docs(prd): MIRA × FactoryLM Phase 1 — read-only machine evidence handoff by Mikecranesync · Pull Request #3048 · Mikecranesync/MIRA

Mikecranesync · 2026-08-02T03:48:14Z

Adds the Phase-1 PRD for handing FactoryLM's canonical PLC snapshot to MIRA's existing audited technician context as read-only live evidence.

Docs-only (one file, +291). No code, no migration, no dependency — exempt from the VERSION bump per docs/versioning.md.

Premise verification (I checked every external claim before opening this)

Branch is current with origin/main (0 behind; merge-base cb49901a = #3041). Every referenced symbol and file resolves on origin/main:

PRD claim	Verified
`materialized_evidence/context_contract.py` — `LiveTag`, `LiveStateOverlay`, `live_overlay_from_machine_packet`	✅ present
`TechnicianContext.live: LiveStateOverlay \| None = None`	✅ present, and `to_dict()` serializes it
`mira-bots/shared/technician_context.py` — `build_turn_context`, `manifest_of`	✅ present
`mira-bots/shared/engine.py` — `_build_prior_decisions_context`, `_context_manifest`, `_build_live_data_context`	✅ 2 / 9 / 2 hits
`mira-bots/shared/live_snapshot.py`	✅ present
`build_turn_context` has no live parameter today	✅ confirmed — PR 1's premise is accurate
FactoryLM #188 (draft), #161 (open), #192 (open); MIRA #3046 (open, DIRTY), #3038 (draft), #2989 (open)	✅ all live and accurately described
`/Users/charlienode/factorylm` is dirty/user-owned	✅ confirmed — on `feat/mission-control-demo-pipeline`, ahead 2, modified files. Use a fresh worktree as the PRD says.

Preflight step 5/6 is now answered: an authorized ingress exists

The PRD says to stop after the contract/fixture PRs if no existing ingress can carry this payload. It can:

POST /api/v1/tags/ingest → mira-relay/relay_server.py::tags_ingest
HMAC auth via _authenticate_http; the HMAC tenant is authoritative and a caller-supplied body tenant is only honored on the non-HMAC dev/bench path — exactly the "never default or infer tenant_id" rule the contract asks for
→ mira-relay/tag_ingest.py::ingest_batch, fail-closed against approved_tags

This is also mandatory, not merely available: .claude/rules/one-pipeline-ingest.md forbids any transport from forking its own normalizer/allowlist/persistence/batch-shape. So PR 3 is buildable, and a second relay would be a rule violation rather than a design choice.

Four gaps found against the real ingest contract

These are not blockers, but PR 3 cannot be built as literally specified without resolving them.

source_system: "factorylm-plc-modbus" will be rejected. ingest_batch validates against VALID_SOURCE_SYSTEMS = {"ignition", "plc_bridge", "relay", "simulator"} and raises invalid_source_system otherwise. Reuse-first answer: send plc_bridge (semantically exact) and carry the FactoryLM identity in provenance — rather than widening the vocabulary at the single enforcement point.
The PRD never mentions seeding approved_tags — without it every tag is dropped. The allowlist is fail-closed on normalize_tag_path(tag_path). All seven canonical tags need rows (conv_simple.vfd_speed_hz → conv_simple_vfd_speed_hz, etc.) for the tenant + source_system, or ingest_batch returns not_allowlisted for all of them and the overlay is silently empty. Precedent exists: tools/seeds/approved_tags_conveyor.sql, gen_approved_tags_simulator.py, migration 035_approved_tags.sql. This is a missing work package.
Envelope fields have no home in the canonical batch shape. build_ingest_batch carries {source_system, tags[], tenant_id?, source_connection_id?} and build_tag_entry carries {tag_path, value, value_type, quality, ts?, equipment_entity_id?, metadata?}. The envelope's machine_state, active_conditions, snapshot_id, captured_at, schema_version, provenance, proposed_uns_path have nowhere to go — yet LiveStateOverlay requires machine_state and active_conditions. They ride in metadata or the design needs an explicit decision. This is the most consequential unresolved question in the doc.
The relay persists; it does not carry a request-scoped snapshot. ingest_batch writes tag_events + live_signal_cache. So PR 4 cannot "receive the snapshot inline at turn time" — it must read back current state at turn time. Worth pinning explicitly, since the current wording ("wire the accepted snapshot from the established ingress/state carrier") permits both readings.

One design note

_build_live_data_context is not relay-based — it HTTP-GETs mira-fault-detective /current_fault. So the PRD's duplicate-[LIVE EQUIPMENT STATUS] concern is real: these are two genuinely independent live paths, and PR 4's dedup requirement is load-bearing.

Also: #3041 added retrieval evidence via a separate augment_with_retrieval(ctx, chunks) rather than extending build_turn_context. PR 1 proposes extending build_turn_context instead. Matching the augment_with_* precedent would be a smaller diff and a consistent shape — worth a decision before PR 1 starts.

Scope

Doc only. No code was written and nothing was built against this PRD yet.

…ntract Verified every claim against origin/main before any code was written. Four corrections, each inlined and marked: - source_system 'factorylm-plc-modbus' would have been rejected outright — ingest_batch validates VALID_SOURCE_SYSTEMS {ignition, plc_bridge, relay, simulator} (tag_ingest.py:59). Use plc_bridge; FactoryLM identity moves to provenance.producer. - The envelope->canonical-batch mapping was unspecified, and machine_state / active_conditions had nowhere to go in build_tag_entry despite being required to construct a LiveStateOverlay. They now ride metadata.factorylm_snapshot, with the two quality vocabularies mapped explicitly (downgrade only). - Seeding approved_tags is a PREREQUISITE of PR 3, not a follow-up. The allowlist is fail-closed, so without it a valid snapshot returns accepted=0 and every acceptance test passes vacuously against an empty overlay. - PR 4 reads state back at turn time. ingest_batch persists to tag_events + live_signal_cache; it never hands a request-scoped snapshot to the engine. Also records that the preflight's 'stop if no ingress exists' contingency does not fire (the ingress exists and is mandatory under the one-pipeline law), and recommends the augment_with_* shape #3041 established over growing build_turn_context's signature per evidence family.

Mikecranesync · 2026-08-02T05:10:33Z

🔒 Slice claim — MIRA PR 1 (contract adapter + augment_with_live + fixtures)

Who: Claude Code (Opus 4.8) session on node BRAVO (FactoryLM-Bravo.local, user bravonode)
Slice: PRD docs(prd): MIRA × FactoryLM Phase 1 — read-only machine evidence handoff #3048 PR 1 — MIRA-side factorylm.machine-snapshot.v1 → LiveStateOverlay → TechnicianContext.live, following the augment_with_* shape established in feat(spine): WS1/G6 — retrieval evidence family on the context contract (audit-only) #3041 (live-state evidence family).
Worktree (isolated): feat/factorylm-machine-snapshot-adapter off origin/main @ 5691f0d0c. No shared-checkout edits.
Not touching: PR 2 (FactoryLM repo feat: Notebook Intelligence Layer — Open WebUI + MIRA Pipeline + Docling #161/chore: update Ollama on Bravo + pull Gemma 4 E4B #188), PR 3 (ingress/seeding), PR 4 (serving path), or fix(ignition): PEP 263 cookie is a guaranteed SyntaxError in a WebDev resource — remove it, keep the live proof #3046. Pure Python only — no ingress, no infra, no plant writes.

If another session already has PR 1 in flight, ping here and I'll stand down. Otherwise I'm building it now.

Mikecranesync · 2026-08-02T05:19:02Z

🔒 Slice claim — PR 2 (FactoryLM canonical-source correctness)

Who: Claude Code (Fable 5) session f985f0fe on the PLC laptop (LAPTOP-0KA3C70H, Tailscale 100.72.2.99) — the machine with the bench Micro820 + GS10 + Ignition 8.3.4 gateway.
Slice: PRD docs(prd): MIRA × FactoryLM Phase 1 — read-only machine evidence handoff #3048 PR 2 — FactoryLM repo (Mikecranesync/factorylm): inspect/fix feat: Notebook Intelligence Layer — Open WebUI + MIRA Pipeline + Docling #161 (RegisterData strips named VFD register fields → diagnosis reads zeros), rebase + test the canonical-tag mapping from chore: update Ollama on Bravo + pull Gemma 4 E4B #188 in a clean branch, and generate factorylm.machine-snapshot.v1 from canonical tag output. Not published remotely — envelope + fixture generation only.
Checkout: fresh clone of factorylm on this machine (no shared dirty checkout touched, per the PRD's constraint).
Not touching: PR 1 (claimed by BRAVO — I will consume its fixture location once it lands, or generate against the PRD §"Contract" spec until then), PR 3 (ingress/seeding), PR 4 (serving path), PR 5, MIRA fix(ignition): PEP 263 cookie is a guaranteed SyntaxError in a WebDev resource — remove it, keep the live proof #3046 (already merged as fee69d1b4 btw — the "until fix(ignition): PEP 263 cookie is a guaranteed SyntaxError in a WebDev resource — remove it, keep the live proof #3046 is resolved" precondition in PR 5's DoD is now satisfied), and nothing write-capable toward the PLC (read-only per .claude/rules/fieldbus-readonly.md).
Bench note: this machine can run read-only sanity against the real Micro820/GS10 if useful, but PR 2's acceptance tests are deterministic/offline per the PRD.

If another session already has PR 2 in flight, ping here and I'll stand down. Otherwise building now.

Mikecranesync · 2026-08-02T10:21:17Z

✅ PR 2 slice code-complete (PLC laptop, session f985f0fe)

factorylm#197 — feat: Notebook Intelligence Layer — Open WebUI + MIRA Pipeline + Docling #161 fixed (RegisterData placeholder fields dropped every named VFD value; drift-guard test pins the model to REGISTER_NAMES). 154 tests.
factorylm#198 — factorylm.machine-snapshot.v1 producer: chore: update Ollama on Bravo + pull Gemma 4 E4B #188 rebased onto main (cherry-pick, authorship preserved — close chore: update Ollama on Bravo + pull Gemma 4 E4B #188 when docs: auto-research / continuous eval loop design #198 merges) + machine_snapshot.py (source_system=plc_bridge, comm-loss-beats-fault state derivation, downgrade-only quality, no clocks, observation-only validator) + the shared fixtures from feat(spine): FactoryLM machine-snapshot → TechnicianContext.live adapter (PRD #3048, PR 1) #3052 vendored verbatim into contracts/machine_snapshot/ so both repos test identical payloads. 180 tests (24 new).

@BRAVO/#3052: if the fixtures change in your review, ping here — factorylm re-syncs its vendored copy and re-runs test_machine_snapshot.py. Producer output is asserted shape-identical to your golden payload.

Not done (per PRD, not mine): remote publishing (PR 3, unclaimed — includes the fail-closed approved_tags seeding), PR 4, PR 5 (integration proof wants this bench; I can take it once 1–4 land). Both factorylm PRs await Mike's review/merge.

…ter (PRD #3048, PR 1) (#3052) First slice of the read-only machine-evidence handoff: the live-state evidence family enters the context contract, after prior-decision (#3032) and retrieval (#3041). Pure Python — no ingress, no infra, no plant writes. - overlay_from_factorylm_snapshot(snapshot) -> (LiveStateOverlay | None, violations): validates the factorylm.machine-snapshot.v1 envelope and maps it to the MachineContextPacket dict live_overlay_from_machine_packet already consumes, then delegates — no duplicated LiveTag/freshness/render logic. Quality {good,bad,stale,uncertain} -> Freshness, always downgrading toward less confidence (unknown never becomes good/live); simulator source marks tags simulated. Read-only by construction (asserted: no pymodbus/pycomm3/socket/ httpx/write). - technician_context.augment_with_live(ctx, snapshot): folds the overlay into TechnicianContext.live on the SAME context (augment_with_* shape from #3041), re-validates, one manifest. Accepts an envelope dict or a prebuilt overlay. Fail-open. - contracts/machine_snapshot/: shared cross-repo fixture (valid + 4 invalid + spec) — the compatibility boundary both repos test against. - Exported overlay_from_factorylm_snapshot + FACTORYLM_SNAPSHOT_SCHEMA. Flag-gated serving-path wiring is PR 4; this changes no engine behavior. Tests: test_factorylm_snapshot_adapter.py (9) + test_technician_context_live.py (6). Regressions green (context_contract, technician_context, ws1_engine_wiring, retrieval). Built by: Claude Code (Opus 4.8) on node BRAVO (FactoryLM-Bravo.local, user bravonode). Co-authored-by: Mike Harper <bravonode@FactoryLM-Bravo.local> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Mikecranesync · 2026-08-02T11:22:51Z

🔒 Slice claim — PR 2 repair pass (PLC laptop LAPTOP-0KA3C70H, fresh Claude Code Fable 5 session)

Slice: adversarial-review repair on factorylm docs: scope mira-web → mira-pipeline cutover (ADR-0008 Phase 4) #197/docs: auto-research / continuous eval loop design #198 per CLAUDE_PR2_REPAIR_BRIEF.md — factorylm issues fix: guarantee PIPELINE_CALL log lines emit from uvicorn-managed process #199–feat: continuous eval — Celery task + /eval/latest dashboard #205 (fixture parity, comm_down vocab, validator hardening, lazy imports, e-stop precedence, register docs, PLC CI).
Not touching: PR 3 (relay publishing / approved_tags), MIRA-side adapter (feat(spine): FactoryLM machine-snapshot → TechnicianContext.live adapter (PRD #3048, PR 1) #3052), no merges, no deploys, no prod config.
Continues code-complete slice from session f985f0fe (same machine).

Mikecranesync · 2026-08-02T11:35:32Z

✅ PR 2 repair pass complete (PLC laptop, per CLAUDE_PR2_REPAIR_BRIEF.md)

All seven adversarial-review findings repaired:

factorylm#198 commit 2926e38 — fix: guarantee PIPELINE_CALL log lines emit from uvicorn-managed process #199 (seven-tag fixtures + path-set test), Fix Open WebUI P0 UX: continue/regenerate/PDF routing #200 (comm_down), feat: auto-research eval loop — Week-1 MVP #201 (validator: source_system + canonical paths + quality-downgrade acceptance), docs: v0.5.4 release notes and rollback anchor #202 (lazy __init__ — import factorylm_plc.machine_snapshot provably pulls no pymodbus), bug: cross-vendor hallucination — GS20 query returns Allen-Bradley PowerFlex content #203 (estopped outranks running; comm_down still dominant). 188 unit tests pass (was 180).
factorylm#197 commit 59abf9a — bug: no honesty signal for out-of-KB equipment (Yaskawa V1000) #204 (CLAUDE.md registers 101-105 documented) + the optional feat: Notebook Intelligence Layer — Open WebUI + MIRA Pipeline + Docling #161 IOResponse serialization acceptance test. 155 pass (was 154).
factorylm#206 (new) — feat: continuous eval — Celery task + /eval/latest dashboard #205 CI: plc-modbus-tests.yml runs the unit suite on PRs; the workflow ran on its own PR and passed (unit ✅ 34s).
MIRA#3058 (new) — fixture-sync half of fix: guarantee PIPELINE_CALL log lines emit from uvicorn-managed process #199: MIRA's contracts/machine_snapshot/ now byte-identical to factorylm's seven-tag copies; adapter freshness count 5→6 live. 15 consumer tests pass; CI green so far.

Merges remain Mike's call (#198 → close #188 on merge, per standing instruction). No relay publishing, no approved_tags seeding, no plant writes — PR 3 scope untouched.

… ingress (PRD #3048, PR 3) PR 1 (v3.240.0) landed the envelope -> TechnicianContext.live adapter; this lands the transport that actually delivers a snapshot. No new endpoint exists or could: the one-pipeline law forbids a transport from defining its own normalizer, allowlist, persistence, batch shape, or enforcement path, and test_architecture Contract 5 fails the build on a violation. So the work is decode -> build_tag_entry -> build_ingest_batch -> ingest_batch, the shape simlab/publishers.py::RelayIngestPublisher established. - mira-relay/factorylm_snapshot.py: snapshot_to_ingest_batch() (pure decode) + FactoryLMSnapshotPublisher (HMAC-signed POST to POST /api/v1/tags/ingest). Snapshot-scoped fields with no column of their own ride per-tag metadata.factorylm_snapshot; machine_state and active_conditions are REQUIRED to build a LiveStateOverlay, so losing them silently would give PR 4 a permanently "unknown state" overlay while every tag looked healthy. value_type is derived per tag (bool before int -- isinstance(True, int) is True). source_system is plc_bridge; FactoryLM identity rides provenance.producer because VALID_SOURCE_SYSTEMS rejects factorylm-plc-modbus. The envelope's tenant_id is never read -- X-MIRA-Tenant is authoritative and the signed body omits it. - tools/seeds/approved_tags_factorylm_conv_simple.sql: the 7 canonical conv_simple.* tags for source_system='plc_bridge'. A PREREQUISITE, not a follow-up: the allowlist is fail-closed with no permissive mode, so without the seed a valid snapshot is accepted with accepted=0, every tag not_allowlisted, nothing stored -- HTTP 200, no error, and every downstream check passes vacuously against an empty overlay. test_unseeded_allowlist_rejects_every_tag pins that failure mode explicitly. UNS identity comes from the seeded uns_path, never the envelope's proposed_uns_path (provenance only). - mira-relay/auth.py: added sign_hmac_headers() and factored the signed string into one _signed_string() that both verify_hmac (server) and the signer (client) call, so the two cannot drift. Four producers already inline their own copy; rather than add a fifth, the new publisher uses the shared helper and a test round-trips its headers through the real verify_hmac. Migrating the existing four call sites is out of scope here. - Runbook: staging-first validation with the seed as step 1 and "accepted=0 rejected=N" called out as the headline failure mode. Read-only throughout: a snapshot is observation data; no fieldbus client is imported, no command/actuator field is honored, no plant write occurs. Tests: mira-relay/tests/test_factorylm_snapshot.py (28). mira-relay suite 191 -> 219 passed, zero regressions; test_architecture 13 passed (Contract 5 green -- the new module is NOT added to _ONE_PIPELINE_ALLOWLIST); PR 1's 15 tests still green. Pre-existing on origin/main and untouched here: test_document_compiler determinism (1 failure) and 84 ruff errors in files this PR does not modify. The seed has been applied to NO environment. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0191hTx4dwKR45av8njFyEPz

… ingress (PRD #3048, PR 3) (#3059) PR 1 (v3.240.0) landed the envelope -> TechnicianContext.live adapter; this lands the transport that actually delivers a snapshot. No new endpoint exists or could: the one-pipeline law forbids a transport from defining its own normalizer, allowlist, persistence, batch shape, or enforcement path, and test_architecture Contract 5 fails the build on a violation. So the work is decode -> build_tag_entry -> build_ingest_batch -> ingest_batch, the shape simlab/publishers.py::RelayIngestPublisher established. - mira-relay/factorylm_snapshot.py: snapshot_to_ingest_batch() (pure decode) + FactoryLMSnapshotPublisher (HMAC-signed POST to POST /api/v1/tags/ingest). Snapshot-scoped fields with no column of their own ride per-tag metadata.factorylm_snapshot; machine_state and active_conditions are REQUIRED to build a LiveStateOverlay, so losing them silently would give PR 4 a permanently "unknown state" overlay while every tag looked healthy. value_type is derived per tag (bool before int -- isinstance(True, int) is True). source_system is plc_bridge; FactoryLM identity rides provenance.producer because VALID_SOURCE_SYSTEMS rejects factorylm-plc-modbus. The envelope's tenant_id is never read -- X-MIRA-Tenant is authoritative and the signed body omits it. - tools/seeds/approved_tags_factorylm_conv_simple.sql: the 7 canonical conv_simple.* tags for source_system='plc_bridge'. A PREREQUISITE, not a follow-up: the allowlist is fail-closed with no permissive mode, so without the seed a valid snapshot is accepted with accepted=0, every tag not_allowlisted, nothing stored -- HTTP 200, no error, and every downstream check passes vacuously against an empty overlay. test_unseeded_allowlist_rejects_every_tag pins that failure mode explicitly. UNS identity comes from the seeded uns_path, never the envelope's proposed_uns_path (provenance only). - mira-relay/auth.py: added sign_hmac_headers() and factored the signed string into one _signed_string() that both verify_hmac (server) and the signer (client) call, so the two cannot drift. Four producers already inline their own copy; rather than add a fifth, the new publisher uses the shared helper and a test round-trips its headers through the real verify_hmac. Migrating the existing four call sites is out of scope here. - Runbook: staging-first validation with the seed as step 1 and "accepted=0 rejected=N" called out as the headline failure mode. Read-only throughout: a snapshot is observation data; no fieldbus client is imported, no command/actuator field is honored, no plant write occurs. Tests: mira-relay/tests/test_factorylm_snapshot.py (28). mira-relay suite 191 -> 219 passed, zero regressions; test_architecture 13 passed (Contract 5 green -- the new module is NOT added to _ONE_PIPELINE_ALLOWLIST); PR 1's 15 tests still green. Pre-existing on origin/main and untouched here: test_document_compiler determinism (1 failure) and 84 ruff errors in files this PR does not modify. The seed has been applied to NO environment. Claude-Session: https://claude.ai/code/session_0191hTx4dwKR45av8njFyEPz Co-authored-by: MIRA Beta Orchestrator <beta-orchestrator@factorylm.local> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

Mikecranesync · 2026-08-02T12:43:06Z

🏁 PR 2 slice MERGED (factorylm repo)

docs: scope mira-web → mira-pipeline cutover (ADR-0008 Phase 4) #197 merged (74a5d52) — feat: Notebook Intelligence Layer — Open WebUI + MIRA Pipeline + Docling #161 RegisterData fix + register-map docs; feat: Notebook Intelligence Layer — Open WebUI + MIRA Pipeline + Docling #161 and bug: no honesty signal for out-of-KB equipment (Yaskawa V1000) #204 closed.
docs: auto-research / continuous eval loop design #198 merged (4521224) — factorylm.machine-snapshot.v1 producer over canonical tags, incl. all repair rounds (fix: guarantee PIPELINE_CALL log lines emit from uvicorn-managed process #199–bug: cross-vendor hallucination — GS20 query returns Allen-Bradley PowerFlex content #203 + the unsourced-tag quality-downgrade P1). Issues fix: guarantee PIPELINE_CALL log lines emit from uvicorn-managed process #199–bug: cross-vendor hallucination — GS20 query returns Allen-Bradley PowerFlex content #203 closed with evidence.
feat: GET_DOCUMENTATION intent + cross-vendor contamination guard (v2.4.0) #206 merged earlier — PLC Modbus Tests CI live (feat: continuous eval — Celery task + /eval/latest dashboard #205 closed); unit check ran green on both PRs pre-merge.
chore: update Ollama on Bravo + pull Gemma 4 E4B #188 closed as superseded (cherry-picked into docs: auto-research / continuous eval loop design #198 with authorship preserved).
Post-merge verify on factorylm main: pytest tests/unit -q → 195 passed.

Still open: MIRA #3058 (fixture-sync half of #199, all 27 checks green — awaiting merge). PR 2 slice is now fully landed; PR 3 (ingress + approved_tags seeding) and PR 5 (bench integration proof) remain unclaimed.

…greed (#3060) Two gaps from reviewing the PRD #3048 PR 1-3 work. An empty tags list built a live overlay with no evidence in it. The FactoryLM producer's validate_envelope already rejects tags: [] ("must be a non-empty list"); the MIRA consumer accepted it and returned a LiveStateOverlay that still asserted machine_state -- a live block capable of telling a technician the machine is "running" with nothing behind it. The producer being stricter than the consumer is backwards: the consumer faces untrusted input. tags:empty is now a violation, so the turn degrades to "no live evidence" rather than a fabricated plant claim. The cross-repo fixture had no drift guard. contracts/machine_snapshot/ is vendored verbatim in TWO repos and the PRD calls it the compatibility boundary both projects must test the exact same payload against. It was byte-identical at review time by care, not by mechanism -- and a one-sided edit drifts them silently on BOTH sides, since each repo keeps passing against its own copy. CHECKSUMS.sha256 + test_machine_snapshot_fixture_integrity.py make that fail loudly; the guard also asserts every fixture on disk is listed, so a new unguarded fixture cannot slip in behind a green check. Proven by execution: 8 pass -> mutate a fixture -> checksum test fails with the regeneration instructions -> restore -> 8 pass. materialized_evidence/tests/ 195 passed + 8 new; ruff clean. No behavior change for any populated snapshot. Co-authored-by: MIRA Beta Orchestrator <beta-orchestrator@factorylm.local>

Wire the read-only FactoryLM live overlay into the ONE technician context path. The snapshot is READ BACK at turn time from the persisted state carrier (live_signal_cache) for the confirmed asset — never threaded inline from the ingress, since ingest_batch persists rather than handing the engine a request-scoped object. Additive, OFF by default (MIRA_FACTORYLM_LIVE); no behavior change until enabled. - mira-bots/shared/factorylm_live.py: fetch_live_signal_cache(tenant, prefix) reads current cache rows scoped to the asset UNS subtree (uns_path <@ ltree, tenant-scoped), mirroring ctx_enrichment; overlay_from_cache_rows(rows) builds a LiveStateOverlay — pure + deterministic. Freshness/timestamps come from the stored row, never now() (observed_at is the absolute last_seen_at), so two reads of the same rows produce a byte-identical overlay + stable manifest hash. simulated -> SIMULATED (never real telemetry); stale -> STALE (not dropped). - Supervisor._build_factorylm_live_overlay reads the overlay back; the serving seam folds it into turn_ctx via augment_with_live so ONE manifest carries it, and renders [LIVE MACHINE STATE (FactoryLM)] ONLY when the fold succeeded — the prompt can never show live evidence the manifest lacks. When present it SUPERSEDES the legacy [LIVE EQUIPMENT STATUS] block (dedup). - technician_context.live_prompt_block(overlay) renders only the live overlay. Flag-gated + fail-open: any intake/contract error returns an ordinary answer without live evidence. Read-only. Gated under contract_enabled() too, since the overlay only reaches prompt+manifest via the contract's turn_ctx. Tests: mira-bots/tests/test_factorylm_live_serving.py (14). Regressions green: context_contract 27, technician_context/live/retrieval 54, ws1_engine_wiring 13, materialized_evidence 185, architecture 13, legacy live wiring 15. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Mikecranesync · 2026-08-02T13:11:44Z

🔒 Slice claim — PR 5 (controlled integration proof) (PLC laptop LAPTOP-0KA3C70H, Claude Code Fable 5 session)

Slice: PRD docs(prd): MIRA × FactoryLM Phase 1 — read-only machine evidence handoff #3048 PR 5 — the verification runbook + executable proof harness for the 7-point proof (simulated snapshot → ingress accept → TechnicianContext.live → prompt/manifest agreement → cited answer → fail-safe controls → no-write assertion). Analysis-first: PR 3 (ingress transport + approved_tags seeding) and PR 4 (serving path) are not yet built, so the deliverable will prove everything provable in-process today and precisely document what remains blocked on PRs 3/4.
Not touching: relay publishing to deployed envs, prod/staging seeding, engine serving-path implementation (PR 4's scope), no merges without Mike.

Mikecranesync · 2026-08-02T13:20:54Z

✅ PR 5 slice built — MIRA #3062 (PLC laptop)

The verification slice per the PRD ("a separate verification PR/runbook update, not an implementation bundle"):

Harness tests/integration/test_machine_evidence_proof.py — all 7 PRD proof points chained in-process through the REAL merged modules (PR 3 transport → fail-closed ingress over the real seed file → PR 1 adapter → one manifest). 20 passed + 3 skipped on main; the 3 served-path cases importorskip(shared.factorylm_live) and activate when PR 4 feat(spine): FactoryLM live machine-state serving path (PRD #3048, PR 4) #3061 merges — full chain already verified 23/23 green against a scratch-merge of feat(spine): FactoryLM live machine-state serving path (PRD #3048, PR 4) #3061.
Runbook docs/runbooks/factorylm-machine-evidence-integration-proof.md — proof matrix, dependency table, and the supervised staging live-probe procedure (seed approved_tags via apply-approved-tags.yml → MIRA_FACTORYLM_LIVE=1 on staging → bench HMAC publish from THIS laptop's real tick() → persistence check → cited bot answer → control cases). Live probe is gated on feat(spine): FactoryLM live machine-state serving path (PRD #3048, PR 4) #3061 merging + Mike supervising; nothing claimed as production-proven.

Merge order: #3061 (PR 4) → #3062 (PR 5) → then the supervised live probe. Remaining unclaimed: staging seed application + the probe itself (needs Mike).

… 4) (#3061) * feat(spine): FactoryLM live machine-state serving path (PRD #3048, PR 4) Wire the read-only FactoryLM live overlay into the ONE technician context path. The snapshot is READ BACK at turn time from the persisted state carrier (live_signal_cache) for the confirmed asset — never threaded inline from the ingress, since ingest_batch persists rather than handing the engine a request-scoped object. Additive, OFF by default (MIRA_FACTORYLM_LIVE); no behavior change until enabled. - mira-bots/shared/factorylm_live.py: fetch_live_signal_cache(tenant, prefix) reads current cache rows scoped to the asset UNS subtree (uns_path <@ ltree, tenant-scoped), mirroring ctx_enrichment; overlay_from_cache_rows(rows) builds a LiveStateOverlay — pure + deterministic. Freshness/timestamps come from the stored row, never now() (observed_at is the absolute last_seen_at), so two reads of the same rows produce a byte-identical overlay + stable manifest hash. simulated -> SIMULATED (never real telemetry); stale -> STALE (not dropped). - Supervisor._build_factorylm_live_overlay reads the overlay back; the serving seam folds it into turn_ctx via augment_with_live so ONE manifest carries it, and renders [LIVE MACHINE STATE (FactoryLM)] ONLY when the fold succeeded — the prompt can never show live evidence the manifest lacks. When present it SUPERSEDES the legacy [LIVE EQUIPMENT STATUS] block (dedup). - technician_context.live_prompt_block(overlay) renders only the live overlay. Flag-gated + fail-open: any intake/contract error returns an ordinary answer without live evidence. Read-only. Gated under contract_enabled() too, since the overlay only reaches prompt+manifest via the contract's turn_ctx. Tests: mira-bots/tests/test_factorylm_live_serving.py (14). Regressions green: context_contract 27, technician_context/live/retrieval 54, ws1_engine_wiring 13, materialized_evidence 185, architecture 13, legacy live wiring 15. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * style(spine): ruff format the PR-4 serving-path additions Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(spine): preserve FactoryLM snapshot evidence at serving * fix(spine): narrow FactoryLM snapshot metadata --------- Co-authored-by: Mike Harper <bravonode@FactoryLM-Bravo.local> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: MIRA Beta Orchestrator <beta-orchestrator@factorylm.local>

…runbook The verification slice: chains the real merged modules in-process (fixture → PR 3 transport → fail-closed ingress over the real seed → PR 1 adapter → augment_with_live → one manifest) across all seven PRD proof points. Served-path cases importorskip shared.factorylm_live so they activate when PR 4 (#3061) merges — verified 23/23 green against a scratch-merge of #3061; 20 passed + 3 skipped on today's main. Runbook documents the proof matrix, dependency state, the supervised staging live-probe procedure, and the deltas the proof must not paper over. Hermetic: no network, no DB, no clocks. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dgej4vU7mMnYTmU6QXsH3n

test(spine): PRD #3048 PR 5 — controlled integration proof harness + runbook

Mikecranesync · 2026-08-02T15:38:12Z

🏁 PR 5 MERGED — #3062 (0cc22b4a3, v3.242.1). PRD #3048 implementation arc is now fully landed: PR 1 #3052 → fixtures #3058 → PR 2 (factorylm #197/#198) → PR 3 #3059 (+#3060) → PR 4 #3061 → PR 5 #3062.

In-process proof runs 24/24 with zero skips on main — the served-path cases picked up the merged (hardened) PR 4 reader: identity-strict metadata.factorylm_snapshot requirement, source event_timestamp (never receipt time), machine_state served from stored metadata, generic cache rows never relabeled as FactoryLM evidence.

Remaining: the supervised live probe (runbook docs/runbooks/factorylm-machine-evidence-integration-proof.md). Two environment facts to decide on first, both verified today:

Staging runs no relay — probe design runs the relay locally on the PLC laptop against the staging Neon branch.
Staging bot tenant is the slug staging but this whole path is UUID-keyed (approved_tags, live_signal_cache casts) — the probe window needs MIRA_TENANT_ID=<UUID> on the staging bot (e.g. the stg quickstart tenant 78917b56-…). Mike's call, since staging KB scoping keyed to the slug may be affected.

Nothing is claimed production-proven until the probe passes under supervision.

…NS resolution + QR certification Live-probe finding (PRD #3048 PR 5): the vendor/model resolver returns uns_path=None for equipment names like CV-101, and nothing else populated uns_context.uns_path — so the FactoryLM live overlay's gate failed on EVERY real chat turn AND every QR deep-link turn. PR 4's engine tests mocked resolve_uns_path, which hid this. - factorylm_live.uns_prefix_for_asset(tenant, *candidates): tenant-scoped cmms_equipment lookup (exact equipment_number first, then unambiguous description match only) — the same physical-subtree identity source the relay allowlist seed and the QR deep-link use. Fail-open, read-only. - engine._build_factorylm_live_overlay: falls back to that lookup when the turn carries no uns_path (tag first, display label second). - telegram/start_command: a QR scan now seeds the certified uns_context (source=direct_connection) with cmms_equipment.uns_path, per .claude/rules/direct-connection-uns-certified.md — previously it seeded only asset_identified/asset_tag, so even QR turns failed the gate. - tools/factorylm_live_soak.py: N-iteration deterministic soak of the real engine seam + context fold (Celery-registrable callable), target 100%. - tools/seeds/staging-cv101-probe.sql: idempotent staging CV-101 row so the probe subtree resolves (prod already has the garage bridge seed). 30 tests pass (test_factorylm_live_serving 24, test_start_command 6). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dgej4vU7mMnYTmU6QXsH3n

…Rs 1-4 merged, PR 5 on bench (#3070) Co-authored-by: Mike Harper <bravonode@FactoryLM-Bravo.local> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…n as live Found by crossing the seam between PR 3 (#3059) and PR 4 (#3061) of PRD #3048 with a degraded tag -- something neither side's suite did, because each was self-consistent on its own. The shared fixture marks conv_simple.height_sensor_mm as quality="stale". Through the DEPLOYED path (ingest -> live_signal_cache -> read-back) the overlay returned it as Freshness.LIVE with freshness_summary {live: 7} -- zero stale. A technician was shown, as current, a reading the producer had already flagged. Cause -- two correct decisions that nobody joined. tag_ingest.persist_batch stamps freshness_status = 'simulated' if simulated else 'live', which is right: that column means COLLECTOR liveness, and a tag whose value is stale is still arriving on time. factorylm_live._freshness_for then read ONLY that column, never latest_quality -- which correctly held "stale" on the very same row. Neither module was wrong in isolation, which is exactly why both suites stayed green. Fix: _freshness_for now takes latest_quality and downgrades on {stale, bad, uncertain}, downgrade-only -- matching what PR 1's overlay_from_factorylm_snapshot already did for the direct path. The band is still computed from stored values; no now() at read time, and the collector-liveness meaning of freshness_status is unchanged. The fix belongs in the reader rather than the writer because liveness and value quality are two different facts and both belong in the band the technician sees. Regression cover, both mutation-verified (reverting the fix turns each red): - test_stale_quality_never_becomes_live_ON_THE_DEPLOYED_PATH crosses ingest -> cache -> read-back with the degraded tag, asserting {live: 6, stale: 1} with all 7 tags still present (downgraded, never dropped). Its sibling test_stale_quality_never_becomes_live runs on PR 1's DIRECT adapter, which is why the defect survived. - test_both_overlay_paths_agree_on_freshness pins the direct and deployed overlays to the same answer, so one snapshot can no longer yield two verdicts depending on plumbing the technician cannot see. Suites: proof 24 -> 26, test_factorylm_live_serving 18, spine/context 242, relay 232. No regressions. Pre-existing and untouched: ruff format drift in tests/integration/test_machine_evidence_proof.py (lines 95/178/192/202 on main, none in the added tests). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0191hTx4dwKR45av8njFyEPz

…n as live (#3067) Found by crossing the seam between PR 3 (#3059) and PR 4 (#3061) of PRD #3048 with a degraded tag -- something neither side's suite did, because each was self-consistent on its own. The shared fixture marks conv_simple.height_sensor_mm as quality="stale". Through the DEPLOYED path (ingest -> live_signal_cache -> read-back) the overlay returned it as Freshness.LIVE with freshness_summary {live: 7} -- zero stale. A technician was shown, as current, a reading the producer had already flagged. Cause -- two correct decisions that nobody joined. tag_ingest.persist_batch stamps freshness_status = 'simulated' if simulated else 'live', which is right: that column means COLLECTOR liveness, and a tag whose value is stale is still arriving on time. factorylm_live._freshness_for then read ONLY that column, never latest_quality -- which correctly held "stale" on the very same row. Neither module was wrong in isolation, which is exactly why both suites stayed green. Fix: _freshness_for now takes latest_quality and downgrades on {stale, bad, uncertain}, downgrade-only -- matching what PR 1's overlay_from_factorylm_snapshot already did for the direct path. The band is still computed from stored values; no now() at read time, and the collector-liveness meaning of freshness_status is unchanged. The fix belongs in the reader rather than the writer because liveness and value quality are two different facts and both belong in the band the technician sees. Regression cover, both mutation-verified (reverting the fix turns each red): - test_stale_quality_never_becomes_live_ON_THE_DEPLOYED_PATH crosses ingest -> cache -> read-back with the degraded tag, asserting {live: 6, stale: 1} with all 7 tags still present (downgraded, never dropped). Its sibling test_stale_quality_never_becomes_live runs on PR 1's DIRECT adapter, which is why the defect survived. - test_both_overlay_paths_agree_on_freshness pins the direct and deployed overlays to the same answer, so one snapshot can no longer yield two verdicts depending on plumbing the technician cannot see. Suites: proof 24 -> 26, test_factorylm_live_serving 18, spine/context 242, relay 232. No regressions. Pre-existing and untouched: ruff format drift in tests/integration/test_machine_evidence_proof.py (lines 95/178/192/202 on main, none in the added tests). Claude-Session: https://claude.ai/code/session_0191hTx4dwKR45av8njFyEPz Co-authored-by: MIRA Beta Orchestrator <beta-orchestrator@factorylm.local> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

…NS resolution + QR certification Live-probe finding (PRD #3048 PR 5): the vendor/model resolver returns uns_path=None for equipment names like CV-101, and nothing else populated uns_context.uns_path — so the FactoryLM live overlay's gate failed on EVERY real chat turn AND every QR deep-link turn. PR 4's engine tests mocked resolve_uns_path, which hid this. - factorylm_live.uns_prefix_for_asset(tenant, *candidates): tenant-scoped cmms_equipment lookup (exact equipment_number first, then unambiguous description match only) — the same physical-subtree identity source the relay allowlist seed and the QR deep-link use. Fail-open, read-only. - engine._build_factorylm_live_overlay: falls back to that lookup when the turn carries no uns_path (tag first, display label second). - telegram/start_command: a QR scan now seeds the certified uns_context (source=direct_connection) with cmms_equipment.uns_path, per .claude/rules/direct-connection-uns-certified.md — previously it seeded only asset_identified/asset_tag, so even QR turns failed the gate. - tools/factorylm_live_soak.py: N-iteration deterministic soak of the real engine seam + context fold (Celery-registrable callable), target 100%. - tools/seeds/staging-cv101-probe.sql: idempotent staging CV-101 row so the probe subtree resolves (prod already has the garage bridge seed). 30 tests pass (test_factorylm_live_serving 24, test_start_command 6). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dgej4vU7mMnYTmU6QXsH3n

…NS resolution + QR certification (#3068) * fix(spine): live overlay reachable from real turns — equipment-name UNS resolution + QR certification Live-probe finding (PRD #3048 PR 5): the vendor/model resolver returns uns_path=None for equipment names like CV-101, and nothing else populated uns_context.uns_path — so the FactoryLM live overlay's gate failed on EVERY real chat turn AND every QR deep-link turn. PR 4's engine tests mocked resolve_uns_path, which hid this. - factorylm_live.uns_prefix_for_asset(tenant, *candidates): tenant-scoped cmms_equipment lookup (exact equipment_number first, then unambiguous description match only) — the same physical-subtree identity source the relay allowlist seed and the QR deep-link use. Fail-open, read-only. - engine._build_factorylm_live_overlay: falls back to that lookup when the turn carries no uns_path (tag first, display label second). - telegram/start_command: a QR scan now seeds the certified uns_context (source=direct_connection) with cmms_equipment.uns_path, per .claude/rules/direct-connection-uns-certified.md — previously it seeded only asset_identified/asset_tag, so even QR turns failed the gate. - tools/factorylm_live_soak.py: N-iteration deterministic soak of the real engine seam + context fold (Celery-registrable callable), target 100%. - tools/seeds/staging-cv101-probe.sql: idempotent staging CV-101 row so the probe subtree resolves (prod already has the garage bridge seed). 30 tests pass (test_factorylm_live_serving 24, test_start_command 6). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dgej4vU7mMnYTmU6QXsH3n * fix(seeds): staging-cv101-probe — unquoted psql default + quoted-tenant repair Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dgej4vU7mMnYTmU6QXsH3n * fix(seeds): staging-cv101-probe keys on the global equipment_number constraint Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dgej4vU7mMnYTmU6QXsH3n * fix(seeds): tenant_id is pre-quoted by apply-seeds — reference it bare Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dgej4vU7mMnYTmU6QXsH3n * fix(soak): live_prompt_block takes the overlay, not the combined context Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dgej4vU7mMnYTmU6QXsH3n * chore(release): v3.242.3 + changelog for the live-overlay resolution fix Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dgej4vU7mMnYTmU6QXsH3n * style(engine): ruff-format after rebase onto main Format-only. The rebase onto current main surfaced drift that the later chain commit used to carry; no behavior change. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dgej4vU7mMnYTmU6QXsH3n --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

…an no longer fabricate plant state Probe finding 2 (PRD #3048 PR 5): 'current state of my garage conveyor' -> general_question (1.00) -> bare 80-token LLM call -> fabricated fault + error log for a healthy machine, KB-gap footnote appended after. - Stage-0 deterministic asset-state override -> diagnose_equipment (tag-query fast-path precedent); UNS-gate precondition ORs the signal (resolver confidence is vendor-only, 0.0 for 'garage conveyor'). - Branch 3b: pre-generation refusal for plant-state questions on the general path - no LLM, honest no-live-data admission + handoff. - Branch-5 prompt states no-live-connection; _INDUSTRIAL_HINTS_RE gains plant nouns; general exits carry citation_evidence for telemetry. - Router prompt: asset-state = diagnose_equipment, CRITICAL RULE 7. 75 tests pass across touched suites. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dgej4vU7mMnYTmU6QXsH3n

…th can no longer fabricate plant state (#3069) * fix(engine): asset-state questions reach the UNS gate; general path can no longer fabricate plant state Probe finding 2 (PRD #3048 PR 5): 'current state of my garage conveyor' -> general_question (1.00) -> bare 80-token LLM call -> fabricated fault + error log for a healthy machine, KB-gap footnote appended after. - Stage-0 deterministic asset-state override -> diagnose_equipment (tag-query fast-path precedent); UNS-gate precondition ORs the signal (resolver confidence is vendor-only, 0.0 for 'garage conveyor'). - Branch 3b: pre-generation refusal for plant-state questions on the general path - no LLM, honest no-live-data admission + handoff. - Branch-5 prompt states no-live-connection; _INDUSTRIAL_HINTS_RE gains plant nouns; general exits carry citation_evidence for telemetry. - Router prompt: asset-state = diagnose_equipment, CRITICAL RULE 7. 75 tests pass across touched suites. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dgej4vU7mMnYTmU6QXsH3n * feat(engine): probabilistic asset-state arbitration replaces the binary override Per Mike: routing is now a scored decision, not a single regex gate. asset_state_probability() combines weak signals as log-odds (state phrasing +2.4, asset noun +1.6, tag token +1.2, educational -2.6, router vote +/-1.2/0.6 x confidence, prior -2.0) through a sigmoid; score >= MIRA_ASSET_STATE_THRESHOLD (0.5) forces the gated path. Deterministic math — same inputs, same score, never sampled — so the behavior stays unit-testable; per-signal contributions are logged for evidence-based threshold tuning. Probe corpus clears the threshold against a maximally-confident disagreeing router; educational questions score <=0.03; a bare asset mention (~0.27) never forces the gate. 76 tests pass. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dgej4vU7mMnYTmU6QXsH3n * style(engine): ruff-format the arbitration + overlay hunks Fixes the Lint & Format failure blocking CI Gate on #3069 — format-only, no behavior change. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dgej4vU7mMnYTmU6QXsH3n --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

docs: add MIRA FactoryLM evidence handoff PRD

8d36fdc

Mikecranesync temporarily deployed to staging August 2, 2026 03:48 — with GitHub Actions Inactive

Mikecranesync temporarily deployed to staging August 2, 2026 04:56 — with GitHub Actions Inactive

Merge remote-tracking branch 'origin/main' into prd-amend-tmp

fb76c4e

Mikecranesync temporarily deployed to staging August 2, 2026 04:59 — with GitHub Actions Inactive

Merge branch 'main' into codex/mira-factorylm-machine-evidence-prd

7d62fb2

Mikecranesync temporarily deployed to staging August 2, 2026 05:02 — with GitHub Actions Inactive

Mikecranesync merged commit 5691f0d into main Aug 2, 2026
23 checks passed

Mikecranesync mentioned this pull request Aug 2, 2026

fix(contracts): machine-snapshot fixtures carry all seven conv_simple.* canonical tags (factorylm #199 sync) #3058

Merged

Mikecranesync mentioned this pull request Aug 2, 2026

fix(contracts): the machine-snapshot boundary is enforced, not just agreed #3060

Merged

Mikecranesync mentioned this pull request Aug 2, 2026

feat(spine): FactoryLM live machine-state serving path (PRD #3048, PR 4) #3061

Merged

Mikecranesync mentioned this pull request Aug 2, 2026

test(spine): PRD #3048 PR 5 — controlled integration proof harness + runbook #3062

Merged

This was referenced Aug 2, 2026

fix(relay): FactoryLM publisher must validate the ingest result, not just HTTP 2xx #3063

Closed

fix(release): tags are one minor ahead of /VERSION — rollback checkpoints restore the wrong commit #3064

Closed

Mikecranesync added a commit that referenced this pull request Aug 2, 2026

Merge pull request #3062 from Mikecranesync/feat/pr5-integration-proof

0cc22b4

test(spine): PRD #3048 PR 5 — controlled integration proof harness + runbook

Mikecranesync mentioned this pull request Aug 2, 2026

fix(spine): a reading the producer marked stale reached the technician as live #3067

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

docs(prd): MIRA × FactoryLM Phase 1 — read-only machine evidence handoff - #3048

docs(prd): MIRA × FactoryLM Phase 1 — read-only machine evidence handoff#3048
Mikecranesync merged 4 commits into
mainfrom
codex/mira-factorylm-machine-evidence-prd

Mikecranesync commented Aug 2, 2026

Uh oh!

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Mikecranesync commented Aug 2, 2026

Premise verification (I checked every external claim before opening this)

Preflight step 5/6 is now answered: an authorized ingress exists

Four gaps found against the real ingest contract

One design note

Scope

Uh oh!

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Mikecranesync commented Aug 2, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant