EVIDENCE FILE

PREDICTION TRACK RECORD

Every prediction. Every outcome. Including the ones we got wrong.

2,738 predictions logged since March 2026 - across sports, macro, geopolitics, crypto, and real estate. No backfilled data. No cherry-picked wins. Every run is on the record.

METHODOLOGY CALIBRATION · REBUILT, TIME-GATED, LEAK-AUDITED

Every question below was submitted to the swarm with today_override clamped to a date at least 30 days before resolution, so the model could not see the outcome in its training data. Sampled across 8 real-estate sub-domains. Audit status: passed (2026-05-25; 0/19 anachronisms). See the full anachronism scan →

PER-SUB-DOMAIN · PHASE2D_XDOMAIN

Time-gated cross-domain calibration: every question submitted with `today_override` set to a date ≥ 30 days before resolution. Phase 2D adds skill-weighted v2 archetype clusters to the baseline. Phase 2E (supervisor v1) is excluded — known temporal leak.

WHERE THE SWARM HAS AN EDGE

Brier scores below 0.25 (coin-flip baseline) where the swarm is meaningfully calibrated. Lower is better; 0.0 is perfect.

DEPTH FIRST, BREADTH SECOND

Holodeck is built depth-first for real estate, macro, and private markets - the domains with the structured ground-truth data Gray Capital cares about. We've built dedicated archetype clusters for those. They perform.

Crypto, broad sports, and geopolitics use a general-purpose archetype mix — the same 180-expert baseline that runs across all domains. We haven't built specialized clusters there, and the track record on those categories reflects it: a Brier of 0.236 across 18 resolved predictions. Know the difference: depth-built domains (real estate, macro) earn their use; breadth domains are tools for when structured domain data isn't available.

REAL ESTATE PHASE 2D BATCH — ARCHIVED, NOT IN HEADLINE

Methodology: skill-weighted aggregation across v2 archetypes. 5 questions sampled at random per sub-domain, prompted with point-in-time context only. Lower Brier = better; 0.25 = a coin flip. Negative delta = beats the prior (Phase 1) baseline.

Why it’s not in the headline: a May methodology audit (W2.2) found that 404 of 436 historically resolved predictions had resolution_date before submitted_at — meaning the model could have seen the outcome in training data. We now exclude any prediction without positive lead time from the public Brier. This Phase 2D table is real and is shown for completeness, but its 0.165 aggregate has been retired from the headline pending a rebuild under the tighter v2 protocol (run with today_override to clamp the swarm’s knowledge horizon to the question’s vintage). Watch this page — the rebuilt number is landing this week.

BREAKDOWN BY DOMAIN (LIVE)

FEATURED OUTCOMES

Our strongest calls - highest conviction predictions with verified real-world outcomes.

When we say “70% likely,” does it actually happen 70% of the time? Each dot below is a bin of resolved predictions; dot size = sample count. Sitting on the diagonal means calibrated. Above the line = under-confident; below = over-confident.

74.2% overall accuracy (46/62). Brier score 0.155 - well below the 0.25 coin-flip baseline. The honest test: when the swarm said "70% confident," did it win 70% of the time? Below: every confidence bucket, including the bucket we deliberately said was a coin flip.

The 50-55% row is the most important. A miscalibrated swarm overclaims uncertainty. Ours said "these 6 are basically coin flips" - and they were.

The first two publicly logged, pre-event predictions with verified outcomes. Run through the live engine before game time; both resolved the same night.

WHY THIS PAGE EXISTS

WHAT WE DON'T HAVE YET

Every public run is logged. This page updates as outcomes resolve.