The last leg of the #80/#95 scene-context leak. llm_correlator._call_block
read call_doc's whole-call transcript for every scene, so on a multi-scene
call every scene's cheap-tier and tiebreaker decision was made against text
that also contained the other scenes.
- intelligence.py: each processed[] scene now carries its own "transcript" —
transcript_corrected, else this scene's segments joined, else (single scene)
the whole transcript.
- _build_context / preview_correlation / correlate_call: take a `transcript`
param; _build_context resolves ctx["scene_transcript"] from it, falling
back to the call doc (sweep, single-scene, tests) — the fallback is kept
here, unlike embedding/severity, because a scene always has real text.
- upload.py: both scene loops pass scene["transcript"].
- llm_correlator._call_block: reads ctx["scene_transcript"] (call-doc
fallback retained for test-built ctx).
- recorrelation_sweep: passes the call doc's text explicitly.
- +1 regression test. Full c2-core suite green (296 passed, sandboxed venv).
NOT for merge until the running correlation measurement window closes and its
dump is analysed — deploying a correlator change mid-window would mix old and
new behaviour in the sample.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
intelligence.py writes only the primary scene's embedding and severity to
calls/{id}. _build_context read them back off the call doc, so every
non-primary scene of a multi-scene call was correlated against scene 1's
semantic vector and severity rung: a scene about a different event scored
on the embedding path against the wrong incident, and could inherit a
minor/moderate/major severity it never had, clearing the creation gate on
borrowed weight. Same defect and same fix as the #87 coords leak.
- _build_context / preview_correlation / correlate_call: take embedding and
severity as params; drop the call_doc.get() fallbacks. A scene that
passes none has none, and is judged thin on its own signal.
- upload.py: both scene loops pass scene["embedding"] / scene["severity"];
_correlate_with_consensus forwards them. The no-scene unclassified branch
passes neither (correct: no scene, judged thin).
- recorrelation_sweep: passes the call doc's stored values explicitly
(whole-call re-link, link-only, so a borrowed severity cannot create).
- intelligence.py: SCENE DETECTION prompt tightened toward one scene
(server-26#5, partial) - MULTIPLE only for genuinely separate events,
"when unsure, one scene", plus a not-a-new-scene list.
- test_incident_identity.py: +2 regression tests mirroring the #87 test.
Full c2-core suite green (295 passed). #5 prompt change is unmeasured -
needs a scoped correlation-only window. Known remaining legs, tracked
separately: llm_correlator._call_block still reads the whole-call
transcript per scene; content-divergence veto skips on a None embedding.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The 2026-08-16 correlation dump showed two failures that looked unrelated and
were the same bug. TG 9048 held one incident of 28 calls spanning 49 minutes --
a prisoner transport, a drone retrieval, a records lookup and a canvass, glued
together -- while 32 other calls on that same channel stayed permanently
orphaned.
Creating an incident required a concrete incident_type. Nothing on a transit
police channel produced one: the extraction prompt said to prefer "other" when
uncertain, extraction then collapsed "other" to None, and the tag-based fallback
had no tags to work with because administrative traffic carries none. So the
channel could never open a SECOND incident. Every later call funnelled into
whichever incident happened to exist first, and every call too substantial for
the thin path had nowhere to go at all. The two symptoms were the same missing
value seen from opposite ends.
Severity now decides incident-worthiness. It is a better fit for the question
being asked -- "is this a real event?" -- than a service label ever was, and
unlike incident_type it is always present. The prompt defines four levels with
no escape hatch (routine/minor/moderate/major, "unknown" is gone) and calls
skipped for a too-short transcript are still recorded as routine, because
downstream code reads a missing severity as "not processed yet" rather than
"nothing happened". Anything above routine, or carrying any extracted content,
opens an incident under the neutral "other" type. "other" is also kept as a real
classification now -- rail operations and public works genuinely are not police,
fire or EMS.
Separately, thin calls no longer refresh updated_at; they write last_thin_at.
updated_at drives every recency gate in the fast path, so each "10-4" was
resetting the idle clock on whatever it attached to, keeping that incident
inside the gate for as long as anyone kept acknowledging. An incident now ages
from its last substantive call. This is what made the 49-minute incident
possible even once buckets existed, so it is fixed independently rather than
being left to the gate change.
The re-correlation sweep also now honours skip_reason. /upload has always
refused to correlate garbage and too-short transcripts, but the sweep did not
apply the same filter, so those fragments came back minutes later through the
thin path and attached to whatever was most recent -- a second, quieter route
into the same over-merge.
Adds tests/test_correlator_gate.py (15 cases), the first tests against
incident_correlator.py in its 1,517-line history. tests/conftest.py stubs
firebase-admin only when it is genuinely absent, so the container's real SDK is
never shadowed; this is what makes the correlator importable in the dev venv.
That stub also made test_mqtt_handler and test_node_sweeper collectable for the
first time, revealing 10 pre-existing failures in them -- test-vs-code drift,
untouched here and catalogued in DEFERRED.md.
No new environment variables, so CI deploys this without an ansible run.
Two independent sources of garbage in the AI pipeline, both visible in the
2026-08-16 correlation dump.
1. Hallucinated transcripts. The Whisper prompt opened with an enumerated run
of ten-codes: 10-4, 10-23, 10-20, 10-97 and so on. Whisper treats prompt
text as preceding transcript, so on noisy or silent audio it continued the
series, emitting transcripts that count upward from 10-4 to 10-99. The
existing no_speech_prob filter could not catch these: the model is highly
confident in text it invented by continuing a pattern.
The prompt no longer contains a series to extend, and _is_degenerate()
rejects the three shapes this failure takes: ascending ten-code runs, one
phrase looping, and near-identical segments across a whole recording.
Verified against 13 transcripts from production: all four known
hallucinations rejected, all nine real ones kept, including terse traffic
containing legitimate codes.
2. Duplicate recordings. node-002 and node-PI-2 both cover TG 9048 and both
uploaded the same transmissions, ~1.1s apart. Nine pairs appeared in one
dump. Each was transcribed, billed and correlated twice, and the resulting
incident listed two units where there was one.
Canonical selection is by earliest started_at, tie-broken on call_id, NOT
by upload order: upload order varies with encode time and network latency,
so it would make the authoritative recording non-deterministic. Call
documents are created from MQTT call_start before uploads arrive, so both
nodes independently reach the same verdict. The loser keeps its audio (it
may be the cleaner capture) but is excluded from STT, correlation, the
re-correlation sweep and the orphan debug view.
Also fixes _sync_transcribe returning a bare None when OPENAI_API_KEY is
missing, where the caller unpacks two values. A missing key surfaced as a
misleading "Transcription failed" instead of the real warning.
Adds tests/test_dedup.py (15 cases). dedup.py reaches Firestore through an
injected callable so it stays importable without firebase-admin present.
Correlator
- Raise fast-path idle gate 30 → 90 min (tg_fast_path_idle_minutes)
- Fix disambiguate always-commits bug: run _call_fits_incident on winner
before committing; fall through to new-incident creation if it fails
- Add unit-continuity path (path 1.5): matches all_active by shared unit
IDs with a reassignment guard, bridges calls past the idle gate
- Add tag-based incident_type inference (_TAG_TYPE_HINTS) as GPT fallback,
rescuing tagged calls that would have been dropped (616 observed orphans)
- Add master/child incident model: _create_master_incident, _demote_to_child,
_add_child_to_master; new incidents stamped incident_type="master"
- Add cross-system parent detection (_find_cross_system_parent): two-signal
scoring (road overlap=0.4, embedding≥0.78=0.3, proximity=0.3, threshold=0.5)
wired into create-if-new path; creates master shell on first cross-system match
- Add maybe_resolve_parent: auto-resolves master when all children close;
called from upload pipeline (LLM closure) and summarizer stale sweep
- Add signal-based auto-resolve via units_active/units_cleared tracking:
GPT now extracts cleared_units per scene; _update_incident moves units
between active/cleared lists and resolves the incident when active empties;
stored on call doc for re-correlation sweep reuse
- Add _create_incident initialization of units_active/units_cleared fields
Re-correlation sweep
- Add corr_sweep_count + MAX_SWEEP_ATTEMPTS=3: orphans get 3 attempts
then are tombstoned as corr_path="unlinked", ending the re-sweep loop
(previously hammering each orphan 29-31 times per shift)
Intelligence extraction
- Add cleared_units to GPT prompt schema and rules
- Extract and propagate cleared_units per scene; merge across scenes;
store on call doc for re-correlation sweep
Token management
- Fix token release bug: remove release_token call on discord_connected=False
in MQTT checkin (transient Discord drops were orphaning bots mid-shift)
- Add PUT /tokens/{id}/prefer/{system_id} endpoint: lock a bot token to a
system; pass _none as system_id to clear; stored bidirectionally on both
token and system documents
- discord_join handler resolves preferred_token_id from system doc and passes
system_name in MQTT payload
Logan CusanoandClaude Sonnet 5