Skip to content

Sprint 080: required CUDA matrix, judge and audit of merged evidence - #28

Merged
jxucoder merged 3 commits into
mainfrom
claude/serene-dirac-ow6g1x
Nov 15, 2026
Merged

jxucoder merged 3 commits into
mainfrom
claude/serene-dirac-ow6g1x

Conversation

@jxucoder

Copy link
Copy Markdown
Owner

Summary

Continues Sprint 080 after the foundation checkpoint merged. All eight required device recipes are implemented, but judged against the sprint's own acceptance, no required cell has complete real-device evidence at the merged source. This PR adds the required matrix, an independent judge and that audit. It executes no CUDA test and changes no production code.

What changed

  • Required matrix (v1-sprints/080-required-device-matrix.json): eight cells derived from the canonical R table (R1 squared, binary, multiclass; R4 Poisson; R5 AFT; R6 Normal; R8 independent and shared), eight obligations each, shared operation groups and CPU-host controls.
  • Judge (benchmarks/v1/required_device_judge.py): standard library only. It rejects a missing recipe, an unexecuted, skipped or failed bound case, and evidence without a retained device observation. The required cells are fixed in the judge, so a matrix edit cannot shrink the scope.
  • Categorical rejection tests: device preparation refused categorical features with no test. Adds CPU-host contracts and three real-device cases.
  • Audit and records: benchmarks/v1/evidence/required-device-080/, the public required-cell table in docs/v1/index.md, the Sprint 080 card, the sprint index, AGENTS.md current position and a learning entry.

Audit result

The repository collects 1,900 GPU-marked cases; the matrix binds 1,362. The T4 run retained for 312d821 executed 441 cases, all passing, of which 322 are bound. The other 1,040 bound cases are unexecuted at that source. None failed or skipped.

Cell Recipe trajectories All obligations
Squared regression 0/13 0/109
Binary classification 11/11 16/89
Multiclass classification 30/30 36/153
Poisson with exposure 11/11 16/88
Event/right-censored AFT 39/39 46/188
Normal, ordinary/Fisher 90/100 91/258
Multi-output, independent trees 14/14 20/63
Multi-output, shared topology 14/14 32/75

Verification

  • 39 new CPU tests pass on Python 3.12 and 3.13.
  • ruff check src/openboost tests/v1 tests/conftest.py benchmarks/v1/required_device_judge.py is clean; mkdocs build --strict passes.
  • Local CPU regression adds 39 passes and no new failure. Five tests fail identically before and after these changes: four on Python 3.13 (dataclasses.replace raises TypeError there) and one resource-limited worker test in the local Python 3.12 container. They were not fixed here.
  • These are development checks in a container without a GPU, not gate evidence.

Not verified

  • No CUDA test, Modal job or upload ran. The three new GPU cases are untested code.
  • Bindings are at test-function granularity, and the judge trusts the supplied JUnit files.

Before this can merge

The v1 Modal Validation Gate is expected to fail on this PR: it binds the committed Modal receipts to every tracked file outside the PR 27 evidence directory, and this PR changes tracked inputs. It needs a refreshed bounded Modal validation of this candidate. That run can also execute the 1,040 unexecuted bound cases, which is Sprint 080's next gate; it needs a concrete hardware and upload allowance.


Generated by Claude Code

Both device preparation entry points refuse categorical features, but no
test covered that required rejection. Add CPU-host contracts that call the
undecorated validation without an execution context, plus three real-device
cases that are collected and unexecuted.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BnXy66RoaZmf6m21fXGm1h
Derive the eight required device cells from the canonical R table and bind
each obligation to collected GPU cases. The standard-library judge rejects
a missing recipe, an unexecuted, skipped or failed case and an absent
device observation.

Judging the T4 run retained for 312d821 leaves every cell open: 322 of
1,362 bound cases executed and passed, and 1,040 are unexecuted at that
source. The public device table is derived from this audit.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BnXy66RoaZmf6m21fXGm1h
Update the sprint card, sprint index and agent guide after the PR 27 merge:
all eight required device recipes are implemented, every required cell is
open against the retained T4 evidence, and the next gate is one frozen
real-device run with a concrete allowance.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BnXy66RoaZmf6m21fXGm1h
@reviewdocs

reviewdocs Bot commented Nov 15, 2026

Copy link
Copy Markdown

📄 Review document changes in ReviewDocs

Open the rendered review →

7 documents changed:

  • AGENTS.md
  • benchmarks/v1/evidence/required-device-080/README.md
  • docs/v1/index.md
  • learnings/2026-11-15-v1-required-device-matrix.md
  • learnings/README.md
  • v1-sprints/080-cuda-required-recipes.md
  • v1-sprints/README.md

See the Markdown as formatted prose — tables, images and all — and comment on any paragraph. Comments and approvals sync back to this pull request.

@jxucoder
jxucoder marked this pull request as ready for review November 15, 2026 05:12
@jxucoder
jxucoder merged commit a638da1 into main Nov 15, 2026
0 of 4 checks passed
@jxucoder
jxucoder deleted the claude/serene-dirac-ow6g1x branch November 15, 2026 05:12
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 15, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-15T05:13:25.678734Z 9b7a0a8 Draft marked ready
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

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