| name | experiment-dossier-builder |
| description | Builds a portable, source-grounded Markdown dossier from a remote research project, including method state, experiment design, result ledgers, figures, failures, provenance, and paper-ready claim candidates. Use on the machine that contains the code and experiment outputs when preparing a handoff for local paper writing, collaborator review, or project archiving. |
Experiment Dossier Builder
Use this on the machine that can read the project repository and experiment
outputs. Produce one compact evidence handoff for a paper-writing agent; do not
write the paper or copy the whole project into the dossier.
Read references/dossier-contract.md before producing the final artifact.
Inputs
- Project repository, including local instructions and documentation.
- Experiment configs, commands, logs, tracker exports, tables, plots, and
checkpoints that are in scope.
- Optional target venue, current paper claims, and result priorities.
- Optional output path; default to
paper-handoff/experiment-dossier.md.
If the available outputs are too large to inspect fully, state the sampling
rule before selecting runs. Never silently summarize only the best runs.
Workflow
- Read project instructions, entry points, configs, evaluation code, and
result-generation paths before interpreting outputs.
- Record the repository snapshot: project name, commit, branch, dirty state,
run date range, environment evidence, and scope inspected. Mark unavailable
fields explicitly.
- Build a project map that explains the research question, method components,
training or control pipeline, datasets/environments, evaluation path, and
important implementation assumptions. Cite repository-relative paths.
- Inventory experiment families before selecting headline results. Group runs
by research question, baseline, ablation, seed, dataset/task, and hardware or
simulator condition where relevant.
- Build a result ledger. For every number or trend that may enter the paper,
record its exact source artifact, config/run identity, metric definition,
aggregation, uncertainty, and provenance status.
- Trace transformations from raw outputs through analysis scripts to tables
or figures. Use
experiment-provenance-auditor when the chain is incomplete
or publication-grade reproducibility is requested.
- Extract only evidence-supported candidate claims. Link each claim to result
rows and record counterevidence, scope limits, and confidence.
- Capture negative results, failed runs, known bugs, confounders, surprising
observations, and missing comparisons. These are paper-planning evidence,
not cleanup noise.
- Inventory useful figures and tables with repository-relative source paths,
generation commands when known, and the scientific message each supports.
- Write the dossier using the contract. End with prioritized evidence gaps
and concrete next experiments, not generic future work.
Evidence Status
Use only these labels:
verified: directly supported by inspected artifacts with a traceable chain;
derived: recomputed from verified inputs with the transformation recorded;
ambiguous: plausible source exists but identity or transformation is not
uniquely established;
missing: required evidence was not found;
not-applicable: the field genuinely does not apply.
Do not use verified merely because a filename, README statement, or tracker
run name resembles the claimed experiment.
Portability Rules
- Use repository-relative paths and stable run or artifact ids in the dossier.
- Sanitize credentials, private hostnames, participant data, and local absolute
paths. Preserve a useful non-sensitive locator instead.
- Do not embed large logs, checkpoints, datasets, or generated binaries.
- If figures must travel with the dossier, list the smallest selected bundle
under
paper-handoff/assets/; do not copy files unless the user requests it.
- Record the transfer mechanism as out of scope. Never assume SSH credentials
or upload project artifacts without explicit authorization.
Output
Produce paper-handoff/experiment-dossier.md by default. Keep the narrative
concise and make ledgers complete enough that the local agent does not need the
remote filesystem to understand the evidence.
The dossier must contain:
- snapshot and inspection scope;
- research question and method map;
- experiment matrix and result ledger;
- claim-evidence candidates;
- figure/table inventory;
- failures, caveats, and negative results;
- provenance and environment summary;
- prioritized evidence gaps;
- source index and handoff manifest.
Validation
- Every headline claim points to one or more result-ledger rows.
- Every result row has a source locator and evidence status.
- Metric definitions, aggregation, seeds/sample counts, and uncertainty are
recorded or explicitly missing.
- Results are not cherry-picked without a declared selection rule.
- Method descriptions point to inspected code/config evidence.
- The Markdown contains no credentials or machine-specific absolute paths.
- A local reader can distinguish fact, interpretation, and proposed narrative.
Completion
Done means one portable dossier exists, the inspected project snapshot is
identifiable, paper-relevant results are traceable or risk-labeled, and the
remaining evidence gaps are explicit.