AirsideLabs/notam-typed-decisions

NOTAM typed decisions

Frozen, hashed evaluation suites for typed decisions about NOTAMs — a choice, a yes/no, or a score, each with a confidence — plus a catalogue of how NOTAMs describe areas in free text. Built by Airside Labs so that any model, served any way, can be scored on the same rows and read with the same per-class table and calibration curve.

Not for operational use. These suites and the numbers quoted here are for research and for triage tooling. NOTAMs are safety-of-life information; nothing here is a substitute for reading them.

What is in it

path	rows	what
suites/deel_notam_subject_test_short.jsonl	4,241	DEEL-AI 13-class subject task, test split, one choice per row; group=disjoint (3,252 rows) is the template-disjoint subset to quote
suites/deel_notam_subject_test.jsonl	4,241	same rows with the 13 definitions inside each row's instruction
suites/deel_notam_subject_test_desc.jsonl	4,241	same rows with the definitions carried on the option keys (for encoder recipes)
suites/notam_multi.jsonl	8,000	1,000 NOTAMs × 8 typed questions; group is the question id
geometry/summary.md, families.csv, by_issuer.csv		the area-description grammar mined from 370,836 NOTAMs; mine_geometry.py reproduces it
baselines.md		every baseline we have run, on the disjoint group, from committed evidence

Each suite ships a .manifest.json with its SHA-256. A suite whose hash does not match its manifest is a different suite; do not compare numbers across them.

Row format

One JSON object per line: case_id, instruction, context (the state), options (list of strings), label (index into options, or null where no truth exists), group, question_type (choice, noul = yes/no with options ["false","true"], or score with ordered levels as options). This is the jevify suite format, so jevify eval <recipe> suites/<suite>.jsonl scores any recipe directly.

The eight questions in notam_multi

question	type	truth source	rows with truth
subject	choice (13 DEEL classes)	ICAO Q-code → class map (structural prior)	760
closure	yes/no	Q-code condition letters	515
aerodrome_traffic	yes/no	Q-line scope contains A	786
enroute_traffic	yes/no	Q-line scope contains E vs plain A	569
daily_schedule	yes/no	item D present	1,000
below_fl100	yes/no	Q-line upper level < 100; null when 999	252
needs_geo_tool	yes/no	weak regex label, a prior not truth	1,000 (weak)
severity	score (none/minor/significant/severe)	none	0

The state is items A/B/C/D/E/F/G rebuilt from parsed fields; the Q-line is never in the state, so the structural labels cannot be read off the header. Score needs_geo_tool only for latency and calibration, never for accuracy; severity has no truth by design.

Provenance and hygiene

• DEEL subject rows: DEEL-AI/NOTAM (MIT), test split, Item E text only. group=disjoint marks rows whose digit-masked template does not appear in DEEL train; 23% of the test split is template-leaked from its own train, so quote the disjoint number.
• notam_multi rows and the geometry catalogue: NOTAMs received from the FAA SWIM feed (public data; international NOTAMs are those redistributed by the FAA), new and replacement NOTAMs only, deduplicated on normalised text, stratified by Q-code subject letter, seed 20260922. Labels are derived from the Q-code and Q-line as stated above and are described as such; they are not human annotations.
• Definitions in the three DEEL suites are DEEL-AI's own class definitions, lightly condensed.
• Anything trained on the source corpus for the baselines below had every row sharing a template with these test rows removed first.

How to read a result

Report the disjoint-group accuracy and the per-class table, never one aggregate. Read off_menu_mass and the scored-vs-cases count before any accuracy: a run whose server died mid-way still prints an accuracy over the rows it reached. Run the shuffled-context and option-permutation controls for any model you quote. A degraded flag on a wide menu can mark confidence rather than failure when label spellings fall outside a top-k window; the off-menu mass is the number that means something.

Baselines

See baselines.md. Headline on the disjoint group: a trained TF-IDF head on the human labels scores 0.936; a ~30B zero-shot model with the definitions in the system prompt 0.898; a 4B model self-distilled on Q-code-derived labels 0.904 at 67 ms per decision on one consumer GPU; headless encoders 0.27 and 0.12. Placement matters: the same zero-shot model scores 0.871 with the definitions in the row and 0.898 with them in the system prompt.

Licence

The suites, labels, manifests, catalogue and card are released under CC BY 4.0 by Airside Labs. The DEEL rows remain under DEEL-AI's MIT licence; NOTAM text is public aeronautical information. Please attribute Airside Labs and DEEL-AI.

Citation

A data paper describing the grammar catalogue and these suites is in preparation. Until then cite this dataset by its Hugging Face URL and commit hash.

Maintainer

Airside Labs — corrections, additional questions and disagreement with any label are welcome as discussions on this repository.