Inherited claims bypass the epistemic gates: label every claim in context files and papers proved/measured/open/wager

[emooreatx]

The gap

CIRISOntology/epistemology.md and GATES.md encode the discipline, and it works — applied to CCA v3 it found 18 defects (#11) on material that had already survived authoring and an external review. But the discipline is being applied to new claims and not to inherited ones, and that asymmetry is a stable failure mode.

Concrete evidence from the 2026-07-31 audit:

1. CLAUDE.md propagated falsified claims into every session in this repo for months. It asserted the GPU array provides "physical validation of CCA theory," citing r = 1.000 (Exp 86) and R² = 0.798 (C1) — both now known to be identity checks that compute k_eff from the formula and compare it against the same formula. It is loaded as context on every task, so it regenerates the claims faster than they can be corrected downstream. (Corrected 2026-07-31.)

2. Corridor Dynamics v6 (coherence-ratchet) credits CCA v3 with establishing the very discipline CCA v3 lacked. Line 226: "inherits the methodology infrastructure established during CCA v3 (pre-registration discipline, matched-control nulls, ...)". CCA v3 had no pre-registration — that is defect C-4/C-9, where the outcome definition was revised after results were seen — and no matched-control nulls. The attribution is backwards: those disciplines were derived from CCA's audited failures.

The pattern: rigor at the point of creation, credulity at the point of citation. A claim that has been restated once acquires the status of settled background.

Proposed

Label discipline. Every claim in a context file (CLAUDE.md), README, or paper carries one of the four strengths from epistemology.md §1 — proved / measured / open / wager — never rounded up. "Validated" is not a label. For measured, the stated domain is mandatory (usually: one RTX 4090).

Unlabelled defaults to wager. Anything inherited without a label is treated as a wager until traced to the artifact that produced it.

Check the artifact, not the prose. This is epistemology.md's own lesson about mechanization ("check the artifact, not the text describing the artifact") applied to citation. Both C-11 (identity check) and C-12 (a print() of the word VALIDATED nine times, with no computation) pass any amount of careful reading and die instantly on tracing to source. Before citing a number, open the code that produced it.

Periodic inherited-claim audit. New-claim gates do not catch claims that entered before the gates existed. The corpus predating CIRISOntology has never been swept.

Candidate mechanization

• A checker that flags "validat*", "confirm*", "proves" in context files and papers without an adjacent four-strength label
• For claims labelled measured, require a pointer to the producing artifact (file path or results JSON), and verify the path exists
• Note the honest limit, per epistemology.md §4: a machine can check that a label and a pointer exist, never that the label is the honest one or that the artifact says what the headline says. This makes dishonesty a deliberate act rather than a lapse, which is the most tooling can do

Scope

RATCHET CLAUDE.md corrected 2026-07-31. Corridor Dynamics lines 226/228/780 need correcting in coherence-ratchet (separate repo — filing there separately, or fold into this issue if preferred). The wider sweep of pre-CIRISOntology material is unstarted.

Refs

#11, #12, #13, CCA_PAPER/CORRECTIONS_v3.md, CIRISOntology/epistemology.md §1 §4, CIRISOntology/GATES.md.