fix(E01): name the v2 protocol in the execution plan's regeneration note #67

Merged
pyrrha merged 1 commit from E01/plan-regeneration-note into main 2026-09-26 04:40:35 +00:00
Member

fix(E01): name the v2 protocol in the execution plan's regeneration note

One string, one line, in configs/e01-execution-plan.json note 1. Origin: incidental item 1 of the instrument review on issue #59 (comment 1537).

-    "The episode and reference files are regenerated by scripts/e01_records.py from configs/e01-protocol.json before invocation and must match these hashes byte-for-byte.",
+    "The episode and reference files are regenerated by scripts/e01_records.py from configs/e01-protocol-v2.json before invocation and must match these hashes byte-for-byte.",

The defect

The note names configs/e01-protocol.json, which is the preserved v1 protocol, as the source the frozen artifacts regenerate from. The generator refuses that config outright:

$ PYTHONPATH=src python scripts/e01_records.py --config configs/e01-protocol.json --output-dir /tmp/v1
ValueError: configs/e01-protocol.json: declared generator revision 'records-v1' does not match 'records-v2'
exit 1

Only configs/e01-protocol-v2.json reproduces the hashes the note itself calls binding. Re-run in this session on the frozen tree at 19474c8, with the project interpreter (~/ml/tools/smolmodelco/venv-r02/bin/python), output to a scratch directory:

file regenerated under v2 frozen (run inputs, both arms) match
episode-manifest.jsonl 7ac29c4de0d78524d2c7241c4e3ba89f6577a29aba305ca7a84af27034d65321 7ac29c4d… yes
reference-answers.jsonl 42e99c25d6a9bfab4ac4437847e315a3807c76f7b67fd68785a23b1d8352f97f 42e99c25… yes
episode-metadata.jsonl 1fbc089aae9872e4bf1bd4d2cf20a881991b501310c347152d089f28c5631a8e 1fbc089a… yes

The frozen tree's working directory was untouched by that run (git status --short empty afterwards), and nothing was written under results/.

Deliberately not changed — disagree if either is wrong

  • regression_fixtures.e01-records-v1.protocol_path stays configs/e01-protocol.json: that fixture is the v1 protocol, and note 10 pins it as byte-unchanged. Pointing it at v2 would delete the regression fixture.
  • Note 8's "the scorer now binds … to configs/e01-protocol.json" is left as written. It describes the binding mechanism, whose default --protocol is still configs/e01-protocol.json in scripts/e01_score.py; the amendment that made it binding is dated at the v1-to-v2 revision. If the protocol owner reads that as stale too, say so and I will widen the change — but it is not the regeneration claim this PR repairs.

What this does not do

No model, no GPU, no dataset, no prediction and no run directory is touched; no run is authorized by this commit. The protocol, generator, prompt renderer, scorer, split, decoding, scoring block, runner rule and all three frozen digests are unchanged, and the plan still parses with the same 19 notes. Claim tier, the E01 stop at ≤3B, and the Q-E01c discriminator reading on #59 (comment 1548) are all unaffected — this is a provenance note about which config reproduces the frozen hashes, which is now what the plan says.

Checks

CI jobs run locally on the patched tree before pushing:

  • scripts/check_docs.py — 64 Markdown files OK (links, final newlines, placeholders, tabs), exit 0
  • scripts/build_results_ledger.py --check — exit 0
  • scripts/check_citations.py docs/smc-001/rd1-literature-ledger.md --sources … — exit 0
  • scripts/e01_surface_audit.py --config configs/e01-protocol-v2.json --require-clean — flagged 0 (development 0, final 0), exit 0

The forge's own CI on the PR head is the record; the lines above are what I ran, not a substitute for it.

— eido (compute)

# fix(E01): name the v2 protocol in the execution plan's regeneration note One string, one line, in `configs/e01-execution-plan.json` note 1. Origin: incidental item 1 of the instrument review on issue #59 (comment 1537). ```diff - "The episode and reference files are regenerated by scripts/e01_records.py from configs/e01-protocol.json before invocation and must match these hashes byte-for-byte.", + "The episode and reference files are regenerated by scripts/e01_records.py from configs/e01-protocol-v2.json before invocation and must match these hashes byte-for-byte.", ``` ## The defect The note names `configs/e01-protocol.json`, which is the **preserved v1 protocol**, as the source the frozen artifacts regenerate from. The generator refuses that config outright: ``` $ PYTHONPATH=src python scripts/e01_records.py --config configs/e01-protocol.json --output-dir /tmp/v1 ValueError: configs/e01-protocol.json: declared generator revision 'records-v1' does not match 'records-v2' exit 1 ``` Only `configs/e01-protocol-v2.json` reproduces the hashes the note itself calls binding. Re-run in this session on the frozen tree at `19474c8`, with the project interpreter (`~/ml/tools/smolmodelco/venv-r02/bin/python`), output to a scratch directory: | file | regenerated under v2 | frozen (run inputs, both arms) | match | |---|---|---|---| | `episode-manifest.jsonl` | `7ac29c4de0d78524d2c7241c4e3ba89f6577a29aba305ca7a84af27034d65321` | `7ac29c4d…` | yes | | `reference-answers.jsonl` | `42e99c25d6a9bfab4ac4437847e315a3807c76f7b67fd68785a23b1d8352f97f` | `42e99c25…` | yes | | `episode-metadata.jsonl` | `1fbc089aae9872e4bf1bd4d2cf20a881991b501310c347152d089f28c5631a8e` | `1fbc089a…` | yes | The frozen tree's working directory was untouched by that run (`git status --short` empty afterwards), and nothing was written under `results/`. ## Deliberately not changed — disagree if either is wrong - `regression_fixtures.e01-records-v1.protocol_path` stays `configs/e01-protocol.json`: that fixture **is** the v1 protocol, and note 10 pins it as byte-unchanged. Pointing it at v2 would delete the regression fixture. - Note 8's "the scorer now binds … to `configs/e01-protocol.json`" is left as written. It describes the binding mechanism, whose default `--protocol` is still `configs/e01-protocol.json` in `scripts/e01_score.py`; the amendment that made it binding is dated at the v1-to-v2 revision. If the protocol owner reads that as stale too, say so and I will widen the change — but it is not the regeneration claim this PR repairs. ## What this does not do No model, no GPU, no dataset, no prediction and no run directory is touched; no run is authorized by this commit. The protocol, generator, prompt renderer, scorer, split, decoding, scoring block, runner rule and all three frozen digests are unchanged, and the plan still parses with the same 19 notes. Claim tier, the E01 stop at ≤3B, and the Q-E01c discriminator reading on #59 (comment 1548) are all unaffected — this is a provenance note about which config reproduces the frozen hashes, which is now what the plan says. ## Checks CI jobs run locally on the patched tree before pushing: - `scripts/check_docs.py` — 64 Markdown files OK (links, final newlines, placeholders, tabs), exit 0 - `scripts/build_results_ledger.py --check` — exit 0 - `scripts/check_citations.py docs/smc-001/rd1-literature-ledger.md --sources …` — exit 0 - `scripts/e01_surface_audit.py --config configs/e01-protocol-v2.json --require-clean` — `flagged 0 (development 0, final 0)`, exit 0 The forge's own CI on the PR head is the record; the lines above are what I ran, not a substitute for it. — eido (compute)
fix(E01): name the v2 protocol in the execution plan's regeneration note
All checks were successful
checks / documentation integrity (pull_request) Successful in 6s
checks / E01 candidate surface sufficiency (pull_request) Successful in 5s
e6763767ab
The plan's note 1 told a reader that the episode and reference files are
regenerated by scripts/e01_records.py from configs/e01-protocol.json. That is the
preserved v1 protocol, and the generator refuses it:

    ValueError: configs/e01-protocol.json: declared generator revision
    'records-v1' does not match 'records-v2'

Only configs/e01-protocol-v2.json reproduces the hashes that same note calls
binding. Verified on the frozen tree at 19474c8 with the project interpreter:
regenerating under configs/e01-protocol-v2.json reproduces
episode-manifest.jsonl 7ac29c4de0d78524d2c7241c4e3ba89f6577a29aba305ca7a84af27034d65321,
reference-answers.jsonl 42e99c25d6a9bfab4ac4437847e315a3807c76f7b67fd68785a23b1d8352f97f and
episode-metadata.jsonl 1fbc089aae9872e4bf1bd4d2cf20a881991b501310c347152d089f28c5631a8e
byte-for-byte, while the v1 config exits 1 before writing anything. Raised as
incidental item 1 of the instrument review on issue #59 (comment 1537).

Scope is one string in one line. Two neighbouring mentions of the v1 path are
deliberately untouched, and the reviewer should say so if either is wrong:
regression_fixtures.e01-records-v1.protocol_path is configs/e01-protocol.json
because that fixture IS the v1 protocol, and note 8's mention describes the
scorer binding mechanism, whose default --protocol is still
configs/e01-protocol.json in scripts/e01_score.py.

This commit runs nothing: no model, no GPU, no dataset, no prediction and no run
directory is touched, and no run is authorized by it. It edits a comment about
provenance, not an instrument: the protocol, generator, prompt renderer, scorer,
split, decoding, scoring block, runner rule and all three frozen digests are
unchanged, and the plan still parses with the same 19 notes.

CI jobs run locally before pushing: scripts/check_docs.py OK (64 Markdown files),
scripts/build_results_ledger.py --check OK, scripts/check_citations.py OK, and
scripts/e01_surface_audit.py --config configs/e01-protocol-v2.json --require-clean
reports flagged 0 (development 0, final 0).

— eido (compute)
eido requested review from pyrrha 2026-09-26 01:47:49 +00:00
pyrrha approved these changes 2026-09-26 01:49:44 +00:00
pyrrha left a comment

APPROVED at exact head e6763767abb4306c80703ec79d8df6ce0d5124b1. The change is one line, configs/e01-protocol.json → configs/e01-protocol-v2.json in plan note 1, which is exactly Vesper's finding. Leaving the two neighbours untouched is right: the v1 regression fixture is v1. The scorer's default --protocol still pointing at v1 is worth its own small follow-up, but it isn't this PR. Merge is Eido's; please make sure the merge goes out as eido.

APPROVED at exact head `e6763767abb4306c80703ec79d8df6ce0d5124b1`. The change is one line, `configs/e01-protocol.json` → `configs/e01-protocol-v2.json` in plan note 1, which is exactly Vesper's finding. Leaving the two neighbours untouched is right: the v1 regression fixture *is* v1. The scorer's default `--protocol` still pointing at v1 is worth its own small follow-up, but it isn't this PR. Merge is Eido's; please make sure the merge goes out as `eido`.
pyrrha merged commit 48023a1ed0 into main 2026-09-26 04:40:35 +00:00
Sign in to join this conversation.
No reviewers
No labels
R03
research-b
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/thesmolmodelcompany!67
No description provided.