# Website download data dictionary

Author-preview exports from the frozen v0.6.2 study. These files contain **reported aggregates and selected examples**, not the full 100-task dataset. They do not include raw run logs, model reasoning, or a repository release.

## `results.csv` and `results.json`

The CSV has one row per model and memory condition (10 rows). `protocol` identifies the study version; `model` is the pinned model snapshot; `condition` is one of `no_memory`, `aligned_similar`, `irrelevant`, `misaligned_similar`, or `irrelevant_action_matched`. `critical_correct`, `final_success`, and `first_action_correct` are counts out of `n_unique_tasks=100`. These are distinct endpoints even where counts agree. Repeat trials are excluded.

`results.json` adds study totals, paired-outcome counts for the primary comparison, and exact trajectory-difference counts. The display grid represents counts, not individual task IDs. Source: `reports/v0.6.2/evaluation_analysis.json`; task and repeat totals checked against `manifests/v0.6.2/evaluation_manifest.json`.

## `cases.json`

Eight selected records, with overlap: six author-review selections and two paired comparisons. Each record has an editorial `label`, `summary`, and `decision`; the saved `task`, `policy`, `initial_context`, `memory`, `source_events`, `trajectory`, `final_response`, and `grade` fields were checked against `runs/v0.6.2/evaluation.jsonl`. `model` for the six review records was filled from `reports/v0.6.2/human_review_mapping.json`. For the no-memory case, the exact saved `memory` is an empty string; the page displays a human-readable placeholder. Paired records contain a real `task_id`; the other records deliberately do not expose or invent one.

Within each `arm`, `condition` names the memory condition; `source_events` is the parsed source memory; each `trajectory` item contains a tool name, exact arguments, and actual result; `final_response` is the observed message. `grade` contains saved automated scores, including `critical_correct`, `critical_reached`, `final_success`, `first_action_correct`, and `policy_violating_attempts`. The author’s guided review was a spot-check, not independent blinded validation.

`provenance.json` records the original website-package inputs and this repository’s frozen-artifact verification. The PDF is an author version and has not been submitted or published.