Radio

Pit radio, classified and verified against lap-scoped telemetry -- instruction compliance and claim corroboration, side by side with the lap-time line they are pinned to.

Fixture replay -- not a live session
SourceUploading your own clip goes to POST /api/radio/transcribe on the gateway; this page replays committed fixtures so it works with no backend attached.
Coverage note: golden-set agreement is 34/38 across the two committed sessions, and only BOX and TRAFFIC have genuine golden-set coverage. Golden-set agreement (docs/RADIO_CALIBRATION.md) is measured with the golden label's own intent and Band.HIGH forced on every transcript (scripts/run_radio_golden_set.py) -- that isolates the verification layer from STT/classification noise. This fixture instead runs the real Tier-1 cascade and each clip's real STT band end to end, so its verdicts are the live pipeline's actual output, not a replay of the 34/38 benchmark. Only BOX and TRAFFIC carry real golden coverage either way -- 10 of the 13 catalogued specs are uncalibrated or placeholder thresholds.

Fixture generated 11/8/2026, 8:42:05 pm by python -m scripts.generate_radio_fixture from data\radio_fixtures\9149, data\radio_fixtures\9157 -- the real Tier-1 grammar cascade and verification engine, run once against each clip's actual STT confidence band, not a replay of the golden benchmark.

Driver

Radio timeline

Lap performance

Lap time · VER

Markers are pinned to the lap the radio message resolved on -- the same X axis as the lap-time line. Click a marker to open its evidence below.

1 marker(s) have no honest lap-time position to plot -- shown here, off the axis, rather than at an invented Y value:

Evidence panel

MAXVER01_1_20230827_152625
STT MODERATEIntent MODERATEVerdict MODERATE
VER · L62

Transcript

“We'll box this lap. No, no, it's too dry here. Wait a second.”

STT confidence 54.5% · 13 words

Intent

Instruction (pit → driver) · BOX · speaker: pit
Tier 1 — deterministic grammar · confidence 95%
Evidence: tier1:BOX:\bbox this lap\b
Slots: {"urgency":"immediate"}

Verification

COMPLIEDInstruction complianceGolden-covered
Window: lap 62–63 (resolved 63)
SignalObservedBaselinevs threshold
pit event present1.00—above 0.50

Requires 1 of 1 signal(s) to agree. A blank observed value means the signal was genuinely not computable for this window, not zero.

S5.1: COMPLIED if a pit event (is_pit_lap) occurs anywhere in the window; NON_COMPLIED if none and the window was fully observed (race continued past it); NOT_VERIFIABLE if the window runs past the session's final lap (RACE_END) before that can be determined. PIT_LAP is intentionally not a guard here -- it is the signal.

spec_version 1.0.1

Cause candidates

Empty by design. Deterministic cause ranking ships in Phase 7 — this is not a missing computation.

Narrative

No SmolLM3 prose yet — Phase 7. The engine decides; the language model only ever narrates a decision that already exists, and every number above already fully evidences this verdict without it.

Phase 0 golden reference

Hand/telemetry-derived label from the calibration dataset — not what the live engine decided above, kept here as a cross-check: INSTRUCTION / BOX → COMPLIED. pit event confirmed lap 63