Replay usage key gemini_usage -> llm_usage (backend + ReplayTab), re-price the
correction/correlation estimates for gpt-4o-mini, no temperature for o-series
or gpt-5 models, and refresh comments still naming Gemini.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
One funded account (OpenAI) now covers extraction, embeddings, summaries,
the ambiguous-scene correlation call and transcript correction; Groq stays
optional for STT. New app/internal/llm.py picks the provider from the model
name, so setting either model back to a gemini-* ID needs no code change.
Usage accounting for replay runs now covers both providers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Resolve upload.py (no-scene fallback uses the new _correlate with #298's
non_event_chatter reason) and chatter_classifier docstring (gate built, off by
default). Fold intelligence._llm_bool into incident_correlator.parse_llm_bool (#88).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
STT_PROVIDER=auto picks Groq whisper-large-v3-turbo only when GROQ_API_KEY is
set (else OpenAI, unchanged); compat covers any OpenAI-compatible server via
STT_BASE_URL/STT_API_KEY. A failed primary falls back to OpenAI STT_MODEL on
the same downloaded audio; the fallback reports to its own ai_health tier so
a dead primary stays red. The non-OpenAI provider gets a token-budgeted
glossary prompt (area places, local_knowledge, vocabulary; never ten-codes),
guarded by a narrow prompt-echo check. A per-30s-window zlib compression
check (Whisper's 2.4) adds a noise gate that survives correction being off
(#83, partial). Call docs record stt_provider/model/audio_seconds/confidence.
The sweep hold-back counts the fallback attempt. gate_chatter added to
chatter_classifier (used by transcription to skip correction when
CHATTER_GATE_ENABLED). GROQ_API_KEY templated with default ''.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The rules engine now commits every confident decision alone. _run_decision
marks a decision ambiguous (new_with_candidates, unfit_tg_candidate,
unverified_unit_link) and only then does upload._correlate make ONE
llm_correlator.decide call; a usable answer is committed, otherwise the rules
decision stands. Removed: _correlate_with_consensus, llm_correlator.tiebreak /
decisions_agree / tiebreak prompt, corr_smart_model, the correlation_smart
health tier, the LLM-orphan gate (_call_is_substanceless, effectively dead:
it could only fire on minor-severity untyped calls) and its sweep budget.
corr_consensus is now rules_only | llm | rules_fallback | rules_guard, with
corr_band naming why; admin debug and replay metrics read the new values.
Section fixes in the same pass:
- #115: no tiebreaker left to manufacture incidents; an LLM orphan of a
rules "new" is honoured only for untyped, substanceless, routine/minor calls.
- #86: divergence veto bar moved to config (embedding_divergence_threshold);
the band between it and the match bar is ambiguous, not a silent merge.
- #103: a unit link with no embedding is "unit_overlap_unverified" ->
ambiguous, instead of skipping the veto and linking as verified.
- #189: unit-only and LLM links refused once an incident is open 90 min;
unit links on incidents open >30 min are ambiguous; LLM links distance-
guarded (incident_max_spread_km) and only to shown candidates.
- #190: severity steps down on an explicit stand-down phrase; new
severity_peak keeps the monotonic max.
- #88: parse_llm_bool fails closed on "false"; scene closures honoured only
on confident links or the incident the scene opened.
- #27: sweep budget is the scan window (wall clock), not 3 attempts.
- LLM prompt: "Reserve new" bullet no longer contradicts the #189 EXCEPT.
- updated_at never moves backward on a back-dated link.
Unmeasured: no live AI or replay (credits unfunded); band frequency, F1 and
thresholds are unverified. Must be re-scored on both answer keys in Replay
before merge.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A day of replay runs (server-26#170) spent ~$5 of Gemini on ~7 two-hour
windows (~$0.70 per 290 calls) — several dollars a day per live deployment
for correlation alone — and nothing could say where it went (#45). Gemini
3.x thinks by default and bills it as output; the deprecated
google-generativeai SDK these calls used cannot set a thinking level.
- app/internal/gemini.py: every Gemini call (correlation + transcript
correction) goes through google-genai with JSON mode, an explicit
thinking level, and logs in/out/thinking tokens. A model that rejects
the level is retried without it once and remembered, so the tier is
never lost to a config param. API failures still raise for ai_health.
- correlator: thinking_level "minimal" (a link/new/orphan choice).
transcript correction: "low" until a replay shows minimal is safe.
- replay: runs record real Gemini token usage (metrics.gemini_usage),
shown in the Replay tab.
- requirements: google-genai.
c2-core: 474 pass. Frontend typecheck not run (no Node on this box).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
First replay (290 calls, 09-22 10:00-12:00 ET) produced 0 incidents and
no errors: every gpt-4o-mini extraction failed and _sync_extract
swallowed it as "no scenes". Same shape as #169 — and the live extraction
tier in /health/ai had no reporter at all, so this has been invisible in
production too.
- intelligence: API failures propagate out of _sync_extract; extract_scenes
reports them to ai_health ("extraction" tier, billing/dead-model
classified) and still returns [] so the pipeline degrades as before.
- ai_health: inside a replay sandbox, failures go to the run's own sink
instead of being dropped.
- replay: aborts after 5 permanent failures on a tier, naming the cause;
run metrics carry ai_failures; UI shows them.
- replay estimate: audio minutes from started_at/ended_at (no duration
field exists on call docs).
- ReplayTab exposes the loaded run on window.__drbReplay for in-page
analysis.
c2-core: 458 pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Correlation has only ever been measured through live AI windows: days of
wall time per change, and the 09-20→22 window was invalidated outright by
unfunded AI accounts (#169). Recordings are kept regardless of AI, so the
traffic to measure against already exists.
- internal/replay.py: runs a time range of real calls through the live
pipeline code in original order, clock pinned per call, into
replay_runs/{run_id}/calls|incidents. Modes: audio (re-transcribe),
transcripts (re-extract), reuse (correlation only from a prior run's
scenes). Simulates the idle-resolve and orphan-recorrelation sweeps on
virtual time. No alerts, summaries, vocab, AI-health alerts or pending
terms. One run at a time, <=5000 calls, <=7 days.
- firestore.py: ContextVar sandbox redirect for calls/incidents.
- clock.py: ContextVar-pinnable now(), used on the correlation path.
- feature_flags.py: ContextVar flag override so replay runs with live AI off.
- upload.py: scene loop extracted to _extract_and_correlate, shared by the
live pipeline and replay so replay measures the code that runs live.
- resolved_via on every incident resolve, so a real clear can be told
from the idle timeout — live and in replay.
- routers/replay.py + /admin Replay tab: estimate, start, compare runs,
drill into incidents with audio.
Reviewed by drb-correlation-review; its leak and fidelity findings are
fixed and covered by tests. c2-core: 456 pass. Frontend typecheck not run
(no Node on the authoring box).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Logan CusanoandClaude Opus 5.5