Events 137+138: gate signal quality + v2.0 epistemic engine#93
Merged
Conversation
Event 137 — audit-driven fixes to the loops that were defeating the kernel's own thesis: - cascade detector: positive-system read-only command exemption ahead of Trigger 2 — wc/ls/grep on sensitive paths no longer classify as cascade:architectural. Root cause of alarm fatigue: 18.8k ops classified high-impact drove the kernel's own repo to advisory mode and 57.9% authoring vs the 80% floor. - session_context: bound NEXT_STEPS.md injection to an 8KB head with a truncation marker (live: 233KB -> 10KB per hook invocation). - E1 self-falsification: SessionStart digest and episteme report (new Protocol Synthesis section, --framework-dir override) evaluate FALSIFIABILITY_CONDITIONS § E1 against live framework state. E1 is FIRED (0 protocols, ~49 days); kernel doc status updated and README marks the compounding arm aspirational per E1's prescribed action. 1274 tests + 54 subtests green (33 new: 17 exemption incl. adversarial bypass shapes, 11 session-context bounds/E1, 5 report E1).
…d verification, lesson synthesis Event 138 — the product's center shifts from deterministic form-gates to a senior-researcher epistemic engine, per operator directive and the 2023-2026 research record (DESIGN_V2_0_EPISTEMIC_ENGINE.md): - epistemic-interrogation skill: decompose a load-bearing decision into tiered claims, verify load-bearing claims in a fresh context against external evidence (CoVe-factored), argue the opposition, name the weakest link, pre-commit disconfirmation, write a verdict artifact. - core/hooks/_interrogation.py: deterministic floors only — freshness via max(ts, mtime) (fixes stale-at-birth), external-evidence requirement, verdict consistency (refuted load-bearing claim + proceed = contradiction; stop fails closed). - reasoning_surface_guard: a fresh verdict is an alternative satisfier to the v1 surface for high-impact ops; advisory text converted to factual statements per the hooks doctrine; audit records carry source: surface|interrogation (E3 measurement). - fence_synthesis: lesson synthesis on successful ops — non-null lessons from verified interrogations become context-scoped protocols (deduped by lesson hash), giving E1's compounding claim a live source; e2e test covers interrogation -> protocol -> guidance. - Identity docs reframed (README lede, THE_WAY_TO_THINK § 5 division of labor, FALSIFIABILITY E3/E4/E5, REFERENCES v2 empirical record). 1298 tests + 54 subtests green (24 new).
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this is
Two events, one arc: the Event 137 audit found the kernel checking form while its compounding loop sat dead and its own falsifiability condition (E1) had fired unnoticed; Event 138 rebuilds the center as the v2.0 Epistemic Engine per operator directive and the 2023–2026 research record.
Event 137 — gate signal quality (audit-driven fixes)
episteme report§ Protocol Synthesis evaluateFALSIFIABILITY_CONDITIONS § E1against live state. E1 status → FIRED (0 protocols, ~49 days); README marks the compounding arm per E1's prescribed action.Event 138 — v2.0 Epistemic Engine (
docs/DESIGN_V2_0_EPISTEMIC_ENGINE.md)Division of labor from the research record: deterministic structure guarantees the epistemic work happened; model judgment — factored, evidence-anchored, in a context that never sees the draft — checks it was real.
skills/custom/epistemic-interrogation/— the senior-researcher protocol: decompose into tiered claims, factored verification of load-bearing claims (CoVe, arXiv 2309.11495), argued opposition (debate, arXiv 2402.06782), weakest link, pre-committed disconfirmation, verdict artifact.core/hooks/_interrogation.py— floors only: freshness via max(ts, mtime), external-evidence requirement (self-correction without external signal degrades accuracy, arXiv 2310.01798), verdict consistency;stopfails closed.sourcefor the E3 falsifiability measurement.Verification
episteme reportrenders § Protocol Synthesis; read-only exemption verified against adversarial bypass shapes (redirection into sensitive paths, command substitution, mixed pipelines, xargs, git mutating subcommands — all still fire).Out of scope (logged in NEXT_STEPS / deferred discoveries)
i18n README parity for the v2 reframe; web/ landing copy; imperative-tone pass on remaining hooks; architecture_v2 diagram regen.