AI providers: review follow-ups to 0060ccd

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>

drb-c2-core/app/config.py

@@ -122,17 +122,17 @@ class Settings(BaseSettings):
     # NWS alerts cache TTL; the issue asks 60-120 s (floored at 30 s in code).
     weather_alerts_ttl_s: int = 90
 
-    # Gemini (intelligence extraction, embeddings, incident summaries)
+    # Gemini: only used if CORR_CHEAP_MODEL / TRANSCRIPT_CORRECTION_MODEL is set
+    # to a gemini-* ID (app/internal/llm.py). OpenAI serves both by default.
     gemini_api_key: Optional[str] = None
     # Correlation LLM (server-26#299). The rules engine decides every confident
     # case; this model is called at most ONCE per scene, and only when the rules
     # flag the decision as ambiguous (incident_correlator._ambiguity). There is
     # no second "smart" tier any more: the cheap/smart consensus tiebreak ran on
     # ~15% of scenes and resolved nearly every disagreement as "new" (#115).
-    # The model ID was once retired by Google and 404'd on every call for days;
-    # a failed LLM call falls back to the rules decision, so nothing broke
-    # loudly. Verify against https://ai.google.dev/gemini-api/docs/models
-    # before changing.
+    # A model ID was once retired by its provider and 404'd on every call for
+    # days; a failed LLM call falls back to the rules decision, so nothing broke
+    # loudly. Verify the ID against the provider's model list before changing.
     corr_cheap_model: str = "gpt-4o-mini"   # OpenAI since 2026-11-08 (was gemini-3.6-flash); a gemini-* ID still works
     # Transcript correction (server-26#36). Runs inside transcription, once per
     # transcribed call above MIN_WORDS_FOR_CORRECTION, so it is priced like STT

drb-c2-core/app/internal/llm.py

@@ -79,11 +79,15 @@ def generate_json(model: str, prompt: str, *, purpose: str,
     from app.internal.ai_clients import openai_client
     # JSON mode requires the word "JSON" in the prompt; both prompts say
     # "Return JSON" / "Respond with JSON only".
+    kwargs = {}
+    # Reasoning models (o-series, gpt-5) reject a temperature; chat models get 0.
+    if not model.startswith(("o1", "o3", "o4", "gpt-5")):
+        kwargs["temperature"] = 0
     resp = openai_client().chat.completions.create(
         model=model,
         messages=[{"role": "user", "content": prompt}],
         response_format={"type": "json_object"},
-        temperature=0,
+        **kwargs,
     )
     usage = getattr(resp, "usage", None)
     record_usage(purpose, model,

drb-c2-core/app/internal/llm_correlator.py

@@ -16,7 +16,7 @@ It returns a decision dict compatible with incident_correlator._apply_decision:
    "reasoning": str}
 
 or None — no API key, nothing to choose from, an answer that fails
-validation, or any Gemini failure — and the caller then commits the rules
+validation, or any LLM failure — and the caller then commits the rules
 decision unchanged, so the pipeline never stalls on the LLM.
 """
 import asyncio
@@ -196,7 +196,7 @@ def _sync_gemini(model_name: str, prompt: str) -> dict:
 
 def _parse_response(raw: dict, ctx: dict, candidates: list[dict]) -> Optional[dict]:
     """
-    Convert raw Gemini JSON to a decision dict, or None when the answer is
+    Convert the model's raw JSON to a decision dict, or None when the answer is
     unusable and the rules decision should stand.
 
     A "link" must name one of the candidates the model was shown — not any

drb-c2-core/app/internal/recorrelation_sweep.py

@@ -84,7 +84,7 @@ def _derive_min_minutes() -> int:
     return max(_FLOOR_MINUTES, math.ceil(pipeline_worst_case_seconds() / 60))
 
 
-# 21 with the defaults (stt 120s, openai 60s x2 attempts, gemini 60s, 3 scenes).
+# 22 with the defaults (stt 120s, openai 60s x2 attempts for every call, 3 scenes).
 MIN_MINUTES_SINCE_PIPELINE_START = _derive_min_minutes()
 
 

drb-c2-core/app/internal/replay.py

@@ -58,9 +58,9 @@ PREFETCH = 6
 # Estimates, not a bill — nothing in DRB reads a real invoice (server-26#45).
 USD_WHISPER_PER_MIN    = 0.006
 USD_PER_EXTRACTION     = 0.0005   # gpt-4o-mini scene extraction + embedding
-USD_PER_CORRECTION     = 0.0003   # Gemini flash transcript correction
+USD_PER_CORRECTION     = 0.0004   # gpt-4o-mini transcript correction (~1.5k in / 300 out tokens)
 USD_PER_GEOCODE        = 0.005    # Google geocode, roughly one per located scene
-USD_PER_LLM_CORRELATE  = 0.0005   # one Gemini flash correlation decision (ambiguous scenes only)
+USD_PER_LLM_CORRELATE  = 0.0004   # one gpt-4o-mini correlation decision (~2k in / 100 out; ambiguous scenes only)
 _LLM_CONSENSUS = ("llm", "rules_guard", "agreed", "tiebreak", "llm_orphan_gate")
 
 # Fields the pipeline writes onto a call doc. Stripped when a call is copied
@@ -569,7 +569,7 @@ async def _run(run_id: str, org_id: str, calls: list[dict], mode: str,
         metrics = compute_metrics(incidents, sb_calls)
         metrics["est_cost_usd"] = _running_cost(progress, metrics, mode)
         metrics["ai_failures"] = dict(Counter(f"{f['tier']}: {f['problem']}" for f in ai_failures))
-        metrics["gemini_usage"] = usage
+        metrics["llm_usage"] = usage
     except Exception as e:
         status = "failed"
         errors.append(f"run: {type(e).__name__}: {e}"[:300])