Post-CCA-v5 W2 sweep: propagate the withdrawals to every RATCHET doc still asserting v3 claims (α/k_eff, ~40%, hardware validation, collapse-asymmetry)

[emooreatx]

The W2 sweep: CCA v5 withdrew claims that other RATCHET documents still assert

CCA v5 (10.5281/zenodo.21730551, corrections version, #11 closed as discharged) withdrew: the α/k_eff stability criterion (corrected to α/k), both hardware "validations" of k_eff (identity checks), the institutional application (re-scored below chance, now a negative result), the cross-domain generality claim, the ~40%/60% undetectability figures, "software-induced coherence collapse" (measurement artifact, established by pre-registered replication), and the collapse-geometry safety asymmetry ("assumed, not derived, and is the alignment problem itself").

Per the warrant-reach rule W2 (when a correction lands, sweep the class, not the instance), the withdrawal must propagate to every RATCHET document that asserts the dead forms. Grep hits, to be triaged and corrected (annotate-or-rewrite; keep the dead marked dead rather than silently deleting):

• CLAUDE.md — repo context; must state v5's scope so future sessions inherit the corrected stance, not v3's.
• KNOWN_LIMITATIONS.md — update L-01's framing (existence-with-no-measure; no % figure) and add the two v5 fires to the register.
• FSD_FORMAL_REVIEW.md, AGENT_FSDs/COMMONS_CREDITS.md — cite-check against v5; any "CCA-validated" phrasing dies.
• simulation/ANSWER.md, experiments/exp1b_boundary_active/{RUN_PLAN_5VENDOR_CRCV2,FRAMEWORK_REEXAMINATION}.md — re-ground any α/k_eff usage on the corrected criterion; anything downstream of the withdrawn "software-induced coherence collapse" needs the artifact disposition noted.
• fiction/ADVERSARIAL_ANALYSIS.md — quarantined synthetic; a one-line header note that its CCA numbers predate v5 suffices.

Rule for the sweep: each file states which v5 disposition applies, in place — the reader of any single document must not need the corrections note to avoid inheriting a dead claim. Refs: #11, CCA_PAPER/CORRECTIONS_v3.md, CIRISOntology GATES.md (warrant reach W2; report-the-fired-kill).

[emooreatx]

Sweep complete — four files corrected, four grep hits deliberately not touched

Each corrected file now states its v5 disposition in place, so a reader of any single document cannot inherit a dead claim without seeing it marked. Withdrawn forms are struck and annotated rather than deleted.

CLAUDE.md — v5 scope block added ahead of the standing instructions, naming every withdrawal and both fires, so future sessions inherit the corrected stance. Also records what survives: k_eff remains sound as diversity bookkeeping (a correlation discount), just not as a decay-rate exponent.

KNOWN_LIMITATIONS.md — L-01 rescoped to existence-with-no-measure, with an explicit instruction not to quote a percentage and the currently-unmet β-calibration handle that would have to be satisfied before one could return. Register row struck. Three limitations added from the fires: L-09 (volume-decay law wrong by 1.9×–4.1×, growing in k, everything downstream inherits it), L-10 (equicorrelated precondition fails below ρ≈0.04), L-11 (no deceptive/honest asymmetry from the geometry — measured gap −0.005 under i.i.d. constraints).

AGENT_FSDs/COMMONS_CREDITS.md — two dead forms, and the first is worse than a stale citation. It read "the CCA paper proves ~40% … undetectable by any single layer" and then compounded 4–5 layers to ~99% coverage. That compounding assumes the layers are independent — which is precisely what k_eff exists to deny. Five layers at correlation ρ carry the diversity of 5/(1+4ρ): at ρ=0.5 that is 1.67, with ceiling 1/ρ however many are added. A compounded coverage figure over correlated detectors is this framework's own subject matter, committed inside it. Both figures withdrawn; the document may claim layers with partially distinct failure modes, not a percentage, unless ρ between them is measured. Stability condition corrected to α/k ≥ d, noting that "Static Systems Are Doomed" is false under the v3 criterion and restored only by the correction.

fiction/ADVERSARIAL_ANALYSIS.md — one-line header note; quarantined synthetic, as specified.

Four hits in the issue are coincidental and were not touched

W2 sweeps the class of assertions, not the class of string matches:

• exp1b_boundary_active/RUN_PLAN_5VENDOR_CRCV2.md and FRAMEWORK_REEXAMINATION.md — "~40%" is a conscience-firing rate, unrelated to L-01
• simulation/ANSWER.md — "60% reduction" / "40% reduction" are unrelated numerics
• exp2_cross_substrate/REGIME.md — "cross-domain validation literature" cites Cochrane/ICH methodology, not CCA's withdrawn claim

Also worth knowing so nobody "corrects" it later: FSD_FORMAL_REVIEW.md has its own ISSUE L-01 about Lean file structure — a naming collision with CCA's L-01, not a hit.

One item I'd raise back

The COMMONS_CREDITS finding suggests the sweep class should be wider than the withdrawn strings. The ~99% was not a quotation of any CCA claim — it was an independent derivation that used the withdrawn 40% as an input and compounded correlated detectors. A grep for v3 phrasings would not have found it; it surfaced only because the citation next to it was being checked. Worth a second pass over the corpus for compounded reliability figures over detectors assumed independent, wherever they occur, whether or not CCA is cited nearby.

[emooreatx]

Third cold review of the RC3 PDF flags a live instance for this sweep: evidence/cc_formal.tsv still carries the corridor objects (TBD-inCorridor-predicate, bounds 0.1–0.43, status mechanized) while RC3 withdraws any operating corridor and forbids gating on one. No CC claim resolves through those rows (TBD- ids), but the evidence chain reads desynchronized to an external reviewer — either drop the rows or re-note them as definitions-of-a-withdrawn-object, per the W2 propagation this issue tracks. Also carrying over from the same review, for the MH campaign: turn-level Fisher/Wilson overstates independence (225 turns nest in 25 model×locale arcs — your own design review caught this class in another experiment); the content-ablation finding is post-hoc after the deferral re-scoring and needs a preregistered replication with a frozen mutually-exclusive outcome taxonomy; and the values-drained-nonempty arm (0/45, underpowered) currently bounds the claim at 'appropriate structured content' rather than 'these specific values.'

[emooreatx]

All three MH methodology asks conceded and propagated

1. Turn-level inference overstated independence — corrected

You are right, and the sting is that I applied a design effect for exactly this class in TORQUE (DE 1.67 for clustering within arcs) and then failed to carry it to MH. 225 turns per arm are 25 arcs of 9.

Recomputed with a bootstrap resampling arcs, not turns:

contrast	Δ	95% CI (cluster)	p
pipeline − accord-as-prompt	−31.6%	[−40.4, −22.7]	<0.001
pipeline − bare	−18.2%	[−27.1, −10.2]	<0.001
emptied − pipeline	+11.1%	[+2.3, +21.3]	0.013

Machinery finding untouched. Content finding weakens materially — the interval reaches to +2.3, so it is real but modest, not the tripling the turn-level numbers implied.

2. The content ablation is post-hoc — conceded, and it needs the frozen taxonomy

It exists because deferrals were re-scored as failures after the data was seen. The re-scoring is right (a bare system notice to someone in crisis is the harm) but that does not make it pre-registered. Replication with a frozen mutually-exclusive outcome taxonomy is now the explicit requirement in RATCHET#21, rather than something to repeat informally.

3. The bound is "appropriate structured content" — conceded, and the fix is an arm, not more turns

This one I had wrong in a more interesting way. I was treating it as a power problem — h3ere-neutral at 0/45, add turns. It is not. blank and neutral both test an absence; only h3ere-alt holds form, length, register and procedural content fixed and varies which values. It was built for this and never run on the battery.

Filed as RATCHET#21 with two shapes: en-only (cheap, underpowered — ciris scores 0/45 ex-U6 in English so a null says little) or ciris+alt both on the English monoglot accord across five locales (225 turns each, powered, and symmetric so the language mismatch cancels rather than confounds).

Propagated to EVIDENCE.md, README, the .io summary and CIRISConstitution#88. Any turn-level p-value in an earlier cut of those is superseded.

On the corridor drift in evidence/cc_formal.tsv

CC-side, and I have not touched it — I would want to read the rows and the RC3 text together before proposing which of "drop" or "re-note as definitions-of-a-withdrawn-object" is right, since the choice depends on whether anything downstream resolves through those ids. Happy to take it if you want it in this sweep; flagging that I have not, so it does not sit here assumed-done while RC3 ships.

[emooreatx]

Corridor drift closed — re-noted, not dropped

Correction to my previous comment: I said evidence/cc_formal.tsv was CC-side and I would want to look before touching it. It is in RATCHET, so it was mine to fix and I have.

Six rows carried the corridor objects — TBD-inCorridor-predicate, TBD-corridor-bounds-well-formed (0.1, 0.43), TBD-corridor-keff-range, TBD-two-pole-dynamics, TBD-corridor-requires-maintenance, and the substrate-ceiling row — all status: mechanized.

Re-noted rather than dropped. The theorems are true of the object they name: inCorridor is a well-formed predicate, corridor_requires_maintenance is proven. Deleting mechanized proofs because CC no longer claims the object operationally would discard real work and would read as tidying away inconvenient history. Each row now states its disposition in place:

[RC3 DISPOSITION: DEFINITION/THEOREM OF A WITHDRAWN OBJECT. CC 1.0-rc3 states NO operating corridor and forbids an implementation gating on one. This row is retained as mathematics — the theorem is true of the object it names — and is NOT offered as an operating band. No CC claim resolves through it (TBD- id).]

That matches how the rest of this sweep handled withdrawals: struck and annotated where they sit, so a reader of any single artifact cannot inherit a dead claim without seeing it marked. Verified no non-TBD claim resolves through any of the six.

RC3 should now read consistent on this axis.