idkmesh

Current-main conjunctive evolution convergence

Date: 2026-08-28
Repository: MSKazemi/idkmesh

Owner direction

Continue strengthening the mathematical and algorithmic foundation and implement it through GitHub-native mechanisms while retaining the public reasoning trail.

What happened concurrently

This pass first merged:

The first canonical portfolio run consumed a trusted Bayesian checkpoint and surfaced parallel evolution-observer PRs as high review-attention candidates.

While a convergence branch was being prepared, PR #144 merged a useful Repository Evolution Observatory onto main. It added a stronger trusted PR event boundary, carrying capacity, graph/reference signals, Shannon diversity, control-energy deficits, replicator-mutator strategy response, hard GUARD, Anti-Goodhart exclusions, and immutable Action pins.

That concurrent merge also replaced the artifact-backed Bayesian update in evolution-loop.yml with a purely recomputed observation. The mathematical pieces were therefore individually useful but no longer composed.

A stale convergence PR (#146) was intentionally abandoned rather than forced across that semantic collision.

Clean convergence from current main

The new current-main branch keeps the merged observatory intact and adds only the missing composition:

persistent Bayesian history
 + current Repository Evolution Observatory
 + live Pareto/UCB Repository Mathematical Portfolio
 -> conjunctive bounded recommendation
 -> independent verification / GitHub governance

New conjunctive controller

scripts/conjunctive_evolution_control.py uses conservative Bayesian confidence bounds together with live observatory blockers/capacity.

A stronger bounded non-integrating experiment is a candidate only when:

No integration, approval, merge, branch mutation, spending, or constitutional authority is created.

Regression tests prove that perfect historical Bayesian confidence cannot override main_unprotected -> GUARD, and that weak history can block escalation even when the live state is clean.

Bayesian persistence restored beside the observatory

The trusted evolution workflow now searches recent successful default-branch runs for the newest actual evolution-checkpoint-* artifact, restores its Bayesian state/ledger, updates history, then recomputes the current observatory and conjunctive decision.

All three evidence layers are retained in the next checkpoint artifact.

The lookup is constrained to successful default-branch runs so PR-generated artifacts cannot become trusted history.

Portfolio trust boundary hardened

The Repository Mathematical Portfolio now mirrors the observatory trust pattern:

Standalone kernel supply chain

The standalone Mathematical Evolution Kernel now uses immutable Action SHAs and no longer needs actions: read.

Remaining truth

GitHub still reports main as unprotected. Therefore the current live observatory must remain in GUARD, and the conjunctive controller must return false for stronger experiment escalation until the external branch/ruleset control in issue #35 is actually configured.

The repository can improve mathematics, verification, simulation, documentation, and bounded experiments in the meantime; it cannot truthfully treat internal scoring as a substitute for the missing external governance boundary.