Status: current-state audit after rapid integration of the Verified Swarm Runner, ACE safety/evidence stack, and mathematical foundations.
IDKMesh is no longer bottlenecked by lack of theory or by a large PR queue. The repository has converged to one open pull request, PR #91, containing the canonical real local worker. Its exact head has green Node/Phase-0 CI and completed controlled-Docker positive/negative acceptance. A fresh real node bundle has also been replayed through the merged EvaluatorPlan metadata-only verifier path. The remaining blocker for #91 is intentionally a separate human/reviewer inspection of the exact-head runtime evidence.
The highest-value sequence is now:
1. protect main in GitHub settings
2. obtain independent human review of PR #91 exact head
3. integrate PR #91 unchanged
4. connect the real node behind the merged two-attempt orchestrator
5. render/replay that real multi-attempt run through the merged Evidence Report layer
6. freeze 5–10 real repository benchmark tasks
7. measure diversity, verifier correlation, backpressure, cost, and human attention on real evidence
8. only then increase autonomy, fan-out, federation, or community reproduction
Only PR #91 remains open at this audit.
PR #91 exact head:
520ad2c9aa5825476de4957da4702d6823f4edb3
Observed state:
The project should not manufacture independence by treating the same repository owner, proposing automation, or acceptance harness as a separate reviewer.
The following previously competing/stale paths have been replaced by clean current-main convergence PRs and integrated:
Old superseded PRs should remain public provenance rather than being revived.
Public branch metadata still reports:
protected: false
protection.enabled: false
required_status_checks.enforcement_level: off
This is now the largest governance gap.
Repository-side safeguards are substantially stronger than the external integration boundary. No additional autonomous repository-write authority should be granted until GitHub rulesets/branch protection actually enforce the project invariants.
Minimum external controls remain:
ACE should remain fail-closed while this is unresolved.
The repository has executable evidence for:
canonical WorkUnit
-> real Docker-isolated node execution
-> ResultManifest + patch/log evidence
-> verifier-owned EvaluatorPlan
-> metadata-only independent patch/log verification
-> canonical VerificationResult
-> explicit human integration decision remains required
This is a meaningful milestone: worker execution and independent verification now exist as separate trust boundaries on real runtime output.
The missing end-to-end demonstration is now narrower:
one WorkUnit
-> two isolated real node attempts
-> independent verification per completed candidate
-> preserve worker/verifier failures and disagreements
-> combined non-selecting Evidence Report
-> deterministic/semantic replay from saved metadata
-> human decision outside worker/verifier authority
Do this before adding sophisticated routing or a large benchmark inventory.
After the two-real-attempt path is stable, add one deliberately simple second real adapter. The adapter should prove coordinator neutrality, not maximize model sophistication.
Only then prioritize mini-SWE-agent/OpenHands/A2A/MCP integrations as product-critical work.
The verification architecture is currently one of the strongest repository areas.
Preserve these invariants:
The verification-backpressure and correlated-evidence work is mainly synthetic today. Once real multi-attempt runs exist, collect measured:
Use those measurements to replace hand-authored controller priors.
The mathematical foundation is broad enough for the current phase. Do not add formulas merely for completeness.
Operationalize the existing model in this order:
Every promoted metric should specify:
The ACE stack is now structurally coherent:
GitHub activity
-> trusted cohort/exposure observation
-> explicit causal lineage receipts
-> recoverable live review-capacity measurement
-> shadow generational controller
-> external conjunctive activation gate
-> bounded future actuator only if all evidence/authority gates pass
Current interpretation should remain conservative:
capacity recovered != permission to act
activity != verified descendant
infrastructure merged != community reproduced
repository rules != GitHub-enforced protection
Do not increase Cohort-2 or Phase-B write activity merely because open-work pressure falls. Require external-participant and verified-descendant evidence.
The root remains documentation-heavy. This is a real navigation cost, but it is not the highest-value blocker while the first real product loop is one integration step from completion.
Recommended sequence:
CONSTITUTION.md, governance, human-flourishing constraints, mathematical foundations, evolution model, and project goals should remain easy to discover from the root/README even if their canonical bodies later move under docs/.
Open issues should increasingly be treated as one of four classes:
Avoid opening new architecture/research issues unless they either unblock the real runner, create measurable evidence, or provide a high-quality community contribution surface.
Close/supersede issues when their canonical deliverable has landed rather than letting historical checklist language make the tracker appear less converged than the repository is.
Protect main. This cannot be replaced by repository documentation or workflows.
Independent human review of PR #91 exact head. If accepted and unchanged, mark ready and merge.
Real two-attempt orchestration + independent verification + Evidence Report/replay. This is the shortest path to an honest v0.1 claim.
Freeze 5–10 real tasks only after the complete real multi-attempt loop works.
Replace synthetic/hand-authored priors with measured runtime, verification, queue, and human-attention data.
Obtain the first external verified descendant and measure the full contribution funnel before increasing reproduction.
Heterogeneous adapters, official A2A/MCP conformance, larger R1/R2/R3/R4 experiments, structural migrations, federation, stronger autonomous routing.
The next important repository iteration is not another formula, workflow, or simulation.
It is:
A separately reviewed canonical real node lands, two real isolated attempts run through the existing orchestrator, each completed candidate receives independent verification, the combined run is rendered/replayed through the Evidence Report layer, and the human integration decision remains external.
That iteration directly improves product quality, verification evidence, architecture convergence, contributor clarity, and the credibility of the IDKMesh thesis at the same time.