Skip to content

Assistant prompt trials

These records say which assistant was tried, whether its generated artifact ran, and how its results compared with the La₀.₅Ba₀.₅CoO₃ on HRPT tutorial. Another assistant run can take different steps; check the generated work and its results.

Published prompts, 2026-11-13

Codex assistant (exact backend model not exposed) read the actual published Python and command-line prompts, the linked worked references and the two input downloads. It produced a notebook and a native CLI project. Both executed all five cumulative fit stages and completed with findings.

Artifact Completed fits Final reduced χ² Final Rwp
Python notebook 5 1.30 7.23%
Native CLI project 5 1.30 7.23%

Both results match the published staged notebook's displayed results and all six unchanged FullProf scientific-reference rows within their original tolerances. The separate single-fit regression fixture's precise χ² and iteration count do not match this five-fit trajectory. The recorded comparisons retain those mismatches and every original reference value and tolerance; this is not an all-regression-pins pass. Display rounding is not a scientific tolerance.

The La/Ba displacement correlation was −1.00. The worked tutorial describes their opposite signs: “move together” in the prompt should be read as strong correlation in magnitude. Both artifacts then applied the requested equal-displacement constraint.

These executions used Linux diagnostic artifacts from Edi 80dc974c431aceee6f1dd4f3468847466b56d27a and SDK 2b6b981e21194f9a76074742dd0e148736ca4d9b. They do not qualify a published release binary, the current SDK candidate, or Windows. The documentation source was b16f9867ef1de464dfed731796848eb5ba60e739; the route was published-release, not a prerelease preview.

Trial provenance

Recorded evidence: tests/fixtures/tutorial_cases/published_assistant_trials.json, SHA-256 f9d892f37e88d7db1a10e7a095130c7de377595e8e84185d0eb53bfcba015ebe. It retains exact fetched prompts, documentation/input identities, execution captures and comparisons. The generated artifacts are retained as authoring evidence; no public download is claimed here.

  • Ready notebook lbco_hrpt_prompt.ipynb: SHA-256 d5f516b4e96086f3c37ebc9e21d4e145f74558fad4b4a5417e7cd2f9f535e960.
  • Ready project cli-ready-project.zip: SHA-256 4d89fd52f404b0b29cc92d8d709c503b6d0f96419a612cf265e7856f6d67caa7.
  • Python final raw result: SHA-256 8c3c606b017f59754f5530af4921e07388bb8ebf8bdc8be4d13e2ac88e75d1ac.
  • Native final machine report: SHA-256 c57958c6014291afe3a092c4de27ae6b9f70c37f24894e100a5ddb074dacd6bf.

Earlier attempts, 2026-11-09

These historical prerequisite failures did not execute a generated fit. They remain part of the record.

Interface Date Model Outcome Scientific comparison
Python notebook 2026-11-09 GPT-6 (Codex; exact backend variant not exposed) Failed: required tutorial pages returned HTTP 404 Not reached
Command line 2026-11-09 GPT-6 (Codex; exact backend variant not exposed) Failed: required tutorial pages returned HTTP 404 Not reached

In both attempts, the documentation root, assistant index and collection downloads were reachable. The required tutorial index and worked LBCO reference were unavailable. The attempts stopped there: no substitute checkout was used, no generated artifact was executed, and no fitted values were compared with the regression references. No preview trial is recorded by this evidence.

The prompts remain available for Python and the command line. A successful trial requires fetching the actual published references, executing the generated notebook or project, and comparing its scientific results with the tutorial references. A documentation build alone does not supply that evidence.

Historical provenance

The original trial records name prompt commit 557d6e703. Their prompt SHA-256 digests are:

  • Python: a5d78847928809368f87ed8260439bc4218fd1837248fc678b03dbac0eb84e1c.
  • Command line: e958568b03ff3447ab6ed7394fc6cfb5544ddcdd08dde7555f63b56cd18c521c.

The captured summary is identified by source commit 37e53a6cf1e38b88549a241717f1eb49d57499bb and SHA-256 85b40bef0e02d3b13b077e5d056c3d51cfc6f70f88072f8d1dec96a48a5da0ff. These identify the historical source; they do not identify the later trials above.