Agentic Development · Governance snapshot

The governance runs, acts, and records — reproducibly.

This snapshot measures how the Phionyx self-governance runtime governs the AI that writes Phionyx itself — aggregate numbers only, reviewer-reproducible. It is refreshed when the maintainer re-runs the public audit script (manual, not a continuous feed); the snapshot date is shown below. What is already proven: every governed decision is signed into a tamper-evident, hash-chained record, and the gate refuses, regenerates, or rewrites work that fails its checks — it acts, it does not only log. What is not yet proven: the full closed loop — a signed decision whose real-world outcome is later confirmed — which is honestly near zero today and accumulates forward. Coverage is one input, not the goal.

Snapshot generated 2026-06-16 12:51 UTC · rolling 30-day window · regenerated when the maintainer re-runs the public audit script (manual), not on a fixed schedule.

What the runtime proves today

Signed records (live)
476

tamper-evident and hash-chained within the runtime; the public feed publishes aggregate counts, not the records themselves (10 chains).

Gate refusals
100

the gate refused / regenerated / rewrote work that failed its checks — it acts, it does not only log.

Open method
public

the deterministic measurement script is public; these numbers are its output over the private dev telemetry, and the published snapshot is the audit artifact.

What we don’t claim yet: that a signed decision’s real-world outcome has been independently confirmed end to end. Confirming outcomes takes time and real labels (which we never fabricate), so that long-loop evidence is still accruing. We’d rather show that gap than hide it.

Invocation discipline · rolling 30-day window (one input, not the headline)

47.5%coverage · 344 of 429 commits instrumented · 327 gate calls (expected 688)

Our own rule asks the AI to call the gate twice per commit, so “coverage” is simply how often it did. It is a discipline check, not the point — calling the gate never proves the work is correct (that is what the signed record above is for). Reading the per-day Coverage column:

  •  no governed activity recorded that day — not measured, and not counted as a failure.
  • 0.0%  the gate was available but the AI did not call it.
  • 100%+  the AI called the gate more often than the twice-per-commit baseline — extra discipline, not an error.
Commits
429
in window
Gate calls
327
verify_claim + response_gate only
Expected
688
= 344 instrumented × 2
Sessions
154
MCP sessions in window

Per-day breakdown (30 days)

since 2026-05-17 15:51 UTC

DateCommitsGateAll MCPCoverage
2026-06-1641223100%+
2026-06-1561938100%+
2026-06-141865123100%+
2026-06-1312237100%+
2026-06-12400
2026-06-11200
2026-06-10154813.3%
2026-06-09113264100%+
2026-06-08112441100%+
2026-06-0736142419.4%
2026-06-0620162040.0%
2026-06-052010.0%
2026-06-042030.0%
2026-06-031070.0%
2026-06-0210120.0%
2026-06-0111450.0%
2026-05-31100
2026-05-30500
2026-05-2924101020.8%
2026-05-2867515138.1%
2026-05-2724363675.0%
2026-05-2640152218.8%
2026-05-2534252.9%
2026-05-24264167.7%
2026-05-232700
2026-05-221700
2026-05-21600
2026-05-20300
2026-05-191800
2026-05-18200

Directive distribution

The aggregate of every directive written to telemetry across MCP tools + observability hooks — broader than the gate's five directive dispositions (pass · damp · rewrite · regenerate · reject) because non-gate tools also write entries.

auto_attest: 2942pass: 318n/a: 103regenerate: 87require_tool: 15checkpoint: 8rewrite: 8reject: 5hedge: 1ok: 1

auto_attest — observability hooks (session start, user prompt log, commit attestation, …). Recorded but excluded from coverage math by design.

checkpoint — emitted by phionyx_checkpoint (lightweight context-note tool, not a gate).

n/a — entries without an explicit directive field (e.g. session start, prompt log, attestation writes).

solid / weak / incomplete — emitted by phionyx_causal_trace (chain-quality verdicts, distinct from the gate's five directive dispositions).

ok — emitted by the trust-boundary MCP server on successful third-party-tool attestation writes.

Reproduce

python3 case-studies/agentic-development-2026-05/scripts/runtime_evidence_self_audit.py --days 30

Same audit script lives at case-studies/.../scripts/runtime_evidence_self_audit.py. Generated 2026-06-16 12:51 UTC · schema v0.7.2 · regenerated on maintainer re-run (snapshot date above).

Agentic Development · Reproducible governance snapshot — runtime evidence