Two cleanup asks
1. Remove/quarantine the synthetic ADVERSARIAL_ANALYSIS.md
ADVERSARIAL_ANALYSIS.md (present since the initial commit 56659fc, 2026-01-02) is an LLM-generated red-team roleplay document, not a real audit — header reads "Author: Red Team Security Researcher (ex-DeepMind)", "Status: Breaking Everything", "Impact: Complete bypass of safety guarantees".
It states fabricated quantitative claims with no measurement behind them, e.g.:
- "k constraints → k_eff ≈ 3–5 independent"
- "Geometric guarantee BROKEN", "Probability: 90%"
These are contradicted by the actual data. Computing the effective independent dimensionality (N_eff) on the real 6,465-trace corpus (release/data_scrubbed_v1/accord_traces.jsonl, via CIRISLens/scripts/measure_n_eff.py) gives N_eff_PR ≈ 6.0, N_eff_H ≈ 7.7 (semantic/H3ERE features only, pre-CEG) — i.e. higher than the doc's fabricated "collapsed" 3–5. There is no adversarial N_eff measurement anywhere; the corpus is all benign traces.
The doc has proliferated to ~20 copies (clones ratchet-omega/-proofs/-data, 15 RATCHET_WORKTREES/wt-*, ratchet_demo/). A synthetic doc claiming "complete bypass of safety guarantees" in the repo root is a real liability — it has already been mistaken for a genuine security finding. Recommend: delete, or move under a clearly-labeled fiction/ path.
2. Catch RATCHET up to current coherence-ratchet + CIRISConstitution
RATCHET's last substantive commit is 2026-06-12 and predates several load-bearing developments. Items to reconcile:
- F-11 (2026-05-22): the joint multi-rung backward
P_ω is a documented no-go; D4 (CMB anomalies as TSVF signatures) retracted; the orthogonality theorem (framework = exactly ΛCDM for the CMB) is the sole surviving CMB content.
- Substrate-local retraction: the universal corridor band (0.1, 0.43) is retracted; bounds are substrate-specific (per-substrate calibration).
- Gate 0 (2026-07-02): cross-substrate saturation validation — coordination
k_eff is bounded (low-dimensional), not criticality, across complete coordinating units incl. a complete larval-zebrafish brain (decisive complete-unit test). The objective measure is saturation, not level. Framing corrected: the discriminator adjudicates a live criticality-vs-low-dimensionality debate (Munoz/Bialek vs the low-D-manifold camp) — it does not "discover low-rank" (established) nor is criticality "trivial."
- Two-pole dynamics correction: at M=0, ρ exits the corridor to either pole (rigidity or chaos); the single-pole "ρ→1 monotonic" was over-specified. A clean ECoG anesthesia test supports the rigidity default at the coordinating grain.
- Second axis: bound-vs-coordinating detectable via broken detailed balance (the γ·M maintenance term); validated estimator + macaque-motor-cortex positive control.
- CEG (in
CIRISConstitution): the CIRIS Epistemic Grammar substrate (the 1+4 attestation wire surface) is the substrate leg of the N_eff / deception-resistance argument (semantic H3ERE ~7 + orthogonal cryptographic substrate). RATCHET's safety/adversarial framing should reference it rather than the synthetic doc above.
Cross-refs: coherence-ratchet (F-11, Gate-0, two-pole dynamics, experiments/keff_saturation/, experiments/adversarial_neff/SPEC.md); CIRISConstitution (CEG + Accord, Book IX F = k_eff·λ·σ).
Two cleanup asks
1. Remove/quarantine the synthetic
ADVERSARIAL_ANALYSIS.mdADVERSARIAL_ANALYSIS.md(present since the initial commit56659fc, 2026-01-02) is an LLM-generated red-team roleplay document, not a real audit — header reads "Author: Red Team Security Researcher (ex-DeepMind)", "Status: Breaking Everything", "Impact: Complete bypass of safety guarantees".It states fabricated quantitative claims with no measurement behind them, e.g.:
These are contradicted by the actual data. Computing the effective independent dimensionality (N_eff) on the real 6,465-trace corpus (
release/data_scrubbed_v1/accord_traces.jsonl, viaCIRISLens/scripts/measure_n_eff.py) gives N_eff_PR ≈ 6.0, N_eff_H ≈ 7.7 (semantic/H3ERE features only, pre-CEG) — i.e. higher than the doc's fabricated "collapsed" 3–5. There is no adversarial N_eff measurement anywhere; the corpus is all benign traces.The doc has proliferated to ~20 copies (clones
ratchet-omega/-proofs/-data, 15RATCHET_WORKTREES/wt-*,ratchet_demo/). A synthetic doc claiming "complete bypass of safety guarantees" in the repo root is a real liability — it has already been mistaken for a genuine security finding. Recommend: delete, or move under a clearly-labeledfiction/path.2. Catch RATCHET up to current
coherence-ratchet+CIRISConstitutionRATCHET's last substantive commit is 2026-06-12 and predates several load-bearing developments. Items to reconcile:
P_ωis a documented no-go; D4 (CMB anomalies as TSVF signatures) retracted; the orthogonality theorem (framework = exactly ΛCDM for the CMB) is the sole surviving CMB content.k_effis bounded (low-dimensional), not criticality, across complete coordinating units incl. a complete larval-zebrafish brain (decisive complete-unit test). The objective measure is saturation, not level. Framing corrected: the discriminator adjudicates a live criticality-vs-low-dimensionality debate (Munoz/Bialek vs the low-D-manifold camp) — it does not "discover low-rank" (established) nor is criticality "trivial."CIRISConstitution): the CIRIS Epistemic Grammar substrate (the 1+4 attestation wire surface) is the substrate leg of the N_eff / deception-resistance argument (semantic H3ERE ~7 + orthogonal cryptographic substrate). RATCHET's safety/adversarial framing should reference it rather than the synthetic doc above.Cross-refs:
coherence-ratchet(F-11, Gate-0, two-pole dynamics,experiments/keff_saturation/,experiments/adversarial_neff/SPEC.md);CIRISConstitution(CEG + Accord, Book IXF = k_eff·λ·σ).