# Aggregate data dictionary

Status: Local draft — Jev prerequisite addendum; independent page review and browser QA pending; not published. These are sanitized local candidate downloads, not publication-cleared data.

## Files and schema

- `aggregates.json`: `research-release-aggregate-v1`. Only explicitly selected table cells, aggregate inventory, historical family metrics and source digests are distributed. No per-case rows or provider packets.
- `aggregates.csv`: a lossless long-form scalar projection of that JSON, not a case dataset. `json_pointer` uses RFC 6901 escaping (`~1` for slash, `~0` for tilde); array indices are zero-based. `json_value` is a JSON-encoded scalar; parse it as JSON to preserve strings, numbers, booleans and null. Every scalar appears once. CSV quotes follow standard CSV escaping.
- `tables`: 15 accepted scientific tables with `headers` and `rows`. Display percentages are rounded to one decimal, AUC/Brier to three decimals, repeat spreads to seven decimals. Fractions are displayed numerator/assigned denominator, except representation counts explicitly use complete groups. Display rounding is not extra measurement precision. Tables preserve accepted values and provide evidence-status notes; not every field exists in the full-suite export.
- `table_evidence_status`: per-table provenance scope and missing export endpoint disclosures. Supplemental values remain bounded accepted-source evidence, not independently replay-verified endpoints in this integration.
- `inventory`: historical task instances, valid/invalid counts, eight exported family totals, historical reserves and excluded non-task smoke attempts. Domain controls plus paired mini-fixtures are one exported family but two display cohorts. These are dependent trials across services, presentations and repeats, not independent problems.
- `historical_jev.families`: `available_case_matched` versus `unavailable_matched_cases`; unavailable metrics are null, never zero. Available aggregates contain assigned/valid/invalid, correct, accuracy_assigned, confusion, auc_valid_only and brier_valid_only. Sufficiency adds determined/ambiguous assigned counts and joint_sufficient_correct. News calibration and test are separate.
- `historical_jev.news_policy`: qualifies=false; eligible_for_autonomous_routing=false; routing_metrics=null. No hypothetical threshold-1 policy or time savings are distributed.
- `historical_laya`: summary_only, case_parity=false; per-task n and accuracy. Not matched longer-task data for current arms.
- `new_jev_1_14`: completed reading and one_step only; other family metrics are null/unavailable.
- `source_sha256`: hashes of private aggregate source documents, not public links and not an independent validation verdict. Public source/protocol citations remain pending.

## Metric semantics

`assigned` retains terminal invalids. `valid` counts outcomes usable under the frozen typed contract; invalid outcomes receive no invented probability. Correctness fractions retain assigned denominators. Valid-only ranking and probability metrics exclude invalids. AUC gives half credit to ties; undefined one-class AUC is null. Brier is mean squared probability error. Confusion TP/FP/TN/FN uses p >= .5 and the corresponding binary reference.

Sufficiency positives mean all allowed hidden-state completions determine the same target bit; joint success requires recognized sufficiency and the correct bit. News positives are historical editorial/LLM proxies, not adjudicated ground truth. Routing disagreement uses autonomous decisions as denominator; important misses use valid proxy-positive rows (79 for the all-valid current test split). Reported confidence, chosen-option probability and proposition probability are distinct.

Representation statistics use complete four-member groups, with no imputed base for invalid groups. Coverage reverse wording uses 1-p, initial cells exclude later repeats, and repeat spreads summarize predeclared identical-wire triplets. A forced menu with no correct answer has undefined semantic accuracy, not zero. Accounting values are a historical snapshot: known billing is incomplete and conservative cap charge is not the provider bill.

## Public/private boundary

Only this public directory is a candidate distribution. Do not serve the parent directory: its evidence ledger and pin manifest contain internal audit addresses. No raw news, actor requests, provider responses, case keys, private ledgers or credential material are included here. Sanitization does not grant redistribution rights or publication approval. The public download alone cannot reproduce raw-response replay.

- `addendum_tables`: named Jev prerequisite, historical overlap, usage and allowance display tables; percentages rounded to four decimal places.
- `new_jev_1_14.aggregates`: independently audited aggregate-only values; probability diagnostics are valid-only, paired subsets state their denominators. Usage-known sums exclude missing usage and the earlier compatibility probe. Billing null means unknown; allowance is not actual spend. Phase timestamps/timers do not establish controlled latency.
- `review_status`: accepted historical offline scope is separate from pending review/browser QA of this successor.

- Practical successor: measured fields and all tables retain predecessor semantics. Only release schema/status/review metadata change. Proposed deployment targets appear in explanatory prose, not as measurements or new frozen gates. No trained-model results are added.