intelligence · measured, not claimed

Is GENESIS getting smarter?

Not yet. On every narrative time window with enough outcomes, GENESIS makes bigger errors than a forecaster that only knows how often themes usually grow. v0 is frozen, so it cannot improve on its own; progress has to come from new versions.

24H narratives · Below the baseline6H narratives · Below the baseline1H narratives · Below the baseline

Is it better than just guessing the usual rate?

Each row is one kind of prediction. The diamond is GENESIS's average error; the amber line is the error of the simple baseline. Left is better.

24H narratives: 6% bigger errors than the baseline · 6H narratives: 19% bigger errors than the baseline · 1H narratives: 54% bigger errors than the baseline.

Compared with other simple strategies (always fades, coin flip, momentum rule)

narrative · 24H · n = 752 · happened 27.3%

strategy	Brier	accuracy	precision	recall
GENESIS v0	0.212	63.7%	33.3%	33.2%
baseline (usual rate)	0.200	—	—	—
coin flip	0.250	—	—	—
always fades	—	72.7%	—	0.0%
momentum rule (accelerates iff 15-min share > 1-hour share)	—	56.8%	32.1%	52.7%

narrative · 6H · n = 1024 · happened 22.9%

strategy	Brier	accuracy	precision	recall
GENESIS v0	0.212	62.5%	29.2%	44.7%
baseline (usual rate)	0.178	—	—	—
coin flip	0.250	—	—	—
always fades	—	77.1%	—	0.0%
momentum rule (accelerates iff 15-min share > 1-hour share)	—	56.6%	28.7%	60.0%

narrative · 1H · n = 1100 · happened 16.8%

strategy	Brier	accuracy	precision	recall
GENESIS v0	0.218	62.0%	22.7%	52.4%
baseline (usual rate)	0.141	—	—	—
coin flip	0.250	—	—	—
always fades	—	83.2%	—	0.0%
momentum rule (accelerates iff 15-min share > 1-hour share)	—	56.7%	22.5%	64.3%

token · 6H · n = 530 · happened 2.3%

strategy	Brier	accuracy	precision	recall
GENESIS v0	0.024	97.7%	—	0.0%
baseline (usual rate)	0.023	—	—	—
coin flip	0.250	—	—	—
always dies	—	97.7%	—	0.0%

token · 1H · n = 550 · happened 2.2%

strategy	Brier	accuracy	precision	recall
GENESIS v0	0.023	97.8%	—	0.0%
baseline (usual rate)	0.023	—	—	—
coin flip	0.250	—	—	—
always dies	—	97.8%	—	0.0%

When GENESIS says 40%, does it happen 40% of the time?

Narrative predictions grouped by the probability GENESIS gave. Dots on the diagonal would be perfectly honest probabilities; dots below it mean GENESIS expected more than happened.

Too optimistic. On average GENESIS gave growth a 40% chance; it happened 22% of the time.

Confidence buckets and calibration error

Calibration error (ECE), narratives: 0.182 · all predictions: 0.143

confidence	n	accuracy	Brier
50%–60%	1112	52.3%	0.252
60%–70%	750	53.0%	0.260
70%–80%	440	68.3%	0.224
80%–90%	491	91.6%	0.089
90%–100%	1163	98.1%	0.020

Is it changing over time?

Narrative skill against the baseline, per day of outcomes.

v0 is frozen, so day-to-day changes reflect the market, not learning. Candidate versions are tested in the lab.

Full technical record

All metrics

predictions4232

open276

scored3956

void0

accuracy (decisive)73.6%

correct / incorrect / partial2732 / 981 / 243

Brier (all)0.162

baseline Brier0.137

skill-0.180

log loss0.475

mean confidence74.5%

average lead timenot measured at stage 0

Pooled numbers mix different tasks and are shown only for completeness; read the per-class results above.

By prediction type and horizon

type · horizon	n	accuracy	partial	Brier	baseline Brier	skill	base rate	precision	recall	FPR	mean conf.
narrative · 24H	752	65.1%	96	0.212	0.200	-0.059	27.3%	33.3%	33.2%	24.9%	64.9%
narrative · 6H	1024	64.0%	77	0.212	0.178	-0.192	22.9%	29.2%	44.7%	32.2%	67.0%
narrative · 1H	1100	62.4%	70	0.218	0.141	-0.542	16.8%	22.7%	52.4%	36.1%	68.8%
token · 6H	530	97.7%	0	0.024	0.023	-0.029	2.3%	—	0.0%	0.0%	94.0%
token · 1H	550	97.8%	0	0.023	0.023	-0.036	2.2%	—	0.0%	0.0%	94.0%

By engine and GENESIS version

engine	n	accuracy	partial	Brier	baseline Brier	skill	base rate	precision	recall	FPR	mean conf.
narrative-heuristic-v0	2876	63.7%	243	0.214	0.171	-0.255	21.7%	27.3%	43.2%	32.0%	67.1%
token-heuristic-v0	1080	97.8%	0	0.024	0.022	-0.062	2.2%	—	0.0%	0.0%	94.0%

version	n	accuracy	partial	Brier	baseline Brier	skill	base rate	precision	recall	FPR	mean conf.
0.1.0	3956	73.6%	243	0.162	0.137	-0.180	16.4%	27.3%	41.6%	21.8%	74.5%

Methodology

Brier score: mean squared error of the stated probability (0 perfect, 0.25 coin flip). The baseline forecaster predicts, for each question, the share of positive outcomes among outcomes graded before that question was asked (no hindsight). Skill = 1 − Brier ÷ baseline Brier. Outcomes with under 90% observation coverage are void and excluded. Verdicts need at least 30 scored outcomes and a margin of ±2% skill.