idkmesh

Conjunctive evolution governor convergence pass

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

Owner direction

Continue strengthening IDKMesh with GitHub-native mathematical/evolutionary mechanisms, open public issues for important actions that cannot be completed directly, and preserve useful chat-derived reasoning/decisions in the repository.

Starting point

This turn began by inspecting the existing IDKMesh Evolution Loop instead of creating a new controller blindly.

The repository then evolved concurrently several times, and the correct architecture changed as stronger work landed:

Rather than treating earlier branches as entitled to merge because they were created first, the work was repeatedly reconciled against these new canonical surfaces.

Superseded exploratory branches

#139

The first live-observer proposal conflicted with the stronger mathematical core from #137. It was closed unmerged and retained as public provenance.

#142

A second clean branch correctly preserved #137 and introduced a hard current-state governor, but it became stale as #143/#144 landed. Its unique ideas were ported into the current convergence PR #146 rather than forcing stale ancestry.

Final responsibility separation

The converged design is:

#137 persistent Bayesian history
        +
#146 hard recomputed current-state governor
        +
#143 live Pareto/UCB attention portfolio
        |
        v
bounded recommendation
        |
        v
independent verification + external GitHub governance

PR #144’s stateless scorer remains on main as an offline comparison/falsification baseline. It is valuable because disagreement between stateless and persistent/current-guarded reasoning can expose stale priors or weak proxies; it is not wired as an additional autonomous controller.

Why the hard live governor still adds value

Persistent Bayesian state answers historical/uncertainty questions. Pareto/UCB answers where attention may be informative. Neither should be able to compensate for a failed current hard condition.

The live governor recomputes:

Hard rule:

main_protected = false => GUARD

Historical Bayesian fitness, Pareto rank, UCB focus, stars, forks, reactions, comments, or commit counts cannot override it.

Checkpoint integration bug discovered

After #144 merged, the live Evolution Loop published artifacts named:

evolution-observation-*

while the #143 portfolio expected:

evolution-checkpoint-*
  containing state/evolution-state.json

That meant the portfolio’s Bayesian-health input could silently fall back to the repository seed instead of consuming the trusted persistent Bayesian checkpoint.

PR #146 restores the persistent checkpoint contract while also publishing the live hard-governor evidence in the same trusted observation artifact.

GitHub Actions trust-boundary correction

A security subtlety was explicitly handled: ordinary pull_request workflow YAML is proposed code, so live API/checkpoint authority must not run from that PR-controlled definition.

The converged workflows separate contexts:

Trusted live observation

Proposed-code verification

Artifact-state concurrency

Trusted artifact-backed observers use repository-level latest-state concurrency with cancel-in-progress: true so concurrent event bursts do not restore the same checkpoint and publish forked successor observer state.

This makes the checkpoint a latest-state observer, not an immutable raw event ledger.

Supply-chain hardening

Core mathematical workflows use reviewed immutable commits for:

The hard governor independently measures repository-wide SHA-pin coverage so these local fixes cannot be mistaken for proof that every workflow is already fully pinned.

Privacy/evidence-minimization finding

The #143 portfolio needs issue/PR text ephemerally for deterministic classification and reference extraction. Its original workflow copied the raw snapshot — including issue/PR bodies — into the retained replay artifact.

That was unnecessary retained data.

PR #146 changes the boundary to:

raw issue/PR bodies -> ephemeral /tmp snapshot
                  -> deterministic portfolio calculation
                  -> not copied into retained artifact

The retained portfolio checkpoint now contains only derived portfolio state/output, policy, and Markdown. A workflow assertion verifies that repository-snapshot.json is absent from the retained checkpoint directory.

Clean branch reconstruction

PR #146 initially inherited conflicting/stale history even though its recorded base matched current main.

The branch was professionally reconstructed as one clean commit on exact current main using exactly the intended convergence file blobs. Unrelated benchmark/IDKGraph changes from main were inherited directly rather than copied or overwritten.

All pre-reconstruction CI was intentionally considered stale and must be rerun on the new exact head.

Anti-Goodhart rule

Live hard-governor fitness excludes:

The broader interpretation is:

Bayesian posterior != causality
Pareto front        != correctness
UCB focus           != approval
activity volume     != improvement
popularity          != trust

Chat-to-repository propagation

The existing PROJECT_RULES.md mandatory preservation rule remains canonical.

Substantive project turns should continue to use two layers:

docs/conversations/ structured public record
+
promotion into canonical code/docs/issues when the turn changes the project

The live snapshot can observe that the rule/archive surfaces exist. It cannot prove completeness for a conversation that was never committed.

External blockers surfaced publicly

#35 — protect main

Public GitHub metadata still reports main unprotected. The issue remains the canonical administrative hard gate. Repository workflows cannot substitute for this external enforcement.

#138 — independent human review for PR #91

Automation cannot manufacture a genuinely independent human witness. A bounded expert-review issue was opened instead of self-approving the canonical node.

Remaining completion path for #146

  1. require fresh exact-head PR-head CI after the clean reconstruction and privacy/doc corrections;
  2. verify the PR is mergeable and contains only the intended convergence files;
  3. verify the live/PR trust separation and minimum permissions from the actual diff;
  4. open a bounded independent-review issue for #146;
  5. close #142 unmerged as superseded once #146 is demonstrably clean/green;
  6. do not self-merge #146.

Scientific next step

After the control-plane convergence, the next strong mathematical step should be calibration from delayed outcomes rather than adding a fourth controller. Historical observations can be joined to regressions/reverts, verifier disagreement, review latency, benchmark movement, newcomer completion, external contributor retention, security findings, and time-to-verified-useful-work, then compared with held-out Brier/log-loss/calibration/ranking-regret metrics.

Negative results remain first-class evidence.