idkmesh

Continuation: anytime-valid sequential evidence

Date: 2026-08-28

User request: continue strengthening MSKazemi/idkmesh with a more solid mathematical and algorithmic background and implement the result with GitHub-native automation.

Repository state checked first

The live main branch was read directly rather than inferred from repository-wide commit search. At the start of this continuation it was:

523b10819abe1e88ce0207665098248ac0ed980b
Record successful Phase B2 Task 003 calibration (#201)

GitHub reported main as protected: false with status-check enforcement off. This remains a hard governance boundary; the new work does not increase autonomous write or integration authority.

The open PR set was also checked. Current open work covered ProjectManifest/DomainPack interfaces (#202), a guarded zero-cost compute pulse (#195), and the canonical node evidence path (#159). None implemented sequential statistical evidence under optional stopping.

An earlier repository-wide commit search surfaced commits that were not on canonical main; the direct branch read corrected that immediately. The branch ref is the source of truth for this implementation.

Mathematical gap selected

The existing mathematical layer already includes:

• Bayesian evidence and lower-confidence scoring;
• correlation-aware effective sample size and Bayesian vote aggregation;
• NSGA/Pareto ranking and crowding distance;
• multiplicative weights and UCB exploration;
• entropy/Jensen-Shannon diversity;
• DAG unlock value;
• Lyapunov-style homeostatic potential;
• conjunctive live-governance guards.

The missing primitive was optional-stopping-safe sequential evaluation. Repeatedly applying ordinary fixed-horizon confidence intervals can create false evidence simply because the repository keeps observing.

Implementation decision

Add a small dependency-free Sequential Evidence Kernel rather than modify the live Evolution Loop.

The kernel provides:

1. a union-Hoeffding confidence sequence using a summable time-indexed error budget;
2. paired candidate-minus-baseline sequential evaluation;
3. bounded inverse-propensity off-policy evidence;
4. Kish effective sample size for importance-weight concentration;
5. fail-closed handling when any importance ratio is clipped;
6. a hard non-compensation rule: strong statistics cannot override a failed governance guard.

A positive output is only experiment_candidate. It is not approval, merge authority, activation authority, or proof of causality.

GitHub execution boundary

The companion workflow is intentionally narrow:

permissions: contents: read

It runs only when the sequential-evidence files change (plus manual dispatch), tests Python 3.11 and 3.13, uses immutable Action SHAs already used by the repository mathematical workflows, publishes a deterministic demonstration artifact, and grants no repository mutation permission.

Next evidence step

The implementation should be accepted only if its pull-request checks pass on the exact candidate head. Because main remains unprotected, this mathematical layer must not be used to justify increased autonomous repository writes.

This site is open source. Improve this page.