Earlier formats¶

P2P reads results and configurations written in the earlier result format, so work done with it stays readable. Nothing is migrated or rewritten to be viewed: the readers below are read-only.

Earlier-format runs¶

An earlier-format run directory holds regression_config.yaml, the survey, target and response files it references, exported weights (updated_endowments.csv), summary_aggregate.csv, token_usage_summary.csv and archived figures.

p2p inspect --run <directory> reads such a run and reconstructs the selected model’s question-level predictions from the survey, the target, the responses and the exported weights. It imports the summary and usage tables and embeds the archived prediction and diagnostic figures. It never imports old Python modules or loads a joblib pickle, and a run missing some inputs still shows every view its files support and names what is missing. Write the page outside the source run:

p2p inspect \
  --run path/to/earlier_run \
  --output outputs/inspect/earlier_run.html \
  --advanced

The Studio lists earlier-format runs found under outputs/paper/ with an earlier format badge, and result folders with Markdown summaries and figures under Result folders. Their run pages show the same views as current runs, plus an Original report files tab.

The Python reader is p2p.earlier_runs.load_earlier_run().

Earlier-format configurations¶

p2p.config_conversion.convert_earlier_config() converts regression, pipeline, ablation, panel and simulation YAML written in the earlier configuration format into a central manifest. It replaces evaluated NumPy expressions with structured grids and separates persona-generation settings (batching, attribute caps) from backend arguments. Review the result before a paid or large local run.

Cleaning policies¶

Every run records how incomplete proxy answers were handled. P2P fits with column_drop_v2, which drops proxies that did not answer every question. Earlier-format runs may carry row_drop_v1, which instead dropped any question row with an incomplete proxy answer. Aggregation refuses to pool runs with different cleaning policies.

Earlier option spellings¶

Two option values have behaviour-named spellings; the earlier ones are still accepted and are recorded exactly as written, so existing manifests run unchanged and keep their configuration digests:

Option	Name	Earlier spelling (accepted)
`evaluation.distribution_prediction_policy`	`signed_sum_normalize`	`legacy_signed_normalize`
`evaluation.distribution_route`	`answer_file`	`legacy_answer_file`

max_iter and post_selection_refit in a model block are also accepted, have no effect, and are recorded as ignored_model_options.

Target renormalisation¶

data.target.renormalize: true divides each target vector by its own sum (first by 100 if the sum exceeds 1.5) before scoring, so a target whose shares do not quite sum to one can still be scored; an all-zero target is recorded as degenerate instead of raising. The fit sees the target block as given and only the scoring sees it rescaled. It is off by default: without it, a target that is not a distribution is refused.