idkmesh

Maintainer continuation — contracts, observers, and evidence boundaries

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

This record summarizes the maintainer/integration continuation after the large PR/branch convergence pass. The governing rule throughout was transactional: every accepted merge changes canonical main, so earlier eligibility, branch-plan, and exact-head judgments become historical until revalidated.

1. Project/domain contract boundary

PR #202 merged the first executable Core -> DomainPack -> ProjectManifest boundary.

It added:

The PR was rebuilt on then-current main while preserving the ten reviewed implementation blobs exactly, and fresh combined-tree Project Domain Contracts, Phase 0, IDKGraph, randomness, and Evolution checks passed before merge.

Issue #6 is complete.

2. Branch Steward convergence

Historical planner PR #205 was not merged twice. Merged #211 had already transplanted its six implementation blobs onto current main, so #205 was closed as superseded.

PR #215 added canonical branch-state resnapshotting after every main push.

PR #217 then added PR-lifecycle observation. Maintainer review corrected an initial freshness race: pull_request_target must never use contributor-head code, and its first design also must not allow a recorded pre-merge base.sha scan to supersede a newer main push scan.

The corrected observer used trusted canonical code, but real repository activity exposed a second-order control problem: observation itself consumed too much GitHub API capacity.

Post-#221 Branch Convergence Audit run 33205890139 passed deterministic tests and then exhausted the GitHub App installation API budget during O(branches) live classification.

PR #222 therefore made the Branch Steward homeostatic:

Exact post-merge evidence:

Issue #127 contains the canonical branch-lifecycle record. No fresh branch counts should be inferred from a blocked/incomplete scan.

3. Human review evidence remains human

PR #203 merged a machine-checkable IDKGraph review-session format and validator. It can validate reviewer identity/independence disclosure, anchoring exposure, active review minutes, labels, confidence, evidence and recommendations.

It does not create a reviewer, infer semantic labels, approve work, or grant integration authority.

During integration the IDKGraph workflow was also hardened with immutable action SHAs and persist-credentials: false.

This distinction matters for the remaining PRs: project automation may validate independent human evidence, but cannot manufacture the independence itself.

4. Algorithm Collaboration Fabric

PR #214 merged the Algorithm Collaboration Fabric architecture.

The key non-compensation rule is now canonical:

hard safety / feasibility gate
    cannot be overridden by
soft optimization or collaboration score

Correctness verification and integration governance remain separate authority planes. A future collaboration signal may route attention or nominate bounded experiments, but cannot make unsafe or unverified work acceptable.

5. Active compute pulse remains externally blocked

Draft PR #213 installs a fail-closed zero-cost GitHub Actions compute mechanism, but issue #35 remains an external governance blocker because public GitHub metadata still reports main as unprotected.

Maintainer review found and fixed an activation-ordering flaw: the early workflow gate originally depended only on main protection + repository opt-in, while canonical repository/ref/event validation happened later in the Python adapter. A manual dispatch on a non-main ref could therefore reach dependency installation/repository scripts before final rejection.

The draft branch now requires, before any active step:

The corrected draft head passed Phase 0, IDKGraph and Evolution checks. It remains draft because mechanism readiness is not activation authority.

6. Adversarial evidence envelope and finite numeric domain

PR #216 added a sharp count-contamination envelope for scalar reports under at most f arbitrary reports.

For sorted observations x_(1) <= ... <= x_(n) and f < n:

L_f = mean(lowest n-f reports)
U_f = mean(highest n-f reports)

Every admissible honest-report subset mean lies in [L_f, U_f]; the endpoints are sharp under the count-only assumption.

Maintainer review found a numeric-contract defect before merge: ordinary Python comparisons allow NaN to evade bound checks. A concurrent squash merge captured the pre-fix version, so corrective PR #218 was created from actual merged main and added finite-domain checks for reports, bounds, threshold and margin plus regression tests.

PR #218 merged as 0504eac6274579dd6daf5b7c32ab405a77d9bc09.

The envelope remains report-level decision support only; it makes no truth, Sybil-resistance or Byzantine-consensus claim.

7. Typed Evidence Aggregation Fabric and strict Boolean correction

PR #220 merged a typed, non-scalar evidence lattice. It deliberately does not multiply provenance, discrimination, correlation, contamination, sequential and drift channels into one confidence number.

Focused post-merge review found a fail-open typing bug: control fields were read using Python truthiness, so values such as "false" could behave as true.

Corrective PR #221 made these fields strict required Booleans:

Truthy strings, numeric substitutes and missing fields now fail closed. PR #221 merged as 95660ebd7c72b28a7a81a875b897b685ec0d6d4e after the dedicated Evidence Aggregation Fabric, randomness and Evolution checks passed.

8. Current open integration boundaries

At the end of this record the live PR queue contains three intentionally held paths:

#219 — canonical self-evolution methodology

Open and review-ready, labeled help wanted.

Its own risk section requires independent review before integration. No PR review has been submitted. It therefore remains outside main; author/project automation does not manufacture the required independent approval.

#213 — active zero-cost compute pulse

Draft. Mechanically hardened, but external issue #35 (protected main) remains unsatisfied. Merge must not be interpreted as activation.

#159 — canonical local worker

Draft. Exact candidate already has fresh CI and controlled Docker positive + A–E2 evidence, but issue #138 requires a genuinely separate human reviewer. Automation, worker success, byte equality, or verifier evidence are not that reviewer.

9. Control-system lesson

This continuation produced a useful repository-level control law:

more observation != more knowledge

when observation consumes a shared scarce resource.

The safer rule is:

if state changes but observation capacity is low:
    invalidate old belief
    reduce authority
    wait for enough sensing capacity
else:
    resnapshot
    recompute

This is the same homeostatic principle used in ACE and other IDKMesh controllers: resource pressure suppresses amplification rather than being compensated by activity.

Maintainer invariant

event != action
action != candidate
candidate != verified result
verification != approval
approval != merge
merge != realized improvement
stale observation != current authority
blocked observation != partial permission

The repository should continue optimizing verified durable progress per unit of reviewer, maintainer, compute and observation cost—not branch, PR, workflow or event volume.