Coherence training for forecasts · Kelvin, MATS

How you ask decides whether fitting helps

Asked only "A and B (and C)", Qwen3-32B says rare combinations are too likely. Fitting those answers is worse than just multiplying its singles (0.340 vs 0.298) and hurts the singles. Asked a sign-balanced set, the fit beats independence on joints (0.393 vs 0.480) and repairs the singles (0.668 → 0.602).

839triangles
6weeks
16questions each
$4.63cost

16 questions per triangle

ABCA Iran-run attack abroad kills 5+B Attack on US soil kills 100+C Major terror attack in EuropeRandom signs here: not A, B, C.What happened: none of A, B, C.Qwen3-32B answertrue?Chainnot A and B and C0.15noA and B and C0.05nonot A and not B and C0.15nonot A and B and not C0.15noB and C0.15nonot A and C0.15nonot A and B0.15nonot A0.75yesB0.15noC0.15noHow many of the three happennone0.25noexactly one0.25yesexactly two0.15noall three0.12nosum 0.77, should be 1Mirrors (OR form)A or not B or not C0.95yesnot B or not C0.95yes
One real triangle, week of 23 June 2025. Triangles are three events GPT-6 Luna linked to each other (it picks up to 5 related events per event). Each event is randomly asked as YES or NO, so no answer type always favours "too likely".

The bias follows rarity, not wording

-0.10+0.00+0.10+0.20One event+0.00Two events, and+0.09Three events, and+0.13Three, one flipped+0.12None of three+0.16Exactly one+0.14Exactly two+0.09All three (count)+0.15Not all three (or)-0.07Not both (or)-0.01answer minus outcome rate (right = says too likely)
Joint and count questions are mostly rare, so they are over-stated; the OR mirrors are common, so under-stated.
-0.2+0.0+0.2+0.4+0.60%20%40%60%80%100%how often the statement came truey: answer minus outcome rateOne event, 0 negated: +0.20 (n=1233)One event, 1 negated: -0.19 (n=1284)Two events, and, 0 negated: +0.22 (n=599)Two events, and, 1 negated: +0.16 (n=1268)Two events, and, 2 negated: -0.17 (n=650)Three events, and, 0 negated: +0.17 (n=93)Three events, and, 1 negated: +0.14 (n=320)Three events, and, 2 negated: +0.20 (n=314)Three events, and, 3 negated: -0.13 (n=112)Three, one flipped, 0 negated: +0.18 (n=320)Three, one flipped, 1 negated: +0.13 (n=907)Three, one flipped, 2 negated: +0.17 (n=976)Three, one flipped, 3 negated: -0.14 (n=314)None of three, 0 negated: -0.15 (n=93)None of three, 1 negated: +0.21 (n=320)None of three, 2 negated: +0.19 (n=314)None of three, 3 negated: +0.24 (n=112)Exactly one, 0 negated: +0.22 (n=93)Exactly one, 1 negated: -0.20 (n=320)Exactly one, 2 negated: +0.33 (n=314)Exactly one, 3 negated: +0.56 (n=112)Exactly two, 0 negated: +0.24 (n=93)Exactly two, 1 negated: +0.21 (n=320)Exactly two, 2 negated: -0.22 (n=314)Exactly two, 3 negated: +0.49 (n=112)All three (count), 0 negated: +0.19 (n=93)All three (count), 1 negated: +0.17 (n=320)All three (count), 2 negated: +0.22 (n=314)All three (count), 3 negated: -0.13 (n=112)Not all three (or), 0 negated: +0.24 (n=112)Not all three (or), 1 negated: -0.15 (n=314)Not all three (or), 2 negated: -0.11 (n=320)Not all three (or), 3 negated: -0.04 (n=93)Not both (or), 0 negated: +0.21 (n=219)Not both (or), 1 negated: -0.09 (n=422)Not both (or), 2 negated: -0.07 (n=198)
Every question type, split by how many NOTs it has. Rare statements are said too likely, common ones too unlikely, whatever the words. A YES single (true 22%) is +0.20; a NO single (true 80%) is −0.19.

The raw answers break the rules

1.55the 4 counts add to (should be 1)
18%order checks fail (triple ≤ pair ≤ single)
79%triangles with at least one failure
-0.20+0.00+0.20+0.40+0.60Counts: none+one+two+all − 1+0.55Pair − triple − flipped triple-0.16Triple + its OR mirror − 1+0.06Pair + its OR mirror − 1+0.08NO single + YES single − 1+0.01error; should be 0
NO and YES singles agree on average but are 0.22 off in a typical triangle. The fit below turns these into one coherent set of probabilities per triangle.

Result: the balanced design lets the fit help

0.200.250.300.350.400.450.500.550.500.550.600.650.700.75JointsAND-only (35 weeks)0.3810.3400.298Sign-balanced (839 triangles)0.4580.3930.480SinglesAND-only (35 weeks)0.6110.687Sign-balanced (839 triangles)0.6680.602log loss against outcomes, lower is better; bars: 95% CI
AND-only: held-out joints and all singles over 35 weeks (weekly world model). Sign-balanced: the same model, 839 triangles in 6 of those weeks, joints scored on the balanced forms, YES singles only. Compare within a design; the question sets differ. Fitted = the closest coherent probabilities to all answers. Correcting each type for its bias adds nothing more.