, and it now states TRANSCRIPT SEAL NONE. That also makes the rentosertib OPEN_SLOT earned rather than incidental: clause E5 requires a seal the named program prints, none of the three prints one, so an entry cannot be admitted on the evidence that exists. The manifest slot now names both blockers and the cheap condition that fills it — give those three programs a transcript seal. E12 REFUSED MY OWN ENTRY, AND THE CLAUSE WAS NOT WIDENED. The fatigue entry's MEASURED line read "all three pairwise CONFIDENCE intervals overlap" — the statistical sense, from a published study, not a generator scoring itself. A false positive. The entry lost nothing by saying "all three pairwise intervals overlap" and printing 3.4 to 19.7 beside it, and softening a detector that fires in order to admit one's own entry is how a gate becomes decoration. Recorded in the E12 section rather than filed off. WHAT THE AUDIT GOT WRONG, because it is the same defect it was built to find. Four of its eight "confirmed unfiled" pages carry ZERO seals and no marker — N-of-1 antisense, PM359, Study 33 and Study 31. They are narrative and charter pages that cite other programs' work; there is nothing on them to file. The verify prompt asked "is this UNFILED?", which is trivially true of a page with no seal, so the adversarial pass confirmed the question it was asked instead of the one that mattered — a true fact about a different question. The verdict count below is stated on the four that survived that check, not on eight. Libraries: 20 entries, 15 admitted, 5 held, 0 refused, census 89, seal fa0c43d8. All three ADMITTED, L9 exact in each: proteins 6, compounds 8, materials 6. Index surgery, nine paths, zero deletions, verified with git diff --stat. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

rg78803 committed Sep 9, 2026

The contract is real: the admission law reads structure, not spellings Six hours ago The-Library-Admission-Law named its own weakness as open work — "the refusal detector is keyed to SPELLINGS ... the durable fix is a contract: one declared line every refusing program prints, and a delimited block around quoted reference figures". Another session building two unrelated studies read that, implemented it in both of its programs without being asked, and said so. This is the law reading it. RUN_TERMINAL COMPLETE this run computed a verdict RUN_TERMINAL REFUSED <reason> this run computed nothing, and says why --- BEGIN QUOTED REFERENCE FIGURES (published; NOT computed on this run) --- --- END QUOTED REFERENCE FIGURES --- TWO PROPERTIES. A declared terminal is final in BOTH directions: REFUSED is a refusal whatever the prose says, and COMPLETE is NOT a refusal whatever the prose says. The second half is load-bearing — a completed run whose quoted block carries the sentence "a gate given nothing must not pass" would otherwise be read as a refusal by the fallback and HELD, and a false NOT_KNOWN is a defect exactly as a false ADMIT is. And the half a refusal test cannot reach at all: a figure between the fences was quoted, not computed, EVEN ON A COMPLETE RUN. E4 and E5 now read the computed region — the transcript less every fenced block — independently of each other, so a run can have its figures accepted and its seal refused. An unclosed BEGIN fence is treated as quoted to the end of the transcript, the direction that can only withhold credit, never manufacture it. The spelling list survives as a fallback for programs that have not adopted the contract. A vocabulary can only ever be as complete as yesterday. SIX NEW ARMS, three of them cutting against the rule: a COMPLETE run that also quotes its block must ADMIT (or the computed-region rule is an always-refuse); figures only inside the fences on a COMPLETE run must be REFUSED, not held, because the run finished and does not print them; a declared refusal quoting every figure must HOLD; a declared COMPLETE must beat a refusal spelling in its own quoted block; an unclosed fence must withhold; and figures computed with the seal only quoted must pass E4 and refuse E5. 71 arms, all passing: 43 refuse, 19 admit, 9 hold. Verdicts unchanged — 18 entries, 14 admitted, 4 held, 0 refused, census 89, seal da85dccf. A strengthening that moves no verdict is what a correct repair looks like. Built with index surgery against a private GIT_INDEX_FILE, seven paths, verified with `git diff --stat` before writing. My working copy of _Sidebar.md was STALE and would have dropped the Study 41 line and reverted the Study 40 wording, so that file was rebuilt from the live remote with only the arm count changed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

rg78803 committed Sep 8, 2026

The library reads in order now, and a quoted figure stopped counting as evidence Cures first, as a five-step reading order. Twelve unreachable pages brought back into the sidebar. The front door carries seven medicines and both registry-wide maps. The registry-specificity slot named empty on 2026-09-07 is entry 7, seal 2d74f5c6. 39 new pins on The-Order-Of-The-Bases, checkable with no corpus. And the defect those pins exposed: the admission law's E4 and E5 were crediting figures a run had QUOTED rather than computed, so a screen that ran nothing was graded MEASURED against its own published seal. Both clauses test for a refusal first now; three arms hold the repair, including the reverse control. 65/65 arms. Libraries: 18 entries, 14 admitted, 4 held, 0 refused, seal b9b2ac76. This clone now carries library/ and the admission law, so the wiki can grade its own libraries — it never could. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

rg78803 committed Sep 8, 2026

The eleven pairs become re-derivable, and three libraries open under a law that refused us first Proteins, Compound Cures, Material Systems, under an admission law with 42 control arms — 30 that must refuse, 9 that must admit, 3 that must hold NOT_KNOWN. It refused our own headline result first: the eleven drug pairs appeared in exactly one place in the whole repository, the prose of the study page. A program now prints them, and one changed digit kills the table and the seal. 17 entries graded, 16 admitted, 0 refused, 1 held and named. harness 260 passed / 0 failed Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

rg78803 committed Sep 7, 2026