| # Test transcripts — October 2026 bench |
|
|
| This archive holds every transcript, verdict and locked rubric behind the test results published on the site |
| (prompt library, data report and methodology page). Everything in it is fictional test material: the sample inputs |
| were written for the bench and contain no real clients, patients or cases. |
|
|
| Licence: everything in this archive (prompts, inputs, rubrics, transcripts, verdicts, results and the Claude Skills |
| texts) is published under Creative Commons Attribution 4.0 International (CC BY 4.0, |
| https://creativecommons.org/licenses/by/4.0/). Credit: AI Professions (aiprofessions.ai). Also on GitHub: |
| https://github.com/aiprofessions/prompt-benchmark |
|
|
| ## What is inside |
|
|
| | Folder | Content | |
| |---|---| |
| | `prompts/<prompt_id>/spec.yaml` | The prompt, its fictional input, the 5-criterion rubric, `round`, and the rubric hash locked before the first run | |
| | `prompts/<prompt_id>/<model>-runN.md` | One transcript per run: header (model id, provider, start time UTC, duration, exit code, rubric hash), the exact query, the verbatim output | |
| | `prompts/<prompt_id>/grades.json` | Per-criterion verdicts (pass/fail + one-line reason) for every run, with the judge model | |
| | `prompts-es2/<prompt_id>/` | ES round 2 (10 new Spanish prompts, 2026-10-21): same layout as `prompts/` | |
| | `tools/<prompt_id>/` | Excel / Microsoft 365 task round: same layout as `prompts/` (spec, transcripts, grades, calibration re-grades) | |
| | `prompts/<prompt_id>/calib-<model>-runN.json` | Judge-swap re-grades (only for the runs sampled for calibration) | |
| | `bench/results.json` | Aggregated results per prompt and model | |
| | `bench/rubrics.lock.json` | SHA-256 of (prompt, input, rubric) per prompt and the lock time | |
| | `bench/calibration.json`, `bench/calibration-frpt.json` | Judge-swap calibration, round 1 (EN/ES) and round 2 (FR/PT) | |
| | `bench/spotcheck.json`, `bench/spotcheck-frpt.json` | Spot-check of judge verdicts by the editor agent, round 1 and round 2 | |
| | `bench/*-fable*.json`, `bench/*-argon*.json` | Calibration and spot-check for the added models (Claude Fable 5.1, Gemini 4 Argon) | |
| | `bench/*-tools.json` | Excel / Microsoft 365 round: results, rubric lock, calibration, spot-check | |
| | `bench/*-es2.json` | ES round 2: results, rubric lock, calibration, spot-check | |
| | `skills/<skill>/<lang>/` | Claude Skills round: current transcripts and per-run grades (`grade-<model>-runN.json`) | |
| | `skills/<skill>/<lang>/v1.0/`, `v1.1/` | ARCHIVED earlier skill versions (not the published text): the `SKILL.md` that was tested and its transcripts and grades | |
| | `skills/<skill>/<lang>/stale/` | Runs set aside because the skill text embedded in the query no longer matched the current `SKILL.md`; they were re-run and are kept for completeness | |
| | `skills/_bench/` | Skill specs and rubrics (`specs.yaml`; FR/PT: `specs-frpt.yaml`), rubric locks, results (all four languages), revision findings | |
|
|
| Counts: 80 prompts, 1200 prompt transcripts; 10 Excel / Microsoft 365 prompts, |
| 150 transcripts; 10 ES round 2 prompts, 150 transcripts; 10 skills in |
| English and Spanish: 180 current skill transcripts, 114 from earlier |
| skill versions, 33 set-aside runs; the same skills in French and Brazilian Portuguese: |
| 180 current transcripts, 45 from earlier versions. |
|
|
| ## Rounds and dates (UTC, start times of the runs) |
|
|
| - Round 1 — 30 English + 10 Spanish prompts: 2026-10-19 11:34 to 2026-10-19 12:20. |
| - Round 2 — 20 French + 20 Brazilian Portuguese prompts: 2026-10-19 18:59 to 2026-10-19 20:09. |
| - Claude Skills round, English and Spanish (current versions): 2026-10-19 14:50 to 2026-10-19 16:13. |
| - Fable round — `claude-fable-5-1` added on the SAME locked material: all 80 prompts of rounds 1 and 2 (240 transcripts) and the current Claude Skills (60 transcripts): 2026-10-20 09:05 to 2026-10-20 10:41. |
| - Gemini round — `gemini-4-argon-medium` added on the SAME locked material: all 80 prompts of rounds 1 and 2 (240 transcripts): 2026-10-20 11:16 to 2026-10-20 11:54. Not run on the Claude Skills (a Claude format). |
| - Excel / Microsoft 365 task round — 10 new prompts (4 English, 2 Spanish, 2 French, 2 Brazilian Portuguese; profession `any`, tools `excel` / `microsoft-365`), all 5 models x 3 runs (150 transcripts): 2026-10-20 12:33 to 2026-10-20 12:43. Text-only answers: formulas were judged by their logic and the stated result on the given data, not executed in Excel. These are chat models answering a pasted task, not Copilot inside Excel or Microsoft 365. |
| - ES round 2 — 10 new Spanish prompts (separately locked set `bench/rubrics.lock-es2.json`), all 5 models x 3 runs (150 transcripts, folder `prompts-es2/`): 2026-10-21 04:38 to 2026-10-21 05:17. Same judges, judge-swap calibration (`bench/calibration-es2.json`) and spot-check (`bench/spotcheck-es2.json`) as the other rounds. |
| - Claude Skills FR/PT round — French (France) and Brazilian Portuguese adaptations of the 10 skills, `claude-opus-5-5`, `claude-sonnet-5`, `claude-fable-5-1` x 3 runs, graded by `gpt-6-astra` on rubrics locked separately (`skills/_bench/specs-frpt.yaml`, `skills/_bench/rubrics.lock-frpt.json`); current versions: 180 transcripts, 2026-10-21 05:41 to 2026-10-21 09:55. Four adaptations were revised after the first run and re-tested on the same locked input; the earlier texts and their runs are kept as evidence in `v1.0/` / `v1.1/` folders (construction-daily-report-delay-notice/fr/v1.0, construction-daily-report-delay-notice/fr/v1.1, event-planners-vendor-rfp/fr/v1.0, event-planners-vendor-rfp/pt/v1.0, sales-managers-pipeline-review/fr/v1.0; 45 transcripts). No judge-swap calibration or spot-check was run on this round. |
|
|
| ## Models |
|
|
| | Model id | Provider (as recorded) | Role | |
| |---|---|---| |
| | `claude-opus-5-5` | anthropic | tested model; judge of `gpt-6-astra` outputs | |
| | `claude-sonnet-5` | anthropic | tested model | |
| | `gpt-6-astra` | openai | tested model; judge of Claude outputs (prompts and skills) | |
| | `claude-fable-5-1` | anthropic | tested model (added 2026-10-20, prompts and skills) | |
| | `gemini-4-argon-medium` | google | tested model (added 2026-10-20, prompts only) | |
|
|
| Claude and GPT runs were sent through an agent command-line client in safe mode (tools, web access, memory and custom |
| instructions off). The client adds its own generic assistant system instruction, identical for every model. |
| Claude models were reached through a routing service on a subscription, `gpt-6-astra` through an OpenAI subscription |
| sign-in. |
|
|
| `gemini-4-argon-medium` was reached through Google's Antigravity command-line agent, not a plain API call. Each run used |
| a custom agent with no tools, skills or rules loaded and a one-line system prompt ("You are a helpful assistant."), |
| started in a fresh empty folder in sandbox mode with permission prompts not auto-approved. The client still wraps the |
| query with the local time and a model-selection note, and it still exposes one built-in tool (listing or stopping |
| background tasks). After each run the client's own log was read: a run with any tool call would not have counted as |
| completed. In this archive: 0 tool calls, 1 turn per run. The `agy_usage` header line is the token count the client |
| reported. Nothing was edited by hand between the query and the output. |
|
|
| ## How the rubrics were locked |
|
|
| Before any model saw a prompt, its rubric was written and the SHA-256 of (prompt, input, rubric) was stored in |
| `bench/rubrics.lock.json` (skills: `skills/_bench/rubrics.lock.json`, hash of skill, language, input and rubric). |
| The harness refused any run whose hash did not match. Every transcript header carries `rubric_sha256`; skill |
| transcripts also carry `skill_sha256`, the hash of the exact `SKILL.md` text embedded in the query. |
| You can recompute a prompt hash as |
| `sha256(json.dumps([prompt, input, rubric], ensure_ascii=False, sort_keys=True))`. |
|
|
| ## How the judging worked |
|
|
| - Cross-family: `gpt-6-astra` graded Claude outputs and Gemini outputs, `claude-opus-5-5` graded `gpt-6-astra` outputs. |
| The skills round has only Claude models, so all skill runs were graded by `gpt-6-astra`. |
| - One strict JSON verdict per criterion (pass/fail + one-line reason); an unsure judge had to mark fail. |
| - Calibration: a seeded sample of runs was re-graded by the other judge (`bench/calibration*.json`; the added |
| models have their own files `bench/calibration-fable-*.json`, `bench/calibration-argon-*.json`, and the Excel / |
| Microsoft 365 round `bench/calibration-tools.json`). |
| - Spot-check: the editor agent operating the bench (an AI agent, not a human) wrote its own verdict from the full |
| transcript before reading the judge's (`bench/spotcheck*.json`). |
| - No judge-swap calibration or spot-check was run on the skills rounds (English/Spanish incl. the Fable runs, and |
| French/Portuguese). |
|
|
| ## What was removed (sanitizing) |
|
|
| Command lines, session and conversation ids, local file paths and run-meta files were removed; provider names are |
| reduced to the vendor (`anthropic`, `openai`, `google`). Queries and outputs are otherwise verbatim. |
|
|
| ## What a pass does not mean |
|
|
| A pass means an output met these rubric criteria on this fictional input. It is not a professional sign-off, |
| not a test of the consumer apps (ChatGPT, Claude.ai, Copilot), and not a guarantee for other inputs. |
|
|
