Five-league opening-round brief

This completed historical issue shows the full format planned for the five-league brief. It uses the World Cup round-by-round track, not the league benchmark, so the two fields and their scores should not be compared.

The scope is eight round-of-16 fixtures, matches 89–96. Picks were locked before the first kickoff. The analysis below was recomputed read-only from the published normalized files, results and scoring code on 14 August 2026. It did not rewrite a prediction, hash or tag.

1. Scorecard

Metric	Result
Ranked World Cup roster	40 models
Valid round-of-16 model files	38/40
Valid model-match predictions	304 (38 × 8)
Missing predictions against the full roster	16 (2 × 8)
Exact score classifications	1/304
Correct goal-difference classifications	81/304
Correct-outcome-only classifications	85/304
No 90-minute score points	137/304
Correct advancers	195/304
Base score points	250
Advance-bonus points	195
Total points across the round	445

The 90-minute score classes are mutually exclusive: exact score earns 3 points, correct goal difference 2, correct outcome 1, otherwise 0. A correct knockout advancer adds 1.

Two roster models produced no valid round file:

• GPT-3.5 Turbo returned six HTTP-200 responses, but every response failed validation.
• Llama 3 70B returned six HTTP 404 responses saying no endpoint was available.

Round winner: Mistral Medium 3.5 scored 17 points over eight picks. Claude 3 Haiku also scored 17; Mistral won the tiebreak by scoring on seven matches rather than six.

Cumulative round-by-round leader after the round of 16: GLM 4.7 Flash had 50 points. GPT-4o also had 50, but GLM led the tiebreak with seven exact scores to GPT-4o's six.

The round's only exact score came from GPT-5.5 Pro, which predicted Portugal 0–1 Spain in match 93.

2. Largest leaderboard moves

Selection rule

1. Snapshot A is the cumulative round-by-round table after round-of-32 matches 74–88. Match 73 is excluded because it had kicked off before that round was collected.
2. Snapshot B adds round-of-16 matches 89–96.
3. Both snapshots use the published live-track tiebreakers: points, exact scores, matches with points, correct advancers, then a shared rank.
4. Compare only models with valid files in both rounds, sort by rank change, and publish the top three rises and top three falls, including a tie at the cutoff.

The pre-round field had 38 models. The post-round table had 39 models with at least one live-round file: 37 had both rounds and 23 stored picks, GPT-3.5 Turbo had only its 15 round-of-32 picks, and Claude Fable 5 entered with eight round-of-16 picks. Every mover below had 23 stored picks.

Largest rises

Model	Before	After	R16 points	Move
Claude 3 Haiku	#32/38, 26 pts	#16/39, 43 pts	+17	+16
Mistral Medium 3.5	#23/38, 29 pts	#8/39, 46 pts	+17	+15
Qwen 2.5 72B	#37/38, 24 pts	#26/39, 40 pts	+16	+11

Largest falls

Model	Before	After	R16 points	Move
MiniMax M3	#12/38, 32 pts	#27/39, 39 pts	+7	−15
Qwen3.7 Max	#22/38, 29 pts	#35/39, 36 pts	+7	−13
Command A	#13/38, 32 pts	#22/39, 41 pts	+9	−9
GPT-5.5	#8/38, 33 pts	#17/39, 42 pts	+9	−9

Claude Fable 5 was a new entrant, not a mover: it joined at #39/39 with 13 points from eight picks. The repository did not store a historical leaderboard snapshot; these positions are read-only derivations from the locked round files and recorded results.

3. Five highest-consensus calls

Selection rule

For each fixture, count the modal 90-minute outcome (home win, draw or away win) over the 38 valid files. Rank matches by support for that outcome, descending, break ties by match number, and take exactly five. A call succeeds only when the modal outcome matches the recorded 90-minute outcome. The modal scoreline is shown for context but does not decide the verdict.

Match	Outcome split H/D/A	Modal score	Recorded result	Verdict
89: Paraguay–France	0/0/38 away	0–2 (20/38)	0–1	Success
91: Brazil–Norway	38 home/0/0	2–1 (23/38)	1–2	Failure
95: Argentina–Egypt	38 home/0/0	2–0 (25/38)	3–2	Success
93: Portugal–Spain	1/8/29 away	1–2 (26/38)	0–1	Success
94: United States–Belgium	23 home/6/9	2–1 (21/38)	1–4	Failure

The fixed rule selected three successes and two failures. It includes the unanimous Brazil call, which was wrong, rather than replacing it with a better-looking result after the fact.

4. Lock and hash audit

The live manifest at /data/predictions-live/manifest.json says:

• manifest lock: 2026-07-04T07:21:30.187Z;

The stored integrity record at /data/hashes/r16-live.txt says:

• track: round-by-round (live, real fixtures);
• stage: r16;
• models: 38;
• digest generated: 2026-07-04T07:21:51.720Z;
• SHA-256: 9383ed8eb9ac1e8e69b70cb7efb1d0f7e8aceb60f1e98de2a18fc918288e64a3.

Git records the annotated tag predictions-r16-live, tag object 11e14c44e8f6dec7568a4bebdfff4da47b7a1294, pointing to commit db29c10873bc409f629a6a84d07cef04fa5634dd. The commit time was 2026-07-04T10:22:09+03:00 (07:22:09Z). The earliest round-of-16 kickoff, Canada–Morocco, was 17:00Z that day.

The digest covers canonicalized normalized records sorted by model and match. It does not claim to cover raw logs, usage/cost metadata, fixtures or results. Results were committed later, after the matches. Git history shows no later change to the locked R16 prediction directory or hash file.

5. Parser, provider and data issues

The public raw record contains 40 model files and 54 attempts: 38 successful records and 16 failed records. Forty-seven attempts returned HTTP 200, six returned HTTP 404, and one transport/parser record had no HTTP status. The 38 successful records produced 38 normalized files.

Material examples:

• Llama 3 70B made six attempts; every one returned HTTP 404 with "No endpoints found". It has no normalized round file.
• GPT-3.5 Turbo made six HTTP-200 attempts, but invalid team names and contradictions between the score and named advancer prevented every response from passing validation. It has no normalized round file.
• Gemma 2 27B first predicted Portugal 2–1 Spain while naming Spain as the advancer. Its second response corrected the contradiction and passed.
• Llama 4 Scout's first response began with a malformed JSON fragment and could not be parsed. Its second response passed.
• GPT-5.5 Pro returned empty content, then an incomplete JSON error, then passed on its third attempt. The successful call used 10,894 tokens, cost $1.32207 as recorded in the file, and later supplied the round's only exact score.

A coverage caveat matters for the cumulative table: a missing entire round file does not create eight explicit zero-scored matches. A model with an earlier live file stays ranked on its earlier points, while a new model can enter with a later round only. The movement table above limits its listed models to those with valid files in both rounds, although their field ranks still come from the mixed-coverage table.

6. Evidence and correction log

Primary evidence:

• R16 fixtures
• recorded results
• live manifest
• normalized R16 files
• raw R16 logs
• stored hash at the lock commit
• scoring method
• validator source

Result and correction record:

• Matches 89–92 were recorded in 9acdc71d.
• Matches 93–96 were recorded in 9fdf9bd9.
• Commit 62e88aee corrected match 92's weather-delayed kickoff from 00:00Z to 01:00Z. It changed fixture timing, not teams, predictions, scoring or the hash.
• No later edit to the locked R16 prediction/hash tree was found. The current changelog does not carry a dedicated entry for the match-92 timing correction even though the methodology says corrections are listed there. This is a transparency gap; the commit is linked directly here rather than calling the changelog complete.