research: Marathon shadow corpus and Elsie handoff #3

Open
pyrrha wants to merge 7 commits from experiment/marathon-disambiguation-contract into main
Member

Scope

Publishes the shadow-only Marathon experiment, admitted exporter, research plan, bounded public seed and original operator annotations for Elsie to continue. No production changes or model inference.

Head: 2957940ee3e862293f7399056191e37dc4aac82e.

Evidence

127 unique seed IDs/URIs, 127 valid annotation joins: 81 marathon_game / 37 marathon_sport / 9 other. The original CSV, seed, import result, manifests and SHA256SUMS are committed byte-for-byte. All control decisions are unknown and model results are null. No semantic accuracy claim is made.

The corpus README flags source bias, semantic label-scope questions, missing quote contexts and timestamp limitations. Original labels remain untouched. HANDOFF-PROMPT.md is an Elsie-specific continuation brief: audit + adjudication queue + bounded hard-case supplement; stop before model calls until review.

Verification

  • python3 experiments/marathon-disambiguation/verify-corpus.py: integrity PASS, 127 joined rows, zero model calls, semantic labels NOT ADJUDICATED.
  • git diff --check and staged check: passed; scoped .gitattributes preserves original CSV BOM/CRLF bytes.
  • Docker Node 20, isolated VM experiment worktree at 8c620ed: yarn build; candidate helpers (14 PASS lines); test:description (3), test:filters (23), test:validate (19), test:subscription (24): all passed. Publication only adds Python integrity checking, corpus files and docs; TS/src/package/lockfile are unchanged from that tested head.
  • Public-record key audit excluded credential/viewer fields. No .env, database or media blobs included.

These are self-verification results, not an independent review or a classifier benchmark. No automatic merge requested. VM production checkout and live Marathon/running exclusions were not changed.

Publication-time read-only check found the current production filter hash differs from the recorded seed snapshot. This publication did not modify it; cause/time/content of the intervening change are unestablished. Elsie must locate the historical snapshot or mark it unavailable before replay. The production checkout remains b9bcce9 with only untracked matchtest.ts and the container running.

## Scope Publishes the shadow-only Marathon experiment, admitted exporter, research plan, bounded public seed and original operator annotations for Elsie to continue. No production changes or model inference. Head: `2957940ee3e862293f7399056191e37dc4aac82e`. ## Evidence 127 unique seed IDs/URIs, 127 valid annotation joins: 81 marathon_game / 37 marathon_sport / 9 other. The original CSV, seed, import result, manifests and SHA256SUMS are committed byte-for-byte. All control decisions are unknown and model results are null. No semantic accuracy claim is made. The corpus README flags source bias, semantic label-scope questions, missing quote contexts and timestamp limitations. Original labels remain untouched. `HANDOFF-PROMPT.md` is an Elsie-specific continuation brief: audit + adjudication queue + bounded hard-case supplement; stop before model calls until review. ## Verification - `python3 experiments/marathon-disambiguation/verify-corpus.py`: integrity PASS, 127 joined rows, zero model calls, semantic labels NOT ADJUDICATED. - `git diff --check` and staged check: passed; scoped .gitattributes preserves original CSV BOM/CRLF bytes. - Docker Node 20, isolated VM experiment worktree at 8c620ed: yarn build; candidate helpers (14 PASS lines); test:description (3), test:filters (23), test:validate (19), test:subscription (24): all passed. Publication only adds Python integrity checking, corpus files and docs; TS/src/package/lockfile are unchanged from that tested head. - Public-record key audit excluded credential/viewer fields. No .env, database or media blobs included. These are self-verification results, not an independent review or a classifier benchmark. No automatic merge requested. VM production checkout and live Marathon/running exclusions were not changed. Publication-time read-only check found the current production filter hash differs from the recorded seed snapshot. This publication did not modify it; cause/time/content of the intervening change are unestablished. Elsie must locate the historical snapshot or mark it unavailable before replay. The production checkout remains b9bcce9 with only untracked matchtest.ts and the container running.
Owner

Nice, self-checking corpus! verify-corpus.py passes (127 rows, integrity PASS), SHA256SUMS verifies, and testCandidate.ts passes locally. A few small things:

  1. .gitattributes only protects humanedited.csv, but SHA256SUMS also covers the four JSONL and manifest files. On a checkout with core.autocrlf, verify-corpus.py:25 would fail on hash mismatch. Suggest experiments/marathon-disambiguation/corpora/** -text.
  2. verify-corpus.py:19 uses read_text() without an encoding, so it depends on the system locale. Passing encoding='utf-8' would match the CSV read at :29.
  3. tsconfig.json:17 only includes src/**, so candidate.ts and testCandidate.ts are never type-checked, and nothing runs them. A package.json script (e.g. "testMarathon": "ts-node experiments/marathon-disambiguation/testCandidate.ts") would keep them from rotting.
  4. The corpus stores other people's public post text and handles. Since it's committed, it might be worth a line in the corpus README on whether removal requests are honored.
Nice, self-checking corpus! `verify-corpus.py` passes (127 rows, integrity PASS), `SHA256SUMS` verifies, and `testCandidate.ts` passes locally. A few small things: 1. `.gitattributes` only protects `humanedited.csv`, but `SHA256SUMS` also covers the four JSONL and manifest files. On a checkout with `core.autocrlf`, `verify-corpus.py:25` would fail on hash mismatch. Suggest `experiments/marathon-disambiguation/corpora/** -text`. 2. `verify-corpus.py:19` uses `read_text()` without an encoding, so it depends on the system locale. Passing `encoding='utf-8'` would match the CSV read at `:29`. 3. `tsconfig.json:17` only includes `src/**`, so `candidate.ts` and `testCandidate.ts` are never type-checked, and nothing runs them. A `package.json` script (e.g. `"testMarathon": "ts-node experiments/marathon-disambiguation/testCandidate.ts"`) would keep them from rotting. 4. The corpus stores other people's public post text and handles. Since it's committed, it might be worth a line in the corpus README on whether removal requests are honored.
This pull request can be merged automatically.
You are not authorized to merge this pull request.
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u origin experiment/marathon-disambiguation-contract:experiment/marathon-disambiguation-contract
git switch experiment/marathon-disambiguation-contract

Merge

Merge the changes and update on Forgejo.

Warning: The "Autodetect manual merge" setting is not enabled for this repository, you will have to mark this pull request as manually merged afterwards.

git switch main
git merge --no-ff experiment/marathon-disambiguation-contract
git switch experiment/marathon-disambiguation-contract
git rebase main
git switch main
git merge --ff-only experiment/marathon-disambiguation-contract
git switch experiment/marathon-disambiguation-contract
git rebase main
git switch main
git merge --no-ff experiment/marathon-disambiguation-contract
git switch main
git merge --squash experiment/marathon-disambiguation-contract
git switch main
git merge --ff-only experiment/marathon-disambiguation-contract
git switch main
git merge experiment/marathon-disambiguation-contract
git push origin main
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
smolmodelco/homespun-feedgen!3
No description provided.