# HANDOFF v6.13 — 24 October 2026

Release `handoff-v6.13-20261024`, frozen at 2026-10-24T17:00:00Z. It adds files on top of v6.12. Every earlier file stays authoritative at its own revision, and HF revision `8c5d3ad1` is unchanged.

This release exists so that each repair made on 24 Oct is carried by a pinned file Zeno can admit. A repair counts as consumed only once Zeno reports it in the repair ledger (`/api/public/v1/repair-ledger`).

## Files

| File | What it is | Count |
|---|---|---|
| `ecb-labels-panel-ecb-v1.ndjson.gz` | ECB SPF labels with the live answer-date audit for each case (target, vintage, value and clock matched exactly), `supervision_only`, `never_input` | 1,110 cases / 81 outcomes; 366 cases (27 outcomes) clock-verified by the freeze time |
| `ecb-labels-panel-ecb-v2.ndjson.gz` | The same for v2 | 1,216 cases / 91 outcomes; 366 / 27 verified |
| `ecb-labels-census.json` | Totals per slice at the freeze cutoff | |
| `numinous-polymarket-episodes-v2.ndjson.gz` | Numinous miners plus the Polymarket crowd at 24h/6h/1h before resolution. All 112 resolved questions are linked. | 283 admitted episodes / 95 distinct outcomes |
| `numinous-polymarket-census.json` | Drop reasons: 40 no panel, 13 single family, 0 conflicts, 0 late labels | |
| `forecastbench-join-keys.csv.gz` | One join key per resolution: `(question_id, source, direction, resolution_date)`, with every value seen across the daily files | 115,521 keys; 71 have conflicting values (fred 57, acled 6, dbnomics 6, yfinance 2) |
| `forecastbench-join-keys-census.json` | Counts by source | |
| `energy-episodes-v1.ndjson.gz` | NESO national demand forecasts with INDO outcomes, frozen from the live feed at `scoring_at=2026-10-24T17:00:00Z` | 15,554 forecasts; 13,490 with an outcome; 321 distinct outcomes; one family |
| `energy-episodes-census.json` | Pages, scoring time, envelope | |
| `hydrology-flood-runs.csv.gz` | GloFAS river discharge runs, as stored | 600 rows, 2 runs (9 and 19 Oct) |
| `hydrology-flood-capture-status.csv` | Status per daily capture, including the 20–24 Oct days our own write refused | |

## Rules

- **ECB.** Labels are loss targets only. A case is admissible for a fit only when `label_audit.training_eligible_at_cutoff` holds at that fit's own cutoff. The freeze time used here is not a training cutoff. Fixed-EA20 growth, unemployment and core-inflation cases have no identical-target archived vintage and stay unverified. Do not swap in the moving-concept series.
- **Numinous.** `resolved_at` comes from the Polymarket payload. `label_available_at` is when we recorded the answer: our capture time for the 62 answers recovered on 24 Oct. Use only miner rows where `is_backtest` is explicitly false, and treat an unknown flag as practice. All miners are one family.
- **ForecastBench.** Join on the full four-part key, direction included. For the 71 conflicting keys, don't pick a value. Take the first-resolved value and clock from `forecastbench-resolution-clocks` (v6.10) and treat later values as revisions.
- **Energy.** One independent family (NESO), so it supports no cross-family failure claim. INDO outturns are provisional operator figures. The decision is the provider's publication time. Outcomes are supervision only.
- **Hydrology.** There's no stored outcome stream yet, so it isn't trainable. `run_issued_at` is the capture-day key, not a witnessed run time (basis `retrieval_time_run_not_selected`). The 20–23 Oct runs were lost to our own write error and remain a dated gap (defect `hydrology-flood-write-refused`).

## What Zeno reports back

For each file it admits, Zeno reports `training_admitted`, `consumed_by_run` and `in_loss`. Report member rows, panels and distinct outcomes separately at every stage: admitted, encoded, split, sampled, in loss, evaluated.

## Older revisions
Hugging Face revisions quoted for v6.12 and earlier (eb468b89, f7be419e, 29c69283, ce99325a, 3d9c60ca, c77975e and others) were removed when the dataset history was squashed on 23 Oct. They now return 404. Their files are unchanged in their release folders (for example `handoff-v6.12-20261022/`) at revision `8c5d3ad1972fff013a7ec5ffa2fbeb0e9da308ca`, the parent of v6.13. Pin `8c5d3ad1` or `3c16e8fd` only.