Re-base the coherence mathematics on TORQUE (hold-while-paid) and retire the corridor framing — CC 1.0-rc3 cannot cite the current basis

[emooreatx]

Why this is filed now

CIRISConstitution cut a 1.0-rc3 candidate against the open issue set, then put it through five independent adversarial reviews. The Part VI remediation — demote §6.2, re-found the corridor, strip CCA as authority — came back blocking. The reviewer's finding, in short: RC3 removed one unsupported basis (J = k_eff·λ_op·σ, which is monotonically maximised at ρ̄ = 0 and therefore cannot support a two-sided corridor at all) and replaced it with an axiom that posits the endpoints outright, decorated with theorems about a different quantity.

RATCHET's own #12 and #13 arrive at the same place from the data side. Together they mean the corridor is not a claim that can be repaired by better wording, and CC should stop trying.

This issue asks RATCHET to re-base the mathematics on the TORQUE claim (#16) — the harness holds agent behaviour to a declared ethical design — and to let the corridor framing go.

What the evidence actually supports right now

Stated at honest strength, and deliberately conservative:

Dead or unsupported:

• The rigidity arm of the corridor has no measured support. Exp 103 (software-induced collapse) is contradicted 3-to-1 by F2, exp117, and F3 — blocking republication #13 is explicit: Exp 103 is "the empirical anchor for the rigidity arm of the corridor — which F3 otherwise never measured," with max measured ρ = 0.171 against a 0.43 edge (CCA v3 corrections version: Theorem 2.3 criterion (α/k not α/k_eff — verified in sympy), Table 8 defects, validation scoring — fix the Zenodo record, don't delete it #11 defect C-14). And Exp 103 is contradicted 3-to-1 by F2, exp117 and F3, all read from stored JSON. So the rigidity endpoint is both unmeasured and anchored to a result under active challenge.
• The institutional k_eff corpus is inverted on every analyzable row (Data integrity: wgi_processed.csv k_eff is inverted on every row (4933/4933); three inconsistent institutional pipelines #12: 4933/4933, k < 1 and k_eff > k, zero exceptions), k is not a count, rho is not a correlation, three inconsistent pipelines give three values for one country, and the pipeline that defines k is not runnable from the repo. Those are received numbers with no live path back to data.
• CCA v5 has already withdrawn the k_eff hardware validations (reclassified as identity checks), re-scored institutional application below chance, and marked the collapse asymmetry "assumed, not derived."
• Independently: the corridor's pole theorems hold at k = 3, and symmetric noise mints even-order structure from four slots up — so the rigidity pole is not zero in the regime the ratchet argument is actually about.

Surviving, and worth keeping:

• The Kish identity and the 1/ρ̄ ceiling — an identity and its limit, not a discovery, correctly credited to Kish 1965 and Laakso–Taagepera 1979.
• "Scale cannot restore collapsed diversity" — the Möbius/ceiling core CCA v5 still supports.
• The voice-pricing consequence, which is what ciris.ai/safety actually claims: "copies of one voice count as one voice, and no amount of echo can buy more trust." Note this needs no corridor, no healthy band, and no collapse asymmetry — it needs only the ceiling.

The re-base: TORQUE as the load-bearing claim

#16's claim is the right shape and the right verb: the pipeline holds behaviour coherent with a declared, legible value design, by maintaining that design at inference time against the model's trained-in prior — holds while paid, not "forces," not "aligns the weights."

That is a claim about a maintained design, and it is the one CIRIS can actually stake, because:

• It is a hold guarantee with a measured reversion constant (arm B), not a geometry claim about deception feasibility.
• Arm C makes the machinery explicitly value-neutral — the safety content lives in the covenant. That is exactly what CC's own axiomology says ("no physics implies the ethics") and exactly what the corridor framing was smuggling past.
• It carries separable kills per arm and pinned void conditions, so it can fail loudly instead of being quietly re-worded.

Important scope note, so this issue is not misread as an ask to cite an unrun campaign: #16 is a pre-registration. It states plainly that "no stance or website claim moves until RESULTS lands and is reviewed." CC must not cite TORQUE as evidence for anything until Stage 0 passes its planted-dye calibration and the arms report. The correct interim state for CC is a stated gap, not a substituted basis.

Asks

Mathematics / claim structure

• Decide whether the corridor survives as a claim at all. Our read is that it does not, on this evidence, and should be retired rather than re-founded. If RATCHET disagrees, name the object that establishes a two-sided band and the strength it carries.
• If retired: mark the corridor's rigidity arm unmeasured (per Exp 103 (software-induced collapse) is contradicted 3-to-1 by F2, exp117, and F3 — blocking republication #13's own proposed checkbox), and say so in the places that currently assert a band.
• Re-base the downstream framing on the hold/maintenance result (TORQUE campaign prereg: does H3ERE hold agent behaviour to a declared design — five arms, separable kills, live-agent runs #16) plus the surviving ceiling core, at stated strengths.
• Resolve the k ≥ 4 question: does the whole-only-share pole result extend past three slots, or is the even-order floor fatal to the "both poles are zero" framing?

Evidence layer — blocking CC regardless of the above

• CIRISOntology is cited by CC Part VI as the corridor's new basis but is not pinned, not vendored, not in claims.tsv, and its pointer form is outside CC's declared vocabulary — so CC's CI emits zero warnings about its most important new claim. Either publish a spec-map manifest CC can pin, or CC drops the citations.
• The measured forward prediction cited by CC (−3.084 ± 0.219 vs −3.109 staked in advance) has no bench pointer, no population, no sample size in CC. Supply the artifact path or CC removes the number.
• RATCHET.cc_formal.tsv and coherence-ratchet.cc_lean.tsv disagree on 11 objects. The 6.2.1↔6.2.2 assignments are systematically inverted between them; the Core.Dynamics block is filed wrongly by both; remainder_scales_with_k_eff is mechanized in one and open in the other; and Core.EffectiveConstraints — named by coherence-ratchet as the home of the decay law — is published by neither. Because CC's checker resolves by section decimal rather than by symbol, this currently lets a CC claim resolve green on a theorem that does not exist at the pinned commit.
• evidence_pins.tsv backs coherence-ratchet on mechanized only but RATCHET on mechanized or empirical, so an empirical exponent backs a decimal at the same apparent strength as a Lean theorem. Worth a shared convention.
• The exp0_cca_validation bench row still backs CC 6.2.4 in the pinned manifest. Post-v5 it is an identity check; consider re-labelling at source.

Timing — this is not a fire drill

CC 1.0-rc3 will not be released until CIRISAgent 2.9.7 merges and the TORQUE campaign completes. Nothing here is blocking on a deadline, and nobody should shorten a stake or skip a gate to unblock CC. The published constitution is still rc2, which makes no claim this issue disturbs. What CC is doing in the meantime is adjusting the candidate so it stops asserting things whose evidence is still in flight — which mostly means removing claims, not waiting on new ones.

The one thing that would genuinely help sooner rather than later is the evidence-layer hygiene above (manifest decimals, the re-pin, a CIRISOntology spec-map), because that is what lets CC's CI see these claims at all.

What CIRISConstitution will do meanwhile

CC's RC3 candidate is on claude/rc3-cut-summary-lo4me4 and is not being cut until this resolves. The intended direction, subject to your view:

• Retire the pinned (0.1, 0.43) band. It rests on the corridorBounds axiom; corridor_bounds_well_formed certifies only that lower < upper, which is true of any pair; and the "measured, consistent with 1/ρ_lower" corroboration is circular, since 1/ρ_lower = 10 only because ρ_lower was posited at 0.1. Exp 103 (software-induced collapse) is contradicted 3-to-1 by F2, exp117, and F3 — blocking republication #13's ρ = 0.171 ceiling makes the rigidity endpoint unreachable on measured data.
• Keep the Kish identity and the ceiling, labelled identity.
• State the corridor gap openly rather than substituting a new basis, and carry TORQUE as the campaign that would license a claim here — cited as pre-registered, not yet run.

Refs: #12, #13, #15, #16, #9/#10, #11 (defects C-13, C-14, C-18); CCA v5 (10.5281/zenodo.21730551); CIRISConstitution#50, #45, #26; ciris.ai/safety.