Source integrity · October 15, 2026
Keep each game
attached to its evidence.
Fourteen historical loss summaries needed correction. The underlying game IDs, results and numerical ranking inputs pass the checks below. The repair changes presentation and input validation; it does not recalculate NRR.
2024 and 2025 · six checkpoints each
What was checked
1,620team rows
16,964game-evidence rows
14corrected summaries
Every frozen archive row was checked for duplicate game IDs within a team, loss counts, score direction, period totals, metadata and dates beyond its cutoff. All 7,753 available two-team game pairs agree on opponents, reciprocal scores, date, stadium name and designated home team. No failures were found in those structural checks.
A targeted check of 29 provider game records covers all 18 conference championship game IDs in the two selection snapshots and additional games for BYU, Texas and UNLV. It matches 148 archived perspectives on participants, scores, UTC date and home/road/neutral status. This is not provider verification of every historical game or an audit of the 2026 preview.
Four winning teams have one fewer Quality-Win Depth evidence entry than wins: Georgia and Boise State in 2024, Texas Tech and Boise State in 2025. The generator deliberately credits one selected win per distinct opponent in that portfolio. These are not missing wins in the record; the portfolio must not be used as a complete schedule ledger.
Dates in the archive are UTC provider dates. A late Saturday game can have a Sunday UTC date. Eleven sampled archive perspectives have a stadium-name difference from today’s provider response; names can change. Those differences remain recorded for dated-source research, without changing location or scoring inputs.
Confirmed presentation repair
Two games against one opponent are still two games.
BYU lost 7–29 at Texas Tech on November 8, 2025 (game 401756945), then 7–34 at neutral-site AT&T Stadium on December 6 (401777327). Both games remain in the record. The second loss had inherited the first game’s road label and late-game checkpoints because the summary lookup used opponent name instead of game ID. BYU’s championship report confirms the rematch and 11–2 record.
The same lookup issue affected Texas–Georgia in 2024 and UNLV–Boise State in both seasons. Five other team/checkpoint pairs had two summary entries in an order different from their detailed evidence. New summaries derive from each evidence row; future generator lookups use game ID.
| Checkpoint and team | Game ID | Corrected description |
|---|---|---|
| 2024-cfp-2New Mexico | 401643711 | Fresno State (home · Q4 -18 · 6:00 -22 · 2:00 -14) |
| 2024-cfp-2New Mexico | 401643750 | Wyoming (home · Q4 +10 · 6:00 +3 · 2:00 -4) |
| 2024-selectionTexas | 401673469 | Georgia (neutral · title game · Q4 -7 · 6:00 tied · 2:00 -3) |
| 2024-selectionUNLV | 401673468 | Boise State (road · title game · Q4 -21 · 6:00 -14 · 2:00 -14) |
| 2024-selectionAppalachian State | 401643844 | Georgia Southern (road · Q4 -2 · 6:00 -2 · 2:00 -2) |
| 2024-selectionAppalachian State | 401643799 | Marshall (road · Q4 -14 · 6:00 -14 · 2:00 -21) |
| 2025-cfp-2NC State | 401754574 | Pittsburgh (road · Q4 -25 · 6:00 -25 · 2:00 -25) |
| 2025-cfp-2NC State | 401754548 | Virginia Tech (home · Q4 -6 · 6:00 -2 · 2:00 -2) |
| 2025-cfp-3Wyoming | 401760412 | Fresno State (road · Q4 -7 · 6:00 -14 · 2:00 -21) |
| 2025-cfp-3Wyoming | 401760385 | UNLV (home · Q4 -14 · 6:00 -21 · 2:00 -21) |
| 2025-cfp-5North Carolina | 401754600 | Duke (home · Q4 -6 · 6:00 +1 · 2:00 -7) |
| 2025-cfp-5North Carolina | 401754612 | NC State (road · Q4 -15 · 6:00 -22 · 2:00 -29) |
| 2025-selectionBYU | 401777327 | Texas Tech (neutral · title game · Q4 -14 · 6:00 -24 · 2:00 -27) |
| 2025-selectionUNLV | 401777326 | Boise State (road · title game · Q4 -7 · 6:00 -14 · 2:00 -17) |
The original JSON archives stay unchanged so frozen experiments remain reproducible. A hash-verified copy of the original generator is retained for the frozen reader-example harness. The rankings loader derives corrected summaries without changing scores, components, weights, record, detailed evidence or ordering. The downloadable report preserves the original hash and before/after text for each correction. Detailed résumé cards already used the correctly identified evidence rows.
Rank, benefit and input provenance
A protected benefit does not rewrite the final ranking.
The canonical proposal already says that a required title game cannot erase the participant’s minimum benefit: pre-title Nos. 1–2 claim a double bye, 3–10 a single bye, 11–14 a home opener, and 15–18 a berth, subject to fixed capacities. A new final-ranking eligibility floor would change that policy; it is not a correction to the existing wording.
Texas entered its 2024 title game at committee No. 2 and finished No. 3. It keeps its double bye while remaining ranked third. Virginia entered its 2025 title game at No. 17 and finished No. 19. Its retained berth places it in the historical protected field instead of USC. These are proposal counterfactuals, not the actual CFP fields. Inspect both protected reconstructions →
All 50 stored historical candidate rows match the pre-title and final ranks in the versioned input specification. The 2024 and 2025 committee tables provide both releases. The 2025 table repeats pre-title No. 24 for North Texas and James Madison; JMU’s dated December 2 release confirms No. 25. That source conflict and resolution are retained.
| Input | Acceptance rule |
|---|---|
| Paired ranking snapshots | Same ranking family and frozen formula/version; pre-title release precedes final release. Retain source, input hash, code hash, capture time and game cutoff. |
| Team and championship identities | Stable identity mapping, unique ranks, explicit unranked status, completed title-game participants and final champion access list. One scoring bonus does not confer an automatic berth. |
| Minimum benefits | Record the pre-title band and retained benefit separately from final rank. Validate candidate and entrant facts agree. |
| Capacity | Exactly two double byes, eight single byes and eight opening-round teams. Reject conflicting claims for explicit review. |
The paired-input gate rejects mixed ranking families or versions, reversed releases and missing provenance. The historical allocator now rejects duplicate identities/ranks, invalid champion lists and entrant protection facts that disagree with the candidate table. It still stops when protected claims exceed capacity; a universal conflict-resolution policy remains open.
The next NRR comparison must use NRR pre-title and NRR final snapshots under one frozen method. Combining committee protection with an NRR final order would test a separately declared hybrid policy. The earlier 2024 final-only membership sensitivity remains preserved and does not become a protected NRR bracket.
Inspect the correction lineage
Download audit and fourteen corrections (JSON) · Download protection inputs and specification (JSON)
Run node scripts/audit-ranking-source-integrity.mjs against the retained checkout. The report records hashes for all twelve source archives and the 29 captured provider responses. Original research manifests, results and unused forward observations are preserved. No weights were tuned and no new predictive validation is claimed.
2 coherent committee ranking pairs are specified. The protected NRR comparison now reports both 2024 and 2025 paired-snapshot cases. Broader dated venue verification and a general capacity-conflict rule remain future work.