# F9bomber external ledger -> VTSR-C (decision memo) Status: **RATIFIED + SHIPPED** (2026-09-17). Operator sign-offs recorded 2026-09-13 (aliases, exclusions, gates, full K, headline-record inclusion, match-level overlap semantics) in the planning thread; implementation landed with `CMDR_ELO_SCHEMA_VERSION 2 -> 3`. ## Decision F9bomber's hand-kept match ledger (`f9stats/f9stats-20260913.xlsx`, sha256 `9e3b5812db200407...`) is imported ONCE by the standalone `scripts/import_f9_ledger.py` into two committed artifacts (`data/external/f9_ledger.json` + `f9_community.json`), and the ledger's eligible duels enter the **published VTSR-C ladder** as first-class externally-sourced duels. Community rollups (thug team records, faction picks, map counts) ship as display-only surfaces. Nothing here touches VTSR-T, the All Matches aggregator, `map_stats`, or picker-scoped `faction_stats`. ## Why the outcomes are trusted - The operator declared F9's outcomes **definitive** (active player, hand-logged at the table). - Empirical check: every F9 row pairable to one of our DETERMINED telemetry matches (Steam64 commander pair + map + date +-1 day + roster Jaccard) agreed on the winner — **28/28, zero disagreements**. The single "disagreement" in an earlier analysis was a rematch-twin mispairing artifact, not a data error. ## Ratified parameters (frozen; changing any of these needs a new memo) | Parameter | Value | Rationale | |---|---|---| | Eligibility | roster-complete, even 3v3/4v4/5v5, straggler-clean, winner is a commander, duration >= 240s, deduped | Operator spec. Even teams remove the man-advantage confound — the one confounder we can kill for free in a no-telemetry dataset. Stricter than our own 6-player telemetry gate on purpose. | | Overlap policy | match-level: ours supersedes. Import-time pairing (date +-1 day for US-local vs UTC skew, map stem OR alias-resolved registry key, Steam64 commander pair, greedy Jaccard >= 0.4) + runtime guard in `compute_commander_elo` (consumed-ids exempt) | Dual-recorded games must never double-count; corpus-era games we simply failed to record stay in. | | K | `k_factor(games) * CMDR_K_EXTERNAL_SCALE` with `CMDR_K_EXTERNAL_SCALE = 1.0` (full standard 40 -> 20 schedule) | Outcomes are definitive; lobbies pass stricter size gates; ELO self-corrects and recomputes from scratch every run. A discount would encode distrust of data we ratified as trustworthy. The constant exists so a future discount is a one-line retune. | | Team handicap | real λ = 1.0 handicap from a running VTSR-T snapshot (`vtsr_t_now`, folded from `elo_history` deltas) over each side's resolved thugs; empty side -> term 0 | Pre-corpus duels see handicap 0 — structurally identical to our own earliest telemetry matches, where everyone sat at the anchor. Corpus-era externals (74) get live handicaps. | | Chronology | day precision; same-day externals sort AFTER telemetry, sheet-row order within a day | F9 logs calendar days, not timestamps. Ordering slop within a day is bounded by K and washes out. Documented approximation. | | Records | headline W-L-D and `matches_commanded_rated` INCLUDE externals; split surfaced via per-rating `duels_external` + ladder chip | Rating and record must tell the same story. `rated_match_count` stays telemetry-only; `external_duels_rated` counts the rest. | | Provenance | per-duel `source: "f9" | "telemetry"`, `external_row`, `map`, `decided_by: "external"`; top-level `external_provider` | UI must never imply a community duel was recorded telemetry. | | Identity | committed alias table `data/external/f9_name_map.json` (exact-normalized names only, never substring); unresolved commanders keyed `name:` (auto-merge on later alias) | Sign-offs: Blue Banana=blue, Bylarge=The_Bylarge, DesUxS=DesUxSAsU, SaggiSponge=sponge, aggressor=Nomad, Muerte=mort, Gravey=Graves, X-Triage=Totokomo, Ultimate=tom. Excluded (participant anywhere -> row dropped): Lone, Mr.Scout, Croatian Knight, General BlackDragon, Systeme, RollyPolly, Jabbapop, Devastator. Scorpion + Echoblammo stay name-keyed. | | Schema | `CMDR_ELO_SCHEMA_VERSION 2 -> 3`; pre-v3 `vtsr_c` / `peak_vtsr_c` values not comparable | Established re-rate signal pattern. | ## Shipped import (funnel, from the importer report) 1309 sheet rows -> 334 no-rosters, 192 uneven, 82 wrong-size, 1 straggler-contaminated, 4 winner-not-commander, 11 no-duration, 3 short -> 682 pre-overlap -> **85 overlaps removed** (ours supersedes; stamped for the adjudication jogger) -> **91 excluded-name rows** -> **506 external duels** (435 pre-corpus + 71 corpus-era at import; the runtime guard conservatively skips 1 more ambiguous row — F9 r1154 "Garden", which collides with an unconsumed same-night Haven recording at 0.70 roster overlap — leaving **505 rated**). 504/506 duels have both commanders Steam64-resolved; thug slots resolve at 96.8%. Map-alias corrections discovered by the first import's runtime guard (1.00 roster-identity collisions, winners agreeing where ours were determined): F9 "Quarry" = `vsrquarry2`, "Alien Dunes" = `st_dunes` (Patton's Proving Ground), "Mort's Wasteland" = `vsrmortwasteland`. ## Interactions - **VTSR-T: untouched.** `scripts/elo.py` never reads the ledger; `elo_history.json` hashed byte-identical across the import (BFB71D9B...46AC, verified twice). The adjudication SESSION (separate step) legitimately moves `wins_*` fields on newly-determined matches; headline `vtsr` is unchanged at ALPHA = 0. - **Adjudication jogger:** ledger `overlaps[]` render one display-only hint line in the outcome prompt; `--adjudicate-f9` widens candidacy to hinted v1/v2-era matches. Operator remains the final authority; F9 hints are never auto-applied. Host-cancelled matches keep Cancelled (attestation of invalidity outranks any winner claim). - **Validator:** `validate_elo.py` §10 replays the enlarged history (telemetry + external audit fields) with replay-integrity ~0.005 ELO. First post-import run: duel prediction 65.1% over n=572 (was ~57% over n=70) — record future runs here as the corpus grows. ## Undo recipe Delete `data/external/f9_ledger.json` (or the specific duels), re-run `python scripts/process_stats.py` — VTSR-C recomputes from scratch every run, so removal is total and immediate. Revert `CMDR_ELO_SCHEMA_VERSION` if reverting the whole feature. --- ## Batch adjudication addendum (2026-09-17) **RATIFIED + EXECUTED.** The operator could not verify the pending outcome-review backlog from memory. Instead of rubber-stamping ~50 prompts blind, the operator ratified ONE rule — *apply F9's outcome to every unclear match whose ledger pairing is machine-provably unambiguous; abandon the rest* — enforced mechanically by the one-shot `scripts/apply_f9_adjudications.py` (dry-run default, `--apply` writes). **Trust basis (measured before applying):** 29/29 winner agreement on every independently-determined paired match, plus a duration fingerprint proving F9 logs the engine timer — median |F9 time − our `duration_sec`| = 0s, p90 = 11s across the 29 calibration pairs. The residual hazard is PAIRING (a rematch-twin cross-pair), not outcomes — and pairing ambiguity is machine-detectable, which human memory of two-year-old lobbies is not. **The six gates (all must pass; pre-registered here):** 1. live `winner.decided_by == "unclear"` (attested / clean_win / contested / cancelled / draw are never touched); 2. `our_team ∈ {1, 2}` (F9 winner resolves by Steam64 to one of our team leaders); 3. roster Jaccard == 1.0 (the 0.8–0.9 relaxed tier was deliberately NOT taken — 2 matches forfeited); 4. |duration diff| ≤ 120s (`DURATION_TOLERANCE_SEC`); 5. no F9-side sibling row (±1 day + same map + same commander pair); 6. no corpus-side sibling, EXCEPT a provable dual recording of the hinted match itself (starts within 300s, durations within 5s) — ignored for uniqueness; the twin id stays unclear on purpose so one physical game rates exactly one duel. **Executed result:** 85 overlap hints → 8 already adjudicated, 21 determined (the validation set), **50 applied**, **6 skipped forever**: r1120 (jaccard 0.778), r1129 (0.636), r1187 (0.818), r1210 (0.9), r1152 (duration contradiction 775s at jaccard 1.0), r1160 (666s + ambiguous sibling). The two duration rejects are exactly the rematch-twin hazard the gates exist for. Applied-set duration diffs: sub-second on 47 of 50, worst 10.7s. Entries live in `data/match_outcome_adjudications.json` with `note: "auto-applied from F9bomber ledger row (…)"` — **the undo handle**: delete every entry whose note starts with `auto-applied from F9bomber ledger`, re-run the pipeline. **Effects verified:** VTSR-C telemetry duels 70 → **120** (`matches_skipped_undetermined` 74 → 23); headline VTSR-T (`vtsr` + `thug_elo`) byte-equal for all 42 players (wins-ladder `wins_*` fields moved legitimately); validator §10 n = 625 duels, accuracy 66.2%, log-loss 0.622. **Two repairs shipped with the batch:** - `compute_commander_elo` gained a same-id **dual-recording guard** (additive `matches_skipped_duplicate_recording` counter): two source recordings of one game sharing a match id produce two `elo_history` entries, and the newly-determined `2026-05-04T03-45-41` exposed VTSR-C rating that duel twice (+4.06 and +4.07). One game = one duel; the second entry now skips. (VTSR-T's own double-walk of same-id twins is pre-existing and deliberately untouched from this module.) - `_investigation/smoke_alpha_override.py` blend tolerance 0.1 → 0.15: it recomputes the blend from independently-ROUNDED emitted fields, so its own error budget reaches ~0.15; the adjudication's legitimate wins_elo movement pushed one player onto the boundary. **UI honesty note:** these 50 matches render the reviewer check-mark badge (`decided_by: "adjudicated"`). The review was rule-level (this memo), not per-match; each entry's `note` field carries the mechanical provenance for anyone auditing the store.