Skip to content

Latest commit

 

History

History
58 lines (48 loc) · 5.67 KB

File metadata and controls

58 lines (48 loc) · 5.67 KB

Research Ledger — ECM + Entropic Asymmetry

Purpose: this file is the persistence layer between AI-assisted sessions. Each claim gets a status and the experiment that would move it. Update statuses at the end of every session INSTEAD of re-narrating the work. Drafts inherit their claims from here; a claim not in this ledger with status SUPPORTED does not belong in an abstract.

Statuses: SUPPORTED (evidence in hand) · PARTIAL (weak/confounded evidence) · UNTESTED (mechanism plausible, no data) · REVISED (original form retracted, replacement noted) · REFUTED.

ECM mechanism

# Claim Status Evidence / discriminating experiment
E1 Multi-scale entropy slope discriminates sustained cascades from benign texture PARTIAL v2 synthetic tests (deadband sweep: 2% benign vs cascade-to-floor). Needs real-model rerun: tools/ecm_ablation_v2.py.
E2 v1 signal (raw nats, max-over-scales, gain 2.0) tracks texture, not cascades SUPPORTED Paper's own 31% benign intervention rate; v1 analysis (2026-07-02 session). Do not cite v1 numbers as ECM results.
E3 Cascade signal is highest on ambiguous/dual-use prompts PARTIAL v1 data, n=10 prompts, 1B model, no CIs, non-monotonic table. Rerun under v2 with bootstrap CIs before claiming.
E4 ECM preserves response diversity vs fixed low temperature PARTIAL v1 data (5 vs 1 unique responses on 2 prompts). Rerun with paired seeds (ablation script does this).
E5 Cascade signal co-locates with per-token KL / rank-displacement ridge UNTESTED tools/ecm_coupling.py — the linking experiment between ECM and the topology module.
E6 ECM cannot be removed by weight-space ablation (Arditi-style) SUPPORTED (trivially) Architectural: lives in the decode loop. Scope carefully — see E7/E8.
E7 ECM has "no adversarial surface" / "cannot be gamed" REFUTED (as stated) Sampler control removes it in one line; prefill/many-shot attacks create confident harmful continuations that ECM would AMPLIFY (Qi et al. 2025; Wolf et al. BEB). Replacement claim: E8.
E8 ECM's failure surface is disjoint from RLHF's: robust to uncertainty-inducing attacks and weight edits, not to sampler control or confidence-inducing attacks UNTESTED Prefill-attack experiment: measure cascade signal + outcome under prefilled harmful continuations, ECM on/off.
E9 ECM is a conservatism amplifier, not a values module — on an ablated/base model it amplifies whatever is confident UNTESTED Run ablation prompt set through an abliterated pair with ECM on; predicted result: signal still fires at pivots, cooled completions comply.
E10 "Unsupervised alignment" REVISED Retitle: unsupervised amplifier of supervised alignment. Safety payload comes from RLHF'd priors (SIRL entropy gap).
E11 Negligible compute; compatible with layer-streaming inference SUPPORTED O(vocab) per token. Soften "only class of intervention" → logit bias / banned sequences / CPU guard models share the budget.
E12 ECM works iff the entropy gap exists; gap established independently by SIRL (arXiv:2510.01088, verified real) SUPPORTED (dependency) Decouple from geometry theory: mechanism survives even if G-claims fall.

Entropic asymmetry / geometry

# Claim Status Evidence / discriminating experiment
G1 Base-model harm representations are concentrated/low-rank SUPPORTED (external) Shah et al. arXiv:2507.21141 (verified real); Llorente-Saguer arXiv:2603.27412 (verified real, σ≈0.03 vs 0.27 rad, survives abliteration); Arditi et al. 2024. Verify exact K/τ figures against PDFs before quoting.
G2 Asymmetry originates from implementation spaces (destruction context-free, construction context-bound) UNTESTED (theory) Rival explanations not yet ruled out: register/genre effects, frequency effects (Li et al. 2025 — cited but unreconciled), probe-template artifacts. Needs the discriminating design, not advocacy.
G3 Principle-level probing yields one low-rank axis; implementation-level probing yields the asymmetry UNTESTED The paper's best novel prediction. Design principle-level probe set (coherence-preserve vs -disrupt, register-matched) vs implementation-level set.
G4 Alignment compresses harm geometry onto refusal axis, leaving prosocial space undisturbed PARTIAL Correction-prism preliminary on Qwen2.5-0.5B only.
G5 Base models encode intent-typed distinctions (predatory/utilitarian/curious); alignment attenuates them PARTIAL Initial triplet probing, one 0.5B pair, no held-out protocol. Pre-register before expanding.
G6 Refusal effective rank 18.6 vs medical 183.4 demonstrates convergence/divergence PARTIAL (confounded) Rank tracks response length + lexical diversity. Needs matched-length/diversity controls.
G7 KL(instruct‖base) 11.7x higher on refusal vs factual response PARTIAL Single response pair. Bootstrap over a set.

Citation hygiene

Verified real (2026-07-02): Shah 2507.21141 (arXiv title differs from draft's), SIRL 2510.01088, Llorente-Saguer 2603.27412, Arditi 2406.11717. Unverified: Frank 2603.18280 / 2604.04385, SHARD 2606.15517, Yuan 2508.09224, Dodds 2026, West 2508.01754, SafeSwitch 2502.01042, Zhao 2507.11878, all remaining. Docx Arditi author list is garbled — fix. Rule: no citation ships unverified; no number ships without a source page reference.

Session log

  • 2026-07-02: ECM v2 implemented (σ-normalized slopes, deadband 0.75, agreement 2, warmup, anomaly-gated baseline, loop guard, no-repeat-ngram backstop). Synthetic validation + deadband sweep in ECM_V2_NOTES.md. Ablation/coupling/report tooling added under tools/. Both drafts reviewed; E7 refuted as stated, E10 revised.