Task folder (the PR)

Upload this before or with the first run. Same required files as experiment-task-guide.md §3.

experiments/<task-slug>/
  task/
    README.md
    experiment.json
    persona_strategy.json
    instruction.md
    task.toml
    input/
      context.md
      questionnaire.yaml           # survey
      self_report_schema.yaml      # web / self-report
    tests/
      test.sh
    variants/<levelId>/expected.json   # optional
  <experiment-id>/
    …

Do not put jobs/, .env, or trial output inside task/. If you later open a public GitHub PR, copy this task/ tree to experiments/tasks/<task-slug>/.

The four shipped exemplars are already here under task/ (survey-onboarding-framing, survey-upgrade-copy, survey-trust-by-segment, web-compare-options). New private studies add a sibling <task-slug>/task/.

Experiment vs model (reports)

A Playground experiment run is one persona model. That complete packet lives under models/<model-id>/ — including its JSON, CSV, and Save as PDF.

The study root is only shared contract + people. A report at this level is cross-model ablation, and only when you actually ran more than one model.

	One model	Several models (ablation)
`models/<id>/report.pdf`	The full experiment PDF	Each model's own experiment PDF
Study-root `report.pdf` / `report.json` / `scorecard.csv`	omit	Required — compare models on the same cohort and contract

Do not copy the same Playground PDF to both places.

Experiment-level files

These sit directly under experiments/<task-slug>/<experiment-id>/. The contract itself lives in sibling ../task/.

File / folder	Required	What it is
`README.md`	yes	Owner, hypothesis, N, who × what, which models, status
`personas/`	yes	Frozen people — see Who × what
`experiment.json`	only if launched ≠ `../task/experiment.json`	Snapshot as launched
`persona_strategy.json`	only if launched ≠ `../task/`	Snapshot as launched
`report.json` / `report.pdf` / `scorecard.csv`	multi-model only	Cross-model ablation

Who × what (one layout)

Do not invent a second persona tree per stimulus. Who is personas/<group>/. What is models/<m>/arms/<arm-id>/ — one Playground / Harbor arm (one job). The engine manifest also calls this object a cell (cells[].cellId); the folder is arms/ so it matches the UI. Do not invent stimulus__group names.

The experiment contract decides which persona cohort an arm used (personaArm / persona_strategy). That mapping is not a folder-naming problem.

Design	Who (`personas/`)	What (`models/<m>/arms/`)
Explore — one group, one version	Only `personas/shared/`	one arm
SUT — same people, several versions	Only `personas/shared/`	one folder per arm (`feature`, `outcome`, …)
PUT — several groups, one version	`personas/<group>/` (ids must not overlap)	usually one arm per group
Crossed — persona × stimulus	`personas/<group>/` (same as PUT)	system arms (`feature`, `outcome`). Group is `personaArm` + the YAML folder, not the path

personas/manifest.json lists who. Arms live under the model.

{
  "schema": "matraix.experiment.who-what/v1",
  "assignment": "shared",
  "seed": 42,
  "source": "matraix-persona-1m",
  "groups": { "shared": ["0042", "0108"] }
}

PUT / crossed: "assignment": "split" and one key per group in groups.

YAML lives only under personas/<group>/<id>.yaml. Never copy YAML into arms/. SUT must not have personas/feature/ or personas/outcome/.

# <experiment-id>

- Owner: <name>
- Task: survey-onboarding-framing
- Axis: SUT | PUT
- Hypothesis: (one sentence)
- N / assignment: 12 shared
- Models: gpt-5.6-sol, claude-opus-5.5, gemini-3.7-flash
- Status: complete for gpt-5.6-sol; opus in progress

Layers

Layer	Path	What it is
Task	`experiments/<slug>/task/`	GitHub PR
Experiment	`…/<experiment-id>/`	Shared people + optional ablation report
Model	`…/models/<model-id>/`	One complete experiment (that persona model)
Arm	`…/arms/<arm-id>/`	One experiment arm (Harbor job)
Trial	`…/arms/<arm-id>/trials/<persona-id>/`	One person on that arm

<model-id> = short name (gpt-5.6-sol, claude-opus-5.5, gemini-3.7-flash). Slash in personaModel → --.

Full tree

experiments/<task-slug>/
  task/                                # GitHub task PR
  <experiment-id>/
    README.md
    personas/
      manifest.json                    # who: groups only
      shared/0042.yaml                 # SUT: only this group
      cost-sensitive/0201.yaml         # PUT / crossed: one folder per group
    models/<model-id>/
      model.json
      manifest.json
      report.json
      report.pdf
      scorecard.csv
      arms/<arm-id>/
        arm.json                       # cellId + arm + factorLevels + personaArm + jobName
        job.json
        result.json
        report.pdf                     # Playground job PDF
        trials/<persona-id>/
          result.json
          verifier/...
          artifacts/<outputArtifact>

Model-level files

File	Required	What to put there
`README.md`	yes	Model id, provider, N completed, exclusions
`model.json`	yes	`personaModel`, `provider`, `deployment`
`manifest.json`	yes	`experiments/runs/<id>/manifest.json` for this model
`report.json`	yes	Full experiment JSON for this model (`matraix.experiment.report` v1)
`report.pdf`	yes	Playground Experiments → Export → Report PDF → Save as PDF
`scorecard.csv`	yes	`matraix experiment analyze --format csv`

{
  "personaModel": "azure/gpt-5.6-sol",
  "provider": "azure",
  "deployment": "gpt-5.6-sol"
}

Arm-level files

One Harbor job per experiment arm. Not a second copy of the people. Folder name = the arm id you already see in Playground (feature, outcome, …). Which cohort it used is personaArm.

File	Required	What it is
`arm.json`	yes	Manifest cell for this arm: `cellId`, `arm`, `factorLevels`, `personaArm`, `jobName`
`job.json`	yes	`matraix results <jobName> --format json` (`MatraixJobResults.v1`)
`result.json`	yes	Harbor job tokens / cost
`report.pdf`	yes	Playground job PDF (`*-persona-task-batch-report.pdf`)
`scorecard.csv`	optional	`matraix results --format csv`

A single trial has no report.

If the workbench launched one job per group for the same system arm, keep them under that arm id and record each job on arm.json. Do not mint feature__cost-sensitive paths.

Trial-level files

File	Required
`result.json`	yes
`verifier/structured_output.json`	yes
`verifier/reward.txt`	yes
`artifacts/survey_result.json`	survey
`artifacts/signup_decision.json`	web
`user_feedback.json`	if produced
`trajectory.json` (`steps[]` only)	web

No screenshots, no agent/ dumps, no .env.

Worked trees

SUT — one persona folder, several stimulus arms:

personas/shared/0042.yaml
personas/shared/0108.yaml
models/gpt-5.6-sol/arms/feature/trials/0042/
models/gpt-5.6-sol/arms/outcome/trials/0042/

PUT — one persona folder per group, one arm per group:

personas/cost-sensitive/0201.yaml
personas/premium-seeking/0310.yaml
models/gpt-5.6-sol/arms/cost-sensitive/trials/0201/
models/gpt-5.6-sol/arms/premium-seeking/trials/0310/

Crossed — same persona folders as PUT. Arms stay the experiment's system arms. Who is in which group is the YAML folder, not the arm name:

personas/cost-sensitive/0201.yaml
personas/premium-seeking/0310.yaml
models/gpt-5.6-sol/arms/feature/trials/0201/
models/gpt-5.6-sol/arms/feature/trials/0310/
models/gpt-5.6-sol/arms/outcome/trials/0201/
models/gpt-5.6-sol/arms/outcome/trials/0310/

What stays on disk (not this dataset)

agent/ is /logs/agent/ (raw logs, screenshots). Survey path is already in survey_result.json. Web: copy stripped trajectory.json only.

Leave local: agent/, images/, trial.log, config.json, job.log, compute.json, _generated/, .env, Persona 1M parquet.

License

MIT for the packaging and docs in this repo. Persona YAML is a frozen subset of the simulated catalog; treat it as study material, not records of real people.

Downloads last month: 214

Total file size:

757 MB