The data gates for Stage 1 role-training sets: one standard-library Python package that decides whether a row, a
batch or a frozen set of training items is fit to train on. The trainer (aspire-si)
and the dataset generator (role-set) both call it, so what a generator
accepts is exactly what the trainer will accept.
Every gate raises or returns a reason you can act on. Nothing is fixed silently.
Install
pip install aspire-si-gates
Python 3.10 or newer. No dependencies.
Use
from aspire_si_gates import DataError, load, validate
rows = load("batch3-t05.jsonl") # one JSON object per line
try:
validate(rows) # every row check, then every whole-set check
except DataError as err:
print(err.code, err.message, err.hint)
A generator that wants to hand refusals back to its writer one item at a time uses the incremental checks instead:
from aspire_si_gates import check_against, check_item
check_item(row) # raises DataError on the first refusal
problems = check_against(row, accepted_rows) # [(code, message), ...] against rows already accepted
The gates
| Call | What it checks |
|---|
validate(rows) | check_item on every row, then check_batch on the whole file. Raises DataError with the first refusal. |
check_item(row) | The checks that need no other row. It refuses an apply item (split: apply or an -AP- id): apply items are never trained. |
check_apply_item(row) | An apply item's checks: the same content and shape rules as check_item, plus the apply rules (split: apply and an -AP- id, kind C or D, one turn, no pair, and the lesson's twin field). Refuses any row that is not an apply item. |
check_trace_leaks_label(row) | Refuses a kind C row whose checker trace names its own planted flaw, or any flaw stem outside the two-step carve-out. Wired into check_batch. |
check_against(row, accepted) | [(code, message)] for the exclusivity rules (spans, framings, material, puzzle shapes, openings, phrases) against rows already accepted. The codes are in REASONS. |
check_batch(rows) | The checks that need the finished set: shares, floors, pairs, phrase and formula limits. |
batch_report(rows), freeze_report(rows) | The figures behind those checks, as dicts. |
gates_identity() | The package version and a sha256 of its source, to record with each generated item. |
The verifier role's data (lessons, trait names, role words and the pre-amendment lessons) is in constants.py. A
different role changes that module, not the checks. The sealed-set leakage leg is not here; it stays with the trainer.
Security and data
- Reads: the rows you pass in, and the file you give to
load.
- Does not touch: the network, the file system beyond that one file, environment variables, or subprocesses. No
telemetry.
- Permissions: none beyond reading that file.
Report a vulnerability privately; see SECURITY.md.
Status
Version 1.7.6, first published to PyPI. The source is byte-identical to the package that lived in aspire-si at
6a603baa, so gates_identity() reports the same hash as before the move. Coverage of this repository's own tests is
68% with the floor at 68 (target 90); more gate cases are exercised by aspire-si's Stage 1 tests.
Quality gate
The hard gates (security, errors, docs, shipping hygiene) pass; see SHIP_GATE.md and SCORECARD.md.
Tests
python -m pytest -q
Or run everything the release gate runs: bash scripts/verify.sh.
License
MIT, © mcp-tool-shop.
Built by MCP Tool Shop
