# Offdiagonal > Point-in-time evidence platform: what independent forecasters, models, crowds and markets predicted at the moment of decision versus what happened. Frozen content-addressed artifacts, explicit clocks, no synthetic zero-fill. MCP server (public, read-only): https://offdiagonal.space/mcp Sources: Physical systems and NWP (air, sea, urban, agriculture, earth; originating producers via Open-Meteo), CDC FluSight / COVID / RSV hubs, ECB and Philadelphia Fed surveys, crypto (Coinbase references, Kalshi, Deribit), prediction markets and wallet trades, ForecastBench, Numinous, NESO/Elexon energy, dated Jev evidence, Zeno30 and reference archives. ## Rules - Point-in-time: every row carries clocks (available_at, forecast/outcome times). Never use a field before its available_at. - Missing means not measured: null is never zero-filled, empty hours are omitted, nothing is backdated. - Append-only: corrections are new content-addressed records (supersedes_sha256); sealed artifacts never change. - Field counts: 24 fields are tracked (everything we collect); 13 are admitted for training in the current release. Both numbers are correct; tracked is not admitted. - Fields under archival_only are never model inputs. - Jev judgments are advisory evidence only, made on or after 2026-09-25; not inputs for older forecasts. - Split training and test by split_group. - Public summaries and MCP tools are read-only without a key. Protected decision records, cutoff labels and partner bundles require the named credential scope (check with /api/public/v1/handoff/whoami). - Separate feature, fit, selection and scoring clocks: features existed by decision_at; training labels existed by fit cutoff; development labels existed by selection cutoff; sealed test predictions may be scored with later answers that never entered fitting or selection. - Numinous (as of 2026-10-24): the provider's market listing and Numinous-1 stopped at snapshot 2026-10-21 (miner beliefs still arrive on part of the markets); all 112 resolved questions now carry the Polymarket settlement (112/112 agree); only is_backtest=false is a live forecast, unknown flags count as practice. Collector health and gaps: /api/public/v1/ingestion-health and the defect register. Zeno30 competes on Numinous: its own miners are Zeno's team, never independent forecasters or outside evidence about Zeno; until their miner identities are confirmed they stay unmarked and the reader must exclude them. - Numinous Financial Graph (numinous-graph-asof-v1, context input, never a label or outcome): ~38k company relationships (head, relation, tail, standing, support, observations with source URLs) and their events, captured every 6 h as full paged passes. Read with a partner key carrying export:rows: GET /api/public/v1/numinous-graph?as_of=&after=&limit<=500 (keyset by edge_key; next_after continues). Each edge carries availability {available_at, basis}: the earliest cited article's own publish time when it has a URL and predates our first capture (cited_article_published_at), else our capture (witnessed_capture). Gate every edge to available_at <= decision_at; edges not yet public are absent, never zero. happened_on and graph first_seen are never availability clocks. Rows are never deleted. ## Endpoints (base https://offdiagonal.space) - /offdiagonal-download/LABEL-CLOCK-REPAIR-20261024.md — READ FIRST for historical panel tests: original publication versus archived vintage, live inventory and separate fit/selection/evaluation clocks; old 26-outcome audit is not a universal test ceiling. - /api/public/v1/handoff/panel-label-clocks?slice=panel-ecb-v1 — Live label identity/value/clock reconciliation, not the stale September audit. Add cutoff= for training eligibility. view=labels requires export:rows and cutoff, pages by after=next_cursor; eligible_only=true excludes unverified clocks and known non-identical target definitions. Answers are supervision-only, never decision inputs; archived vintage does not prove original first publication. - /api/public/v1/handoff/decision-records?slice=panel-ecb-v1&limit=10 — decision-record-v2 (default; version=v1 for v1) requires an export:rows partner key; v2 adds per-revision availability clocks, the forecaster's own range inline and whole-slice totals for capture checks: one row per panel member at decision time with availability age (issue clock and forecast age unknown), own distribution reference, peer panel and event-level history (repeated forecasts of one outcome collapsed into one entry with its revision trajectory and signed error); no labels; follow next_cursor via after=. Public view=agents lists Zeno30 agent identities by base-model family (same family never independent). - /api/public/v1/wasde?after_id=0 — USDA WASDE raw text reports: exact response SHA-256, calendar release date and witnessed retrieval clock; historical availability and first-print evidence unverified, no parsed forecasts or outcome join. Follow next_after_id to an empty page. - /api/public/v1/variables — READ FIRST for weather/sea/air: variable catalogue mapping every observation channel to quantity, unit, measurement kind and truth eligibility; pair forecast and truth only on matching quantity+unit; station_pressure_msl is QNH - /offdiagonal-download/ZENO-SUPPORT-20261021.md — Response to 'Why Zeno is still struggling': pressure/unit anomaly checks, measured airport receipt delays, shared-failure register, 75k temperature-market links, action receipts - /api/public/v1/consensus-register?after=0&limit=100 — Keyset prospective shared-failure register: >=3 producer families agreeing tightly, sealed before target with sha256; adjudications vs station truth are never inputs - /offdiagonal-download/VARIABLES-AND-DOMAINS-20261021.md — Variable and topic audit across weather, hurricanes, energy, health and markets; question topics via question-domain-v1, unclassified never guessed - /offdiagonal-download/DATA-EXPANSION-20261021.md — Regional issued-run additions, repaired airport schedules, independent buoy/gauge measurements and severe-weather raw captures; live additions, not a replacement frozen release - /api/public/v1/independent-evidence?after=0&limit=100 — Keyset raw buoy/gauge/SPC/NWS capture receipts, source clocks and hashes; follow nextAfter to complete=true, never treat unmatched raw reports as scored episodes - /api/public/v1/observations?site=ndbc:42001 — Independent station observations: buoy:42001/41004/46001 and usgs-08070000/01646500/07374000; observedAt and first-recorded availableAt kept apart - /offdiagonal-download/DATA-AUDIT-20261020.md — READ FIRST: full data audit; measured per-source outcomes, live defect register with exclusions; second pass at PIPELINE-AUDIT-2-20261020.md; apply /api/public/v1/data-defects before training on any release - /offdiagonal-download/DATA-CENSUS-20261020.md — exact row count of every table, by kind (source, own, derived, frozen copy, operations) - /offdiagonal-download/AGENTS-CONNECT.md — For outside AI forecasting agents: own key, sign-up, validate, submit, read back, scorecard, evidence tips, truth checks and action receipts. Current intake is 1.11.0; omit per-record schema_version, require a genuine occurred_at for receipts; historical 1.9.0 records remain eligible under timing and scoring rules. - /offdiagonal-download/ZENO-REVIEW-REPLY-20261025b.md — CURRENT RULES: clock-evidence-v2 (event time alone gives no label availability and the row is ineligible), stage-report-v2 funnel at /api/public/v1/handoff/field-counts (used counts only with run_id + code_sha256), refusals idempotent by batch_sha256 or one file at /api/public/v1/handoff/refusal-artifact, hydrology held until hydrology-validation-v2 passes (7-day event blocks per river, 5 USGS rivers). Earlier 25 Oct reply: ZENO-REVIEW-REPLY-20261025.md - /offdiagonal-download/ZENO-REVIEW-REPLY-20261025.md — Superseded by the 25b reply: clock-evidence-v1 and the first lane contracts (/api/public/v1/handoff/lane-contracts) and row reconciliation (/api/public/v1/handoff/reconciliation) - /offdiagonal-download/HANDOFF-v6.22.2-ACK-20261026.md — READ WITH v6.22.1 (HF af652a05, no data changed): answer to Zeno's 26 Oct training report. v6.22 runs stopped/superseded; v6.22.1 intake admitted, not used; Zeno-own forecasters are one dependent family (zeno_own), never peers or inputs; exclude Zeno's Numinous miner by hotkey, not UID 53. - /offdiagonal-download/evidence-contract-v1/README.md — Shared evidence contract v1 for Zeno30 and Zeno Divergent: commit/reveal rounds, assessment-receipt-v2, research actions, family-first scoring; test against examples.json. - /offdiagonal-download/HANDOFF-v6.25-INFO-20261027.md — READ WITH v6.22.1 (HF c496ca78, handoff-v6.25-info-20261027/, no training rows changed): 2,192,195 weather ensemble summaries gated to available_at (702 incomplete, never filled); energy multi-producer listing (no target has two independent families). - /api/public/v1/handoff/notflat-identity-inventory — notflat.world v2 identity inventory (nfw-identity-v1, public, counts + hashes only): per resource of one pinned snapshot — state complete/partial/not_started, received records, pages, terminal, digest failures, sorted record_id sha256 (LF-terminated, bytewise), duplicate ids; plus notflat's own rights fields and the owner's append-only rights declarations. Partial resources never get a hash. - /api/public/v1/handoff/launch-inputs — LAUNCH INPUTS (launch-inputs-v1, public, counts only): per field, newest 1,000 rows judged on Zeno's three campaign inputs: evidenced issue clock (capture time and submission deadlines never count; producer run time older than 48h before capture rejected), distinct lead hours (>=3, lead from issue clock only), and decision-time links (neighbouring site <=800 km, same channel, both sides public at or before issue). pass + reasons per field. - /api/public/v1/handoff/categorical-outcomes — CATEGORICAL HEAD (categorical-outcome-v1, public): resolved Manifold multiple-choice questions, one outcome per question with full option list; winner only when the resolution names exactly one mutually exclusive option; label_available_at = Manifold resolutionTime. Keyset ?after_id=. Play money, own population. - /api/public/v1/handoff/corpus-inventory — v0.90 LEDGER INPUT (corpus-inventory-v1, public, counts only): every v6.35 file with status used/held/evaluation_only/archival_only/metadata, rows, admissible rows, clocks, split rule, reason and unblock date; pin inventory_sha256 in the ledger. full_corpus lists EVERY data file in every HF release (current vs superseded_by) — v6.35/v6.36 are add-ons, the ledger must include inherited packs. - /api/public/v1/handoff/outcome-map — outcome-map-v1: one event id per real outcome across every lane (weather city-day, energy region/target/interval, markets, Zeno30); split ONCE on it after combining old and new packs. Weather confirmation weeks follow the sealed protocol (fifth week freezes 14 Dec). - /api/public/v1/handoff/zeno30-crosscheck — Zeno30 answer cross-check (export:rows): each answered question beside the newest official check (fact_review, official clocks, clock gap in minutes); rows_sha256 pins what was compared. Official checks re-run daily 06:20 UTC and append only on change. fact_review-v2: the newest check that shows a value wins; review_required gives needs_review (no label); annulled questions have no outcome; a correction (supersedes_sha256) replaces the answer it names; the seal sent with a forecast is the only seal. - https://huggingface.co/datasets/mbarbosa1/flatplanet/tree/28bec30c5eeae4544ae8574472f12088456aa3ce/handoff-v6.38.1-20261031-audit — AUDIT ADD-ON (v6.38.1, HF 28bec30c, no row changed): content-hash-v1 fingerprints of every value in v6.37 and v6.38, change report beyond row counts (18 Zeno30 records gained an official answer clock), frozen holdout list HOLDOUT-v1 (never moves; keep holdout out of training, history, calibration, vocabulary and routing). 6/6 files re-verified. - https://huggingface.co/datasets/mbarbosa1/flatplanet/tree/b5f03880e78c2c9bf7f5aeed6e8eb4616e3c789b/handoff-v6.38-20261031 — LATEST (v6.38, HF b5f03880, add-on to v6.37): zeno30-records-v1 re-cut at 2026-10-31T17:00Z (1,704 rows; train only train_eligible=true: 65 forecasts over 6 questions; exposed=true rows are revised, not blind), categorical-outcomes-v1 (1,113 questions, 639 eligible), launch-inputs-v1 field grades (no field passes yet), corpus inventory, AUDIT.json (all checks pass). 7/7 files re-verified. Read README.md first. - https://huggingface.co/datasets/mbarbosa1/flatplanet/tree/3346453a07eb3d9451c2d76a7bcd5df7755f7fd9/handoff-v6.37-20261030 — v6.37, HF 3346453a, add-on to v6.36): zeno30-records-v1 frozen at 2026-10-30T20:00Z (1,101 rows; train only current, unrefused rows with fact_review=accepted: 65 forecasts over 6 questions; cohort per question; family is the independence unit), categorical-outcomes-v1 (1,100 Manifold multiple-choice questions, 629 with one winner; label clock = resolutionTime; never pooled with binary lanes), AUDIT.json (all checks pass). 5/5 files re-verified. Read README.md first. - https://huggingface.co/datasets/mbarbosa1/flatplanet/tree/577c34f3ae43021c1e0583487293ec0a0d820af3/handoff-v6.36-20261029 — PREVIOUS (v6.36, HF 577c34f3, add-on to v6.35): ensemble-daily-max-pairs-v2 (62,280 member forecasts on 126 paired city-days under station-daily-max-v1, re-audited under v2 (unnamed station readings never pair): same 126 days, 0 affected; replaces pairs-v1 for scoring), station-daily-max-days-v1 (all 180 city-days with status; 41 incomplete, 13 mixed_station have no label), append-only Metaculus official checks (not labels), corpus inventory. 6/6 files re-verified. Read README.md first. - https://huggingface.co/datasets/mbarbosa1/flatplanet/tree/24636e4d995223e20e7651bb3440ae7c3087204e/handoff-v6.35-zeno-fixes-20261029 — OLDER (v6.35, HF 24636e4d, add-on to v6.34): Zeno30 with per-question cohort, fact_review and official clocks; 857,536 energy forecast-outcome pairs (325,058 prospective); ensemble daily-max pairs with pair_status; weather quantity crosswalk (domain|site_id|channel|unit); Jev split into active target_match and deprecated streams; hydrology-validation-v2 receipt (passes, provisional gauges); notflat.world archival_held; empty-by-evidence Manifold bridge. 16/16 files re-verified. Read README.md first. - https://huggingface.co/datasets/mbarbosa1/flatplanet/tree/05c4270197aef021a229b92d38ebf4f60fd3200b/handoff-v6.34-20261029 — (v6.34, HF 05c42701, as_of 2026-10-29T01:30Z, no v6.22.1 row changed): Zeno30 lane (1,021 records, no minimum answer count), Jev classifications with timing lanes (185,418; outcome-bound ones all retrospective) and replay (330 clean of 4,000), 98,379 Manifold answers, fresh weather/energy/market/Numinous/crypto deltas, full per-domain inventory, training instructions and audit; 22/22 files re-verified by sha256. Read HANDOFF-v6.34-INSTRUCTIONS.md first. - https://huggingface.co/datasets/mbarbosa1/flatplanet/tree/95f7da56b590e57dc1496b016cb6b110ebedd9e9/handoff-v6.33-structure-20261028 — conviction-influence-v1 (HF 95f7da56, no v6.22.1 row changed): per-bet timing, pool share, personal conviction, price distance and own exposure from strictly earlier seconds; price impact and time-before-answer archival_only; world/information influence unknown until evidenced; one global timeline gate; train-week-only weather scales; strict manifest checker (40-hex pins only). - https://huggingface.co/datasets/mbarbosa1/flatplanet/tree/be9c2fb661b90cc2b056cbdcc7cd8537a1bfbbe1/handoff-v6.32-belief-20261027 — belief-record-v1 (HF be9c2fb6, no data changed): sealed belief fields per Zeno30 forecast (expected other-family probabilities, dated sources, countercase, change-my-mind observation) checked at /api/public/v1/zeno30/belief-check; exposure log at /zeno30/exposure; countercase_probe vs evidence_refresh arm randomised by event; action records at /api/public/v1/action-records/submit never labelled averted or false alarm. Countercases are hypotheses, never labels. - https://huggingface.co/datasets/mbarbosa1/flatplanet/tree/fd36ea3fe23dd3c385db311d729f208246b882f6/handoff-v6.31-fixes-20261027 — READ WITH v6.22.1 and v6.30 (HF fd36ea3f, no training rows changed): week-block-v2 scorer (drawn week copies tagged; <5 weeks = insufficient_evidence), venue lanes (Manifold separate from Polymarket), row-level answer clocks for 13,151 Polymarket questions and 82,667 weather outcomes, sealed-week protocol from 2 Nov, Zeno SN6 miner by hotkey not UID 53. - https://huggingface.co/datasets/mbarbosa1/flatplanet/tree/ea968917a615a68db74562ea8606441e0b737945/handoff-v6.30-weeks-20261027 — READ WITH v6.22.1 and v6.26 (HF ea968917, no training rows changed): week-block-v1 — event_week + outcome_group on every weather outcome row, frozen whole-week partitions (82,667 matched outcomes over 8 weeks; confirmation is one week), Polymarket dev_selection/dev_confirmation split, overlap checker and declared outcome-weighted scorer (day weighting is sensitivity only), 98 frozen content-addressed quantity IDs, consumption ledger. - https://huggingface.co/datasets/mbarbosa1/flatplanet/tree/d3b5bab980bcd5aa1ddcb19c631ed4e52df17d9c/handoff-v6.29-audit-20261027 — READ WITH v6.22.1 (HF d3b5bab9, no data changed): decision-clock audit (every shared row has its real decision time, nothing later used) and per-source roles (supervised / input / diagnostic / held). Weather outcomes = 8 event weeks, not 1.14M outcomes. - /offdiagonal-download/HANDOFF-v6.28-INFO-20261027.md — READ WITH v6.22.1 (HF 4bf3d04d, handoff-v6.28-info-20261027/, no training rows changed): temporal-correction-v1 rival (only earlier-public outcomes), 59,438 Kalshi nested-threshold links (outcome relations, not error correlation), Zeno30 participation, provenance/research-queue/decision-context checks, capability sheet. Say: additional value beyond simple calibration has not yet been demonstrated. - /offdiagonal-download/HANDOFF-v6.27-INFO-20261027.md — READ WITH v6.22.1 (HF 509bada9, handoff-v6.27-info-20261027/, no training rows changed): matched calibrated comparison (pair by case + decision time, seeds then outcomes; tie-pooled isotonic), CPU recovery recipe, shadow reference_only until verified outcomes, Grok perspectives one family, count vs result claim levels, v6.26 weather = 8 event weeks. - /offdiagonal-download/HANDOFF-v6.26-INFO-20261027.md — READ WITH v6.22.1 (HF 150d9bfa, handoff-v6.26-info-20261027/, no training rows changed): outcomes attached to 2,198,394 weather ensemble summaries (1,138,371 matched, 664,053 unmatched, 395,900 refused for unrecorded unit, 70 incomplete); assessment-receipt-v3 with evaluated decision choices; Zeno30 shadow workflow. - /offdiagonal-download/HANDOFF-v6.24-INFO-20261027.md — READ WITH v6.22.1 (HF bb5c412b, handoff-v6.24-info-20261027/, no training rows changed): source scorecard (freshness separate from seven equal preparation checks), ECB SPF past-only indicators gated to 2026-04-29 (431 rows, 15,374 pair dependence rows), weather error diagnostics (never inputs), weather panel behaviour per case. - /offdiagonal-download/HANDOFF-v6.23-INFO-20261027.md — READ WITH v6.22.1 (HF 3ee2f489, handoff-v6.23-info-20261027/, no training rows changed): 12,145 target cards with locations, 21,506 native forecast shapes (695 complete), 358 producer provenance records, history census, 304,892 market quote-coverage series, research-action schema, and 11 fixes for Zeno's trainer and judging. - /offdiagonal-download/HANDOFF-v6.22.1-ADDENDUM-20261026.md — READ WITH v6.22.1 (HF b2839db6, no data changed): usable check rows now also require no case refusal and a check before our first sighting of the answer; measured 0 of 83,199 rows affected. Reply to the 26 Oct review: reader settled-history clock, saved wallet predictions vs baseline_brier, range block dates, Numinous pin, exclusion for rival paths, no oracle rival. - /offdiagonal-download/HANDOFF-v6.22.1-NOTES-20261026.md — LATEST (v6.22.1, HF revision 46d136eb03647138afb0e4c0fbaea7308e9a4497, folder handoff-v6.22.1-20261026/): polymarket-trader-v4 supersedes v6.22 wallet files. prev_tape_yes now strictly earlier second only (424,500 v6.22 bets used a same-second price); wallet_settled_before counts only questions the wallet traded eligibly and saw answered before the bet (204,959 were overstated). Panels add tape_yes_at_check and baseline_brier (latest-trade rival at the same check time, mean 0.1106). Future-trades test passes; combined overlap audit: 6 multi-source questions, 0 split conflicts, 100 answers public after their split ends. Post-deadline questions are historical evaluation, not a sealed test. - /offdiagonal-download/HANDOFF-v6.22-NOTES-20261026.md — SUPERSEDED (v6.22, HF revision 3bd0b2ec49ca5c8c5bec7a5364724835bf81e0cc, folder handoff-v6.22-20261026/): wallet lane under polymarket-trader-v3 (v2 rules unchanged): 11,147,828 directional trades, 674,741 eligible bets from 79,393 wallets over 13,151 questions, 83,199 usable check rows. New pre-bet columns: prev_tape_yes (market comparator, 661,607 bets), wallet_settled_before (member ranking), forecastbench_overlap (8,490 bets; one question one split; list in fb-polymarket-overlap.csv). venue_yes_before is empty by design (hourly venue recording began 2026-08-18, after these bets). Test split remains date-held-out, not untouched: final judging must be prospective. - /offdiagonal-download/HANDOFF-v6.21-NOTES-20261026.md — PREVIOUS (v6.21, HF revision 1601d283067a373de9616cefc825ccfa27fbb8c0): wallet lane rebuilt under polymarket-trader-v2 from 10,616,400 directional trades: 671,683 eligible bets from 78,998 wallets over 13,150 questions; 83,178 usable check-panel rows; date split by deadline (exact thresholds 2026-04-29 / 2026-06-19 UTC; use the crosswalk). v6.21.1 amendment (HF 13b44a308f94f7e77e90078642808aaf118711ed): exposure ledger shows the test split is date-held-out, NOT untouched (1,995/1,996 test questions shared before, 948 were v6.19 validation) so final judging must be prospective; 41 chronology-purge questions; the 3,064-row panel drop is traced to hedges/market makers v6.20 missed. 8,700,169 trades on markets that hit the venue cap are refused whole. Minute prices stay at v6.20; energy and Numinous graph at v6.17. - /offdiagonal-download/HANDOFF-v6.20-NOTES-20261026.md — (v6.20, HF revision 636d471052dcacc9efbecccfc67f3013a16f46b9): wallet panels now carry the last minute-level Polymarket price at or before each check (at most 12h old): 29,953 panel rows with a market price, up from 3,570. 26,003 price windows fetched, raw responses hash-bound. 11,994 of 12,975 questions have no venue price in their final 48h and stay empty. Episodes and split crosswalk stay at v6.19. - /offdiagonal-download/HANDOFF-v6.19-NOTES-20261025.md — PREVIOUS (v6.19, HF revision 22abf5e0c26934d5f43529caa3040ba565af79c0): Polymarket wallet lane v2 — flags use only earlier trades, 594,529 trades eligible under Offdiagonal rules (73,933 wallets, 13,141 questions); settlement times labelled provider_reported or witnessed_capture; trades are bounds (yes_at_least/yes_at_most), never stated probabilities; decision-time wallet panels with market price at check; split crosswalk with a date split (use it). Supersedes v6.18 trader episodes. - /offdiagonal-download/HANDOFF-v6.18-NOTES-20261025.md — PREVIOUS (v6.18, HF revision e358a74665c31cdcba4930c3833612793319f9d4): Polymarket traders as forecasters — 542,809 eligible wallet trades from 66,870 wallets over 13,137 settled questions; labels dated by on-chain settlement; independence between wallets unknown; hedges and market makers refused. All v6.13–v6.17 files still current. - /offdiagonal-download/HANDOFF-v6.17-NOTES-20261025.md — PREVIOUS (v6.17, HF revision 5b0ec02d6d7222e1810614cd7fac7ae3a19d8da8): energy pairs refreshed (39,716 eligible NESO over 716 outcomes; 25,881 eligible CAISO over 4,140 outcomes) and Numinous Financial Graph context (38,084 edges, 8,663 events, gated to available_at; never labels). Supersedes v6.16 energy files. - /offdiagonal-download/HANDOFF-v6.16-NOTES-20261025.md — PREVIOUS (v6.16, HF revision 679659855fd729b5eee7332afc84584572f83896): paired energy episodes — 25,838 eligible NESO forecasts over 479 outcomes, 25,161 eligible CAISO forecasts over 4,068 outcomes (per-hour witnessed capture clock) - /offdiagonal-download/HANDOFF-v6.15-NOTES-20261025.md — PREVIOUS (v6.15, HF revision 03345819bf24fa2260fa90b800b1b051bb6f67e4, top-up): rows recorded since 24 Oct with original clocks — 49,940 Numinous live forecasts, 24,634 Numinous prices, 127,143 ForecastBench rows, 13,256 NESO forecast points, 43,709 energy outcomes; held/ folder (CAISO without publish clock, river flow, new buoys, Kalshi/MISO raw) is never model input - /offdiagonal-download/HANDOFF-v6.14-NOTES-20261025.md — PREVIOUS (v6.14, HF revision a289edff474495ae572b7bd5961a373cc1b0db97, inherits v6.13 data files unchanged): refusal receipt #1 (7,271,876 rows, 13 fields), defect register, stage funnel, source completeness and lane contracts at freeze - /offdiagonal-download/HANDOFF-v6.13-NOTES-20261024.md — DATA FILES (v6.13, HF revision 3c16e8fde1aca3b12a9d8fe45cbf1898784ab63f, parent 8c5d3ad1): per-case audited ECB answer dates (v1/v2), Numinous+Polymarket episodes with 112/112 settlements, ForecastBench join keys with 71 conflicting values listed, energy episodes (15,554) and river flood runs - /api/public/v1/source-completeness — per source: newest forecast issue and outcome publication clocks (the source's own), cadence, last clean pass, backlog, open dated gaps, whether the source itself stopped, and which pinned artifact holds it - /api/public/v1/repair-ledger — append-only chain per defect: repair, then the pinned artifact carrying it, then Zeno's reported consumption (empty until Zeno reports; fixed never implies trained on) - /offdiagonal-download/AUDIT-3-20261024.md — THIRD AUDIT (24 Oct): every table counted (exact under 2M rows), collector states, live endpoint probes, source freshness, open items with blockers; read before training - /offdiagonal-download/HANDOFF-v6.13-NOTES-20261024.md#older-revisions — OLDER REVISIONS: every HF revision quoted below for v6.12 and earlier was removed when the dataset history was squashed on 23 Oct; those IDs now return 404. The same files survive unchanged in their release folders (for example handoff-v6.12-20261022/) at revision 8c5d3ad1972fff013a7ec5ffa2fbeb0e9da308ca, which is also the parent of v6.13. Pin 22abf5e0 (v6.19), e358a746 (v6.18), 5b0ec02d (v6.17), 67965985 (v6.16), 03345819 (v6.15), a289edff (v6.14), 3c16e8fd or 8c5d3ad1 only. - /offdiagonal-download/HANDOFF-v6.12-NOTES-20261022.md — v6.12 (HF revision eb468b89eb63ce65c30d0778e0b9a922c8aa6016): temperature-market events x 6 ensembles with the last Polymarket CLOB Yes price before each deadline (18,466 of 19,030 priced; rest empty, never zero-filled) - /offdiagonal-download/HANDOFF-v6.11-NOTES-20261022.md — v6.11 (HF revision f7be419e0e6bd454f91ad50edef997bee0916feb): complete Polymarket prices for all 19,030 verified temperature markets (19,019 covered, 11 with no Polymarket history, none zero-filled) - /offdiagonal-download/HANDOFF-v6.10-NOTES-20261021.md — v6.10, HF revision 29c69283c562eb23af88d752881bc555d1592c69): ForecastBench per-question resolution clocks from git history (use first_resolved_commit_at as label_available_at; join includes direction), as-of track-record features for wind, surface pressure and urban (null when unknown), refreshed temperature-market prices - /offdiagonal-download/NOTFLAT-FORECAST-MODELS-20261021.md — v6.9 (HF revision ce99325a4f9b0493e20e72df0f223810bce0c476): notflat.world urban forecasts come from UrbanWatch statistical models (LightGBM v1 + climatology baseline) on apparent temperature, not a weather model; lineage unknown so independence is null. v6.9 also adds Polymarket price history for verified temperature markets (partial at freeze; live coverage at /api/public/v1/temp-market-prices/coverage) - /offdiagonal-download/HANDOFF-v6.7-CORRECTION-20261021.md — v6.8: corrected temperature-market events — bins re-parsed from each market's own question text (temp-market-link-v4); the v6.7 'no pre-deadline prices' statement is withdrawn (Polymarket CLOB prices-history has them; our market_price_yes covers only our captured wallet trades). Frozen at HF revision 3d9c60ca14f77e1fb119b1c44778cbaffb8ea04c; use the v6.8 temperature-market-events.csv.gz, not v6.7's - /offdiagonal-download/HANDOFF-v6.7.md — answers Zeno's 21 Oct audit — full-denominator warning ledger (/api/public/v1/warning-ledger), temperature-market exact events (see the v6.8 correction above before using), notflat.world weather subset, Zeno30 blind-vs-revised pairs (/api/public/v1/zeno30/paired), forecaster report card (/api/public/v1/forecaster-scores) - /offdiagonal-download/HANDOFF-v6.6.md — Previous: additions since v6.5 — checked Polymarket temperature links (only verified_station_day pairs), 2.96M airport readings with full wind/dew/visibility/rain, variable catalogue, NOAA GFS single runs, shared-failure seals, retired FluSight 2021-23 lane; buoy/gauge forecast-truth pairs at /api/public/v1/independent-pairs - /api/public/v1/independent-pairs?entity=ndbc:42001 — Buoy/river-gauge forecasts paired with real readings (independent-pairs-v1); never inputs - /api/public/v1/storm-outlooks?days=3 — SPC Day-1 outlooks scored against severe NWS storm reports (storm-outlook-v1); reports are incomplete truth - /offdiagonal-download/HANDOFF-v6.5.md — Previous: everything Zeno v0.80 listed as waiting on Offdiagonal, in one frozen release: world conditions, shared-truth table, herding market flag, full quantiles and Philadelphia Fed bins, blind-spot counts, Jev judgments with timing lanes, counterworld-lab-v1, air/urban/agriculture forecasts with outcomes - /offdiagonal-download/HANDOFF-v6.4.md — Counterworld Lab addendum: Zeno30 1.7.0 counterworld and probe records (sealed before outcome) and the blind-spot label rule; read with v6.3 - /offdiagonal-download/HANDOFF-v6.3.md — History addendum: hurricanes (NHC ATCF forecast decks, 301 methods in 16 documented groups, best-track truth) and the ranked list of sources coming next; read with v6.2 - /offdiagonal-download/HANDOFF-v6.2.md — Master guide for Zeno's next training run: situations v1.3 (HF c77975e) fixes the split to usable cases with outcomes and adds airport station checks from 1 June (28,623 of 97,196 cases); v6.1 still covers the registry, lineage, readiness and calibration - /api/public/v1/situations-readiness — per-domain, per-capability readiness verdicts, typed source registry and forecaster lineage for situations - /offdiagonal-download/HANDOFF-v6.0.md — History: situations contract v1.1 - /offdiagonal-download/HANDOFF-v5.3.md — History: situation pilot (contract v1.0, superseded) - /offdiagonal-download/SITUATIONS-v1-ADMISSION.md — per-pilot admission report: outcomes, families, context and truth coverage, proposed periods, testable questions - /offdiagonal-download/HANDOFF-v5.2.md — changes since v5.1: defect exclusions, real collector gaps, exact recount 20 Oct; read with v5.1 - /offdiagonal-download/HANDOFF-v5.1.md — complete all-source instructions; additive episodes-v3-20261020 (HF 2a0e218) plus exact pinned parents; 95 energy outcomes and 7,330 open questions - /offdiagonal-download/EPISODES-v3-20261020-PIN.json — exact immutable release revision and manifest; 14 new artifacts anonymously verified - /api/public/v1/handoff/energy — NESO national demand forecasts with INDO provisional outturn vintages, both collected hourly; fixed scoring_at, cursor paging to explicit terminal; one demand family - /api/public/v1/handoff/weather-impact?forecast_id=&scoring_at= — decision-visible local London weather context for energy; not national weather or a second demand family - /api/public/v1/handoff/open-questions — daily sealed open-question freezes (staged 00:00 UTC, sealed within the hour); date/after_id paging; answers checked every 10 minutes and kept separate; open questions are not resolved training examples - /api/public/v1/handoff/asof-panels — decision-time panels; next_after paging; absent_in_page_roster is page-local, not a complete withdrawal history - /api/public/v1/handoff/lineage — model/family/organisation/platform identity evidence (partner key, scope export:rows); unknown lineage is not proof of independence - /api/public/v1/handoff/outcome-quality — outcome status and revisions; next_after paging - /api/public/v1/handoff/market-state — question_uid and decision_at required; venue-stored prices and cross-venue twins with explicit missingness - /api/public/v1/handoff/lane-roles — every source's allowed use; all held data is cleared for training, only leakage/no-outcome/Zeno test set excluded - /offdiagonal-download/ZENO-START-HERE.md — start here: verified training inventory by domain, counts, status and known limits - /api/public/v1/handoff — entry points, scopes, how sealed packages are pulled and verified - /api/public/v1/handoff/catalogue?format=summary — every feed, report and package with measured counts - /api/public/v1/evidence-feed?lane=&limit=&after_id= — keyset-paged lanes; start after_id=0, follow x-next-after-id until x-terminal: true - /api/public/v1/market/census — prediction-market census: scoreable, unmatched, declared limits - /api/public/v1/handoff/source-depth — historical October 1 depth report; inspect measured_at, not current counts - /api/public/v1/handoff/source-usability — historical October 1 usability report; old permission notes superseded by owner clearance - /api/public/v1/evidence-feed?lane=ensemble-forecasts — weather ensembles with their own 5–95% member range and outcome - /api/public/v1/evidence-feed?lane=episodes&lead_hours=24 — one row per outcome at a decision time (&source=ensemble-forecasts, a health hub key, ecb_spf, forecastbench, crypto, markets, sea, air, urban or agriculture): own-range panel, per-forecaster history, dated pair dependence counted on distinct outcomes (matched_outcomes; miss_rule names own_interval_90 vs error_above_own_median); outcome under supervision - /api/public/v1/evidence-feed?lane=forecaster-history — per forecaster per episode: ordered past errors/coverage known at decision time - /api/public/v1/evidence-feed?lane=hide-one — panels of >=4 with one forecaster held out (held-out forecast only under supervision) - /api/public/v1/evidence-feed?lane=ecb-confident-failures — ECB SPF respondents' own bins: confident-and-wrong labels - /api/public/v1/handoff/comiss-graph — which forecasters miss together; effective independent crowd size - /api/public/v1/handoff/jev-cutoff — 4,000 replay judgments split at Jev's release date for the cutoff test - /api/public/v1/handoff/jev-provenance — when each Jev judgment was made, fingerprint, version - /api/public/v1/handoff/validation-bundle — validation bundle - /api/public/v1/handoff/contract-check — reader-agreement check: real row ids per warning-label lane with expected source, strict label, eligibility and split group — run your reader on them and compare - /api/public/v1/corpus/census — corpus readiness, frozen contracts, split manifest, leakage checks - /api/public/v1/source-rights — licences and reuse terms per source - /api/public/v1/zeno30 — Zeno30 submission contract (submit: POST /api/public/v1/zeno30/submit) - /api/public/v1/zeno30/validate — POST records to check them; nothing is stored - /api/public/v1/handoff/fragile-agreement — Proposed fragile-agreement protocol and upstream map; Zeno30 1.11.0 accepts intervention_run, observation_request, pre_post_pair, information_gap, gap_fulfilment via existing authenticated submit. All require never_input:true; blind server receipt must precede advice. Benchmark awaits Zeno confirmation; no experiment results claimed. See ZENO30-SUBMIT.md for examples. - /api/public/v1/zeno30/scorecard — Per-agent track record: forecasts before the outcome, Brier when scored - /api/public/v1/agents/register — POST to request your own agent key (pending until the owner approves) - /api/public/v1/hurricanes — NHC hurricane forecast decks: counts and latest rounds - /api/public/v1/assessment-receipts — append-only assessment receipts (submit: POST .../assessment-receipts/submit) - /api/public/v1 — full endpoint index ## Evidence-feed lanes - pretrain-events — point-in-time forecast events across physical, health, market, crypto sources - pretrain-labels — one row per forecast later scored against its outcome; split by split_group - co-forecast-links — same-event links between forecasters - failure-signals — per-forecast failure inputs: references and decision-time context, gated to available_at - confident-failures — labels only: confident forecasts that missed - health-confident-failures — flu/COVID/RSV confident-and-wrong labels (source_key= or source=hubverse_flusight|hubverse_cdc_covid19|hubverse_cdc_rsv; unknown values return 400; train on strict_label where eligible; split on split_group) - health-join — health forecasts joined to outcomes with forecaster identity - health-cases — health outcome cases - forecastbench — ForecastBench forecasts with named forecaster identities (CC BY-SA 4.0); counts vary by pinned release - crypto-observations — hourly crypto prices (asset=BTC|ETH|SOL|CELO); known only after the hour closes - crypto-forecasts — crypto forecasts - weather-clocks — NWP issuance and availability clocks - jev-training — Jev judgments, accepted by default; include=all for every status - jev-scores — Jev scores incl. 4,000 replay judgments - jev-keys — Jev join keys - shared-failure — shared-failure context cases - training-cases — assembled training cases