v1.10.6

1.10.6 (2026-07-16)

Bug Fixes

• agents: force what the SC regulation makes certain, not a strategy opinion (281fb32), closes #464
• cli: feed the PMV the real overtake gap, and give the augmentation one home (dc9914d)
• N26: scope the TCN window to this stint, up to now, and zero-pad as trained (20a7a06), closes #449
• N28: refuse to score an undercut against a car that is not racing (018f910), closes #462
• orchestrator: stop the LLM overriding the validated undercut target (fc2c42f), closes #462

Documentation

• multi-agent: document the SC rails as rules, not as a forced stop (d5f8709), closes #464

v1.10.5

1.10.5 (2026-07-16)

Bug Fixes

• agents: compute FuelEffect from the stint baseline, not a hardcoded fuel_load factor (5d4ff38)
• agents: compute FuelEffect from the stint baseline, not a hardcoded fuel_load factor (7ac17e5), closes #446
• agents: encode N15's compound_id as the ordinal rank it was trained on (6bb025f), closes #445
• agents: honour the Safety-Car guard-rail on the final answer, and stop simulating free pit stops (b423acd), closes #436
• agents: let the overtake model see the elapsed-time gap it was trained on (46c320f)
• agents: let the overtake model see the elapsed-time gap it was trained on (3b722bb)
• agents: re-key the circuit lookup tables to the slug keyspace the agents query with (02e7a79), closes #448
• agents: Safety-Car rail on the final answer, MC sentinels, and un-freeze the circuit tables (0fd28e7)
• agents: un-invert the undercut model's top feature and stop feeding it the race lap (a68025c)
• agents: un-invert the undercut model's top feature and stop feeding it the race lap (ec0f444), closes #444
• engine: scope the laps frame to the analysed Grand Prix in run_lap (9a8f0cd)
• engine: scope the laps frame to the analysed Grand Prix in run_lap (454db40), closes #429
• simulation: guard the stint-baseline precompute against frames without Stint (4262859), closes #446

Documentation

• agents: fix the entry-point table and make the examples actually run (59b02e0), closes #438
• strategy: correct the pages the engine fixes made wrong (4168b13)
• strategy: correct the pages the engine fixes made wrong (62fa3c5), closes #438

v1.10.4

1.10.4 (2026-07-15)

Bug Fixes

• deps: bump click/kafka-python/transformers past CVEs; waive blocked pillow and setuptools (97d273d)
• deps: resolve OSV Python CVE backlog (bump click/kafka-python/transformers; waive blocked pillow and setuptools) (fbd6df2)

v1.10.3

1.10.3 (2026-07-12)

Bug Fixes

• nlp: DOUBLE YELLOW severity + time-penalty routing (#398 follow-up) (7f3f7a7)
• nlp: RCM parser superset + penalty/red-flag routing + arcade SC persistence (#398) (47cf9a7)
• nlp: render DOUBLE YELLOW at yellow severity + route time penalties to N30 (#398 follow-up) (b45d7f1)
• nlp: widen the RCM parser + route penalty/red-flag alerts + persist SC in arcade (#398) (6a223e1)

Performance

• cli: skip double model load in --no-llm prewarm (#389) (56c24f7)
• cli: skip the tire/situation/pit singleton prewarm in --no-llm mode (#389) (f0ba4b4)

v1.10.2

1.10.2 (2026-07-12)

Bug Fixes

• data: resolve every GP's radio + compound labels via a canonical name (F-01) (2ab70d1)
• data: resolve every GP's radio + compound labels via a canonical name (F-01) (29d1cac), closes #243
• data: validate the parquet read boundary + surface FastF1 quality flags (F-02) (80b12b6)
• data: validate the parquet read boundary and surface FastF1 quality flags (F-02) (dd44451), closes #244
• deps: drop fitz dummy, declare pypdf, override frozendict (DX-05) (3a7612c)
• deps: drop the unused fitz dummy, declare pypdf, override frozendict (DX-05) (794a39b), closes #253
• devex: unbreak the Docker/Streamlit quickstart + redirected-output crash (#388) (75d1b3e)
• devex: unbreak the Docker/Streamlit quickstart and the redirected-output crash (f7f988a), closes #252 #388
• nlp: persist Safety-Car state across laps so the override survives the stint (NR-02) (c713705)
• nlp: persist Safety-Car state across laps so the override survives the stint (NR-02) (215007b), closes #305
• packaging: bundle the telemetry submodule in the release wheel (PK-01) (a047b48)
• packaging: bundle the telemetry submodule in the release wheel (PK-01) (40ae581), closes #289

Documentation

• security: add the Phase A security-boundary design for review (#224) (c71eac9)
• security: Phase A security-boundary design for review (#224) (ed4e829)

v1.10.1

1.10.1 (2026-07-12)

Bug Fixes

• cli: wire the simulation CLI to the shared inference engine (#236) (32dcc4b)
• cli: wire the simulation CLI to the shared inference engine (#236) (0694299)

v1.10.0

1.10.0 (2026-07-11)

Features

• agents: give every sub-agent ChatOpenAI a finite provider timeout (L-1) (3eb0eb8)
• agents: give every sub-agent ChatOpenAI a finite provider timeout (L-1) (401765f), closes #263

Bug Fixes

• ci: skip the tiktoken vocab roundtrip on an upstream network outage (6c26ec2)
• ci: skip the tiktoken vocab roundtrip on an upstream network outage (cdd9f4a)

v1.9.0

1.9.0 (2026-07-11)

Features

• eval: reproduce the pace MAE 0.4104 from the featured-laps holdout (2a2c982)
• eval: reproduce the pace MAE 0.4104 from the featured-laps holdout (998a800)
• eval: reproduce the tire MAE 0.7078 and validate the MC-Dropout sigma (8ce8653)
• eval: reproduce the tire MAE 0.7078 and validate the MC-Dropout sigma (762b030)

Bug Fixes

• eval: correct the circuit_cluster hygiene verdict to a real coarse leak (4328f95)
• eval: correct the circuit_cluster hygiene verdict to a real coarse leak (679208e)

Documentation

• eval: resolve the circuit_cluster hygiene item as an accepted non-target limitation (33ff301)
• eval: resolve the circuit_cluster hygiene item as an accepted non-target limitation (c72c9fe)

v1.8.0

1.8.0 (2026-07-11)

Features

• eval: add LLM-judge alert-precision module over unlabeled radios (b727baa)
• eval: add metrics-registry + calibration harness (f1-eval) (c08780d), closes #206
• eval: add NLP per-stage eval harness (f1-eval nlp) (dbf76b1), closes #304
• eval: add threshold-provenance + leakage hygiene report (f1-eval hygiene) (3788167), closes #207
• eval: compute radio alert precision from gold intent labels (16bd305), closes #304
• eval: LLM-judge alert precision over unlabeled radios (#304 follow-up) (1651c3b)
• eval: ML metrics-registry + calibration harness (f1-eval, #206) (380f1d7)
• eval: NLP per-stage eval harness (f1-eval nlp, #304) (1c8ff7c)
• eval: radio alert precision from gold labels (closes #304) (da1db0c)
• eval: recompute SC + undercut calibration + AUC-PR (part of #364) (5c97461)
• eval: recompute SC + undercut calibration and AUC-PR from in-memory holdouts (7365b50), closes #364
• eval: regenerate pit holdout from raw laps (closes #364) (d66ddae)
• eval: regenerate pit holdout from raw laps + recompute coverage and MAE (ccc213d), closes #364
• eval: reproduce intent setfit-free + #303 NLP hygiene verdicts (6bc5413)
• eval: reproduce intent setfit-free and record #303 NLP hygiene verdicts (275c487)
• eval: reproduce NER entity-F1 + RCM coverage (#304) (bc0350d)
• eval: reproduce NER entity-F1 and RCM coverage in the NLP harness (1efb133), closes #304
• eval: threshold-provenance + leakage hygiene report (#207) (5f1f29d)
• eval: wire f1-eval alert-llm + report the unlabeled-corpus finding (1e9fed1), closes #304

Bug Fixes

• eval: align undercut provenance wording with in-train relabeling (Fable re-verify) (54e5d42)
• eval: correct SC operating threshold + window on honest splits (#363) (1275101)
• eval: correct SC threshold + window on honest splits (closes #363) (e1ee603)
• eval: relabel SC/overtake threshold corrections as in-train (Fable gate) (c38c94e)

Documentation

• eval: regenerate hygiene report with in-train relabeling + correct provenance stamp (c1af9c1)
• roadmap: reconcile published metrics to thesis/IEEE finals (2758ec5)
• roadmap: reconcile published metrics to thesis/IEEE finals (#213) (50b7ecf)

v1.7.0

1.7.0 (2026-07-11)

Features

• strategy: add no-llm engine profile (fixes the --no-llm crash) (f32e38b)
• strategy: add shared inference engine (rich profile) + arcade delegate (43c2717)
• strategy: no-llm engine profile (fixes the --no-llm crash, #166) (91b98d3)
• strategy: shared inference engine (rich profile) + arcade delegate (0e1b793)