# RECORD (what happened, in order)
- 2026-11-18 02:44 IST: the checks run (condition N). 79 cards (70 cases: 42 TN minus the 2 opening posts, 30 national sample A... counted by card files; 9 controls: C01-C03 genuine-rated clips, C04-C09 own-channel uploads), shuffled with seed 20261121 into 12 runner batches (data/runner-batch-NN.json). Runners: 12 fresh Sonnet 5.5 agents, one batch each; they read only their cards; web search with the fact-check outlets blocked; each card timed by the shell clock; the answer keys (data/keys/) were never given to them. Models: the card-writers, controls gatherers and runners are all Sonnet 5.5 (one family: a limit). The runners may know some cases from training (they say so in already_knew).
- 2026-11-18 02:47 IST: first run scored against the keys: 70 cases: NOT REAL 24 (12 settled by a source: 5 the original, 4 a primary report of the event, 3 the leader's or party's own channel; 12 by absence only), REAL 2 (N005, N014: both genuine events or clips with a false caption or date; the runners saw the problem but wrote REAL), CANNOT TELL 44. Controls 9: REAL 8, CANNOT TELL 1 (no genuine clip was called NOT REAL). Tamil Nadu 40 cases: 10 NOT REAL (4 by a source), 30 CANNOT TELL; national 30: 14 NOT REAL (8 by a source), 14 CANNOT TELL, 2 REAL. Fact-check surfaced on 32 runs, never used. 22 cards had only C1 or no check (search limit). FIRST RUN ONLY; to be redone for the 22.
- Search budget: the 12 first-run agents and the re-runs together used the session's 200-search limit; runs after it were refused.
- 2026-11-18 18:17 IST: the 22 thin cards re-run (batches 14a, 14b: 2 fresh Sonnet 5.5 agents, 11 cards each, all three checks required, no search refused; the refused 13a-c files are in data/runs-refused/, not counted). FINAL COUNTS of the first draw (data/results-final.csv, counted by two routes, same numbers): 70 cases: settled NOT REAL by a source 14 (original 5, primary report of the event 5, leader's or party's own channel 4), NOT REAL by absence only 15, called REAL 4 (N005, N014, TN026, TN034: all four are genuine events or clips with a false caption or date, the runners saw the problem but wrote REAL), CANNOT TELL 37. Tamil Nadu 40: source 5, absence 6, REAL 2, CANNOT TELL 27. National 30: source 9, absence 9, REAL 2, CANNOT TELL 10. Controls 9: REAL 8, CANNOT TELL 1, none called fake. Which check settled the 14 by a source: C3 (find the original) 8, C1 (search the claim) 3, C2 (the leader's or party's own channels) 3; TN: C3 3, C1 2; national: C3 5, C2 3, C1 1. Fact-check surfaced on many runs, never used. Minutes are agent-seconds, not a person's time: not reported. Plan's reading: 14 of 70 settled by a source (20%), far under "3 in 4": the outcome is "few settle at all"; adding absence 29 of 70 (41%) is still under half, and absence is weak (it cannot tell a fake from a thing nobody wrote about).
- 2026-11-18 18:22 IST: SECOND DRAW scored (22 cards: sample B's national cases not in A, minus N114 dropped; runners 15a, 15b; all three checks, nothing refused): settled NOT REAL by a source 7 (leader's or party's own channel 4, primary report of the event 3), NOT REAL by absence only 7, called REAL 2 (N006, an edited real clip, and N077, which the key says is an AI-animated still: the runner found the party's real cooking videos and called THIS clip real without matching it), CANNOT TELL 6. Which check settled the 7: C2 5, C1 2, C3 0. FIRST DRAW (national 30 cases): C3 5, C2 3, C1 1; all 70: C3 8, C1 3, C2 3. THE RULE (PLAN.md, second draw): a check is named best only if it leads in both draws. C3 led the first and settled none in the second; C2 leads the second and tied the first. RESULT: NO CHECK IS NAMED BEST; the video says it is not settled which free check works best, only that the few that settled used a primary source (the original, the leader's own channel, a report of the event) and never a fact-check. Also: across both draws 6 of 92 cases were called REAL by a runner (all genuine events/clips with a false frame, plus N077, a fake called real by a near-match): a viewer's own check can mislead in both directions. Not pooled into one rate: the draws overlap by 7 cases.
