Date: 2026-08-28
Repository: MSKazemi/idkmesh
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.
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:
pull_request_target, immutable action pins, and fail-closed GUARD behavior on unprotected main.Rather than treating earlier branches as entitled to merge because they were created first, the work was repeatedly reconciled against these new canonical surfaces.
The first live-observer proposal conflicted with the stronger mathematical core from #137. It was closed unmerged and retained as public provenance.
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.
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.
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.
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.
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:
pull_request_target for PR metadata observation;pull_request;contents: read only;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.
Core mathematical workflows use reviewed immutable commits for:
actions/checkout v7.0.1;actions/setup-python v7.0.0;actions/upload-artifact v7.0.1;actions/download-artifact v8.0.1.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.
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.
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.
Live hard-governor fitness excludes:
The broader interpretation is:
Bayesian posterior != causality
Pareto front != correctness
UCB focus != approval
activity volume != improvement
popularity != trust
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.
mainPublic GitHub metadata still reports main unprotected. The issue remains the canonical administrative hard gate. Repository workflows cannot substitute for this external enforcement.
Automation cannot manufacture a genuinely independent human witness. A bounded expert-review issue was opened instead of self-approving the canonical node.
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.