Skip to content

bench: task-oneshape pack for gold-pack Arm C - #811

Draft
danieljohnmorris wants to merge 1 commit into
cursor/exp08-lean-excerpt-packsfrom
cursor/arm-c-oneshape-pack-7f04
Draft

danieljohnmorris wants to merge 1 commit into
cursor/exp08-lean-excerpt-packsfrom
cursor/arm-c-oneshape-pack-7f04

Conversation

@danieljohnmorris

Copy link
Copy Markdown
Collaborator

Summary

Draft pack so gold-pack Arm C can run on artefact-exp07. Arm B (--context task-modules) is unchanged. This PR does not spend a live key.

Canonical conditional, the only one the loaded excerpts teach:

?h cond then else

Nested else is parenthesised: (?h cond then else). HOF predicates stay >b comparison expressions. Other legal shapes are marked non-canonical in this pack only.

Invoke

Offline (no API key):

python3 scripts/closed-loop-bench.py --dry-run \
  --task-set artefact-exp07 --context task-oneshape

That loads oneshape_modules on the four artefact tasks:

task module context_chars
schedule-window ilo-excerpt-schedule-oneshape 837
typed-rollup ilo-excerpt-typed-oneshape 845
record-normalize ilo-excerpt-record-oneshape 965
csv-sales-summary ilo-excerpt-csv-oneshape 985

Arm B on the same set stays 659–777 chars (--context task-modules). Arm C is the same band (under 1k), not a curated balloon.

Live (parent only, after this offline note is accepted):

python3 scripts/closed-loop-bench.py \
  --provider deepseek --model deepseek-chat \
  --task-set artefact-exp07 \
  --context task-oneshape \
  --retry-cap 2 \
  --repair-shape-hint --header-recovery --fold-meta-stdout \
  --python --lang2-docs bench/closed-loop/context/python-control-docs.md \
  --output-dir bench/goldpack-v1-live-arm-c-YYYY-MM-DD/

Pins

Item Value
Base tip 7b15b28dac876f977d5280387e6ae09c5bb56b8f
Pack tip 2683d13bfa9fd05f67464495a93ef7a00701863c
ilo CalVer 26.5.0
Card bench/closed-loop/context/one-shape-conditional.md (34 lines, not loaded)
Card sha256 e3495fad4dd014e19e3ca34b2a27b1320fc570ec1859f2cbe46aedd63f12f98a
Run note bench/closed-loop/arm-c-run-note.md

The two shape lines in the card are copied byte-for-byte into each excerpt. Re-hash the card if that file changes before the live.

Constraints

  • Artefact-4 bank not widened. modules still names the Arm B excerpts.
  • No new header rail. Rails stay shape-hint + header-recovery + fold-meta-stdout.
  • No compiler change. Arm B's only code shared by two reds was ILO-P009 on @ (record-normalize and typed-rollup). Those two excerpts name the loop form. P007 was one task; P023 did not appear.
  • Pinned goldens-exp07 are untouched. ilo check / ilo run on 26.5.0 are green. schedule-window already uses ?h. csv (match), record (brace exists{cnt}), and typed (braceless guards) still use a non-canonical form. A ?h-only rewrite of those three matches the recorded stdout offline and is not committed — that rewrite belongs in its own PR.
  • ilo-next card mirror is not in this repo.

Tests

python3 scripts/test_closed_loop_context.py (20 tests, including the new arm). Dry-run of --context task-oneshape on artefact-exp07. Four goldens ilo check and ilo run.

Open in Web Open in Cursor 

Load one canonical ?h conditional on artefact-exp07 via --context
task-oneshape. Arm B task-modules stays the multi-shape excerpts.
Goldens are unchanged and still run on ilo 26.5.0.

Co-authored-by: Daniel Morris <danieljohnmorris@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants