[rlntel]

Adds the GovKM fixture contribution invited in #2 and Simone’s October 19 agreement: #2 (comment)

Proposes an experimental track outside Core: deterministic seed/count generator, 12 wholly fictional public cases across six families, separate public-example expected outcomes, source/output hashes, and 17 non-scored validation tests. GovKM is fixture author. The synthetic contribution is offered under this repository’s MIT license. ADEBench retains scoring, metrics, experimental leaderboard and private holdout custody.

All 9 approved v0.2 candidate files are preserved byte for byte; SUBMISSION.md records subsequent submission status and boundaries. The diff adds 10 files only under experimental/authority_resolution/. No Core, CI, dependency, scorer, adapter or leaderboard changes. Directory/schema and public-oracle placement remain proposed for maintainer review.

Validation: Python 3.12.14; python -B -m unittest discover -s tests -v — 17/17 PASS. Public seed 20261022/count 2 reproduces case/oracle bytes. Original metadata records Python 3.13.5; replay changes runtime metadata and its manifest hash only. Cases SHA-256: f6c2515f64ed6ddc2226cfef19c0e9db0426092893ee8d211195cbd119f1e0a5.

No scored benchmark, native AI-memory/model experiment, private holdout, partner records or proprietary GovKM implementation. expected_outcomes.json is a public-example evaluator oracle, never memory/model input. source_inventory_text is unscored inventory; only ADEBench’s independently captured delivered_text_after_cut can support later scores.

For maintainer review: integration/schema and oracle visibility, model-visible coverage proof for no-authority cases, temporal boundaries, semantic matching and cut rules. No claim of adoption before merge.

[adecubed]

Thanks, Rob. This is careful work: the generator is small, dependency-free and reproducible (seed 20261022 regenerates your cases byte for byte here, on 3.11 too), the holdout refusal is the right call, and keeping source_inventory_text and expected_outcomes.json out of anything scored is exactly the boundary we agreed.

Three changes before this can measure anything:

One record per fact, mixed with others. Today each case reaches a memory as a single block (source_inventory_text), so a memory either returns all of it or none of it, and the dimensions you designed (value, authority outcome, source event) can’t come apart. Could the generator emit each grant, event and the coverage registry as separate records, each with its own valid_at/recorded_at, written into one store together with the records of the other cases? The simplest target is adebench’s set format (sets//world.json with facts/episodes, plus cases/; see adebench/sets.py and sets/public2/): every adapter’s importer already reads it, so the track would run on all 15 memories without new import code.
Invented values. Blue, Dawn, Jade and the other values come from fixed lists of 12 common words, which collide with ordinary text under whole-token matching, and a private set drawn from the same lists protects little. Values, names and scopes generated from the seed as invented tokens (e.g. Velmora, Quintrel) would fix both.
Docs. README.md and CONTRIBUTION_NOTES.md still describe a candidate that is “not submitted”, and the hashes live in two places (SOURCE_SHA256.json and the manifest). One README describing what this PR is would be enough. I don’t need the provenance of the earlier drafts in the repo.
The scorer and the delivered_text_after_cut capture are on our side, as agreed. Once the fixtures are per-record, I’ll write the experimental section against them and run it on the board’s memories.