CORE CULP reproduction

A native Codex session reproduced the graph-classification capsule and returned all six requested answers. Its report also explains why four printed predictor labels do not describe the predictors the source actually called.

Dataset	Requested printed label	Actual predictor	Accuracy
Iris	CN	CS	100%
Iris	AA	CS	100%
Zoo	CN	CS	100%
Zoo	AA	CS	100%
Wine	CN	CN	97.22%
Wine	AA	AA	97.22%

The unchanged original CORE-Bench scorer accepts 6/6 answers. The source’s Iris/Zoo label defect is preserved and explicitly interpreted; these rows do not measure actual CN/AA calls. Only the two absolute data paths were adapted, with their required Path imports. All other capsule source/data bytes were retained. See the protected scoring receipt and controller review.

The agent completed the initial brief without subsequent coordinator messages. It preserved a deliberately interrupted Iris attempt, recovered under a new ID, and independently found and corrected a self-loop error in its own graph checker. The source-generated graphs and numerical answers were unchanged by that repair. Fresh-environment execution, raw-only reanalysis, and ordinary script execution without the capture hook agree. The fixed transductive splits provide no cross-split or population uncertainty estimate; no general superiority claim is made. Review combines source inspection and deterministic checks with explicitly provisional model critique, not independent scientific peer review.

Execution and verification

The first fully verified Allagma package in the frozen run order is r01 (replicate 2). It is the illustrative package selected by the prespecified criteria. The 12-run comparison is complete; this individual result does not establish a workflow advantage.

The run consumed 76.26 seconds of scientific computation and 15.43 seconds of setup. Native wall time was 1,921.48 seconds. Exact token and measured resource fields are retained with the native receipt and controller ledger.

Restore the complete research package

The package index identifies the unchanged report, source, protocol, arrays, attempts, claims, review, reproduction commands and software wheels by hash. Restore into a new directory, hydrating exact software dependencies:

python3 evals/research-v0.3/retention.py restore \
  --source evals/research-v0.3/runs/r01/package \
  --destination /path/to/new-culp-package --download-wheels

An exact local wheel cache may be supplied instead of downloading. After restoration, read REPORT.md, REPRODUCE.md and submission.json. Use the standalone broker driver to service the delivered reproduction command under a new finite validation profile. Run the full-study and raw-data recomputation commands separately and retain their outcomes. The clean-checkout release check passed: a new Git clone restored all exact files and dependencies, created a fresh environment, reran the complete study, and recomputed the retained-data results. Both answer sets equal the original package.

---

Canonical source · Generated from the maintained repository document at this build’s commit.