feat(selector): literature-driven disagreement router v2 - #6
Merged
Merged
Conversation
Three-way outcome (TRUST/TRUST_SPOT_CHECK/ESCALATE), forced-escalate categories, Laplace-smoothed confidence, and escalationPriority — each mechanism grounded in a read paper (EDV, Agentic Abstention, OmniOPD, SkillHarness). Adds a CONVOLVE-style skeptic playbook template distilled from the megabench B3 collective-failure traps. Pure domain module; 18 vitest+fast-check cases; full suite 680 green; check:docs unaffected. Codex-reviewed ACCEPTED.
Collaborator
Author
Held-out validation (follow-up to the in-sample caveat)The PR body flagged the skeptic-playbook numbers as in-sample (playbook distilled from the same 60 traps). Ran the held-out check: 20 brand-new trap questions, same 10 categories, different specific traps the playbook never saw.
Conclusion: the CONVOLVE-style category-level rules generalize to unseen traps (they're not memorized questions). The in-sample 58/60 overstates, but the transferable component is real and large. Data/scripts archived alongside the megabench reports. |
This was referenced Jul 2, 2026
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Selector v2 — a verifier-aware disagreement router for best-of-N fan-out, grounded in 6 papers from the local reading notes and validated on the megabench 60-task trap benchmark.
Mechanisms (each ← a paper)
TRUST / TRUST_SPOT_CHECK / ESCALATE— EDV (self-confirmation trap): unverified consensus is never clean trust.security/correctness/impossible) — SkillHarness safe-skill boundaries + B3 data showing consensus fails correlated there.(k+1)/(n+2)— OmniOPD Bayesian smoothing: a 5/5 fleet reads as 0.857, not certainty.Empirical (offline replay on 60-task benchmark, strict dual-judge)
Approaches the best premium single (59/60) at a fraction of premium cost.⚠️ skeptic numbers are in-sample (playbook distilled from these traps) — held-out validation is the next step (deliberately out of scope here).
Test
Pure domain module. 18 vitest + fast-check cases; full suite 680 green;
check:docsunaffected (domain tests not gated). Codex-reviewed ACCEPTED.Not in this PR
Runtime wiring into dispatch/loop, playbook auto-evolution (SelfHarnessLoop), learned TRINITY-style router — all follow-ups.