The SP-TRACE-CFR narration-fidelity verifier is built, tested (270 tests), and merged to main — two-engine recovery, the sp-run narration-gate / attest-run surface, signed run attestations, a live reference runtime, and a real Claude ModelPolicy (PRs #311, #312, #313, #314; sourceos-spec #131). What remains is blocked on resources not available in the build environment.
1. Live end-to-end smoke against a real Anthropic key ⛔ needs ANTHROPIC_API_KEY / ant auth
ModelPolicy (tools/model_policy.py) drives the runtime via a dependency-injected Anthropic client; tests use a scripted fake (no live API). We have not run it against a real model turn.
2. Wire production emitters to emit control-flow ReasoningEvents ⛔ other repos / possibly other lane
sourceos-spec #131 added the controlFlow vocabulary. The real surfaces (TurtleTerm / Noetica / BearBrowser / the governed runner) don't emit these events yet.
3. Deploy hardening ⛔ needs a real secret + CI decision
4. Mellumwork + CBES doctrine grounding ⛔ needs source docs from @mdheller
tools/mellumwork.py grounds the T1/T2 tiering we actually use, but the named external frameworks weren't locatable. CBES's 7 axioms were never provided (not fabricated).
5. Composition primitives blocked on other infra
Doable here without any of the above (not blocked; separate follow-ups if wanted): ReceiptIR base fold additive edit; R_PI 3rd engine; eval strata S2/S3/S6; conformal anytime-valid/PASC.
The SP-TRACE-CFR narration-fidelity verifier is built, tested (270 tests), and merged to
main— two-engine recovery, thesp-run narration-gate/attest-runsurface, signed run attestations, a live reference runtime, and a real ClaudeModelPolicy(PRs #311, #312, #313, #314; sourceos-spec #131). What remains is blocked on resources not available in the build environment.1. Live end-to-end smoke against a real Anthropic key ⛔ needs
ANTHROPIC_API_KEY/ant authModelPolicy(tools/model_policy.py) drives the runtime via a dependency-injected Anthropic client; tests use a scripted fake (no live API). We have not run it against a real model turn.default_client()+AgentRuntime.run_reacton a real task; confirm control-flow events emerge from a genuineclaude-opus-4-8turnattest()gates the model's real narration (truthful → PASS, induced mismatch → FAIL)2. Wire production emitters to emit control-flow ReasoningEvents ⛔ other repos / possibly other lane
sourceos-spec #131 added the
controlFlowvocabulary. The real surfaces (TurtleTerm / Noetica / BearBrowser / the governed runner) don't emit these events yet.reasoning.tool.called/decision.branched/subrun.spawned|joined/run.completedfrom at least one real emittertrace_cfr_reasoning_bridge.py); once an emitter produces the events, gating runs on real runs with zero further change.3. Deploy hardening ⛔ needs a real secret + CI decision
--key-seed/ env)attest-run4. Mellumwork + CBES doctrine grounding ⛔ needs source docs from @mdheller
tools/mellumwork.pygrounds the T1/T2 tiering we actually use, but the named external frameworks weren't locatable. CBES's 7 axioms were never provided (not fabricated).5. Composition primitives blocked on other infra
Doable here without any of the above (not blocked; separate follow-ups if wanted): ReceiptIR base
foldadditive edit; R_PI 3rd engine; eval strata S2/S3/S6; conformal anytime-valid/PASC.