The lab
This page is not GENESIS. It is where its developers test candidate models beside it. Candidates run in shadow: on the same questions, at the same moment, scored on the same outcomes as GENESIS v0. They never influence GENESIS's predictions, and GENESIS cannot see them. MEMORY is still locked; nothing here means GENESIS gained it.
In plain words
The question being tested: can a model tell a theme that is about to grow again from one that is about to fade? The main candidate did better than both GENESIS v0 and the baseline on past data it never saw. Now it has to prove that on the future: 960 of 100 paired 1-hour outcomes collected, errors so far GENESIS 0.216 vs candidate 0.117 (too early to read). Nothing here changes GENESIS, and it could only go live by a deliberate decision once MEMORY exists.
narrative-v1c1-full-360
would require: memory.persistent, learning.signal_evaluationGENESIS v1 CANDIDATE (shadow research, not GENESIS's active intelligence). 6H narrative acceleration: L2 logistic regression on 16 features from up to 6h of observations (base rate + momentum + mean reversion: share, velocity, acceleration, distance from previous peak, persistence, age, concentration, duplicates, new entrants, breadth). Coefficients learned by developers from v0's observation record.
No held-out evaluation was possible with the data available (10712 examples over too short a span for a purged split). Its only evaluation is the live record below.
| n | Brier | Skill vs base rate | |
|---|---|---|---|
| GENESIS v0 (active) | 920 | 0.2128 | -0.174 |
| narrative-v1c1-full-360 (shadow) | 920 | 0.1674 | +0.076 |
| Base-rate forecaster | 920 | 0.1812 | 0 |
Candidate closer to the outcome on 585 of 920 paired questions. 48 pending.
| v0 # | Question | v0 | v1 | Outcome |
|---|---|---|---|---|
| #4032 | "presale" | 15% | 30% | faded |
| #4031 | "tumor" | 8% | 14% | faded |
| #4030 | "timmy" | 8% | 12% | faded |
| #4029 | "worth" | 30% | 22% | faded |
| #4016 | "affiliated" | 39% | 39% | faded |
| #4015 | "team" | 43% | 41% | faded |
| #4014 | "sonny" | 50% | 17% | faded |
| #4013 | "certified" | 11% | 8% | faded |
| #4000 | "usepaid" | 41% | 28% | faded |
| #3999 | "mba" | 18% | 7% | faded |
| #3998 | "bot" | 31% | 16% | faded |
| #3997 | "shift" | 11% | 24% | faded |
narrative-v1c1-full-60
would require: memory.persistent, learning.signal_evaluationGENESIS v1 CANDIDATE (shadow research, not GENESIS's active intelligence). 1H narrative acceleration: L2 logistic regression on 16 features from up to 6h of observations (base rate + momentum + mean reversion: share, velocity, acceleration, distance from previous peak, persistence, age, concentration, duplicates, new entrants, breadth). Coefficients learned by developers from v0's observation record.
| On unseen later data | n | Brier | Base-rate Brier | Skill | PR-AUC | ECE | Calibration |
|---|---|---|---|---|---|---|---|
| 2026-10-09 18:23 UTC → 2026-10-09 21:30 UTC | 8751 | 0.1504 | 0.1756 | +0.143 | 0.452 | 0.015 | none |
Block bootstrap over 20 held-out moments: Brier difference vs base rate 90% CI [-0.0290, -0.0212]; vs v0 [-0.0950, -0.0825]. Negative = candidate better. One market regime only.
| n | Brier | Skill vs base rate | |
|---|---|---|---|
| GENESIS v0 (active) | 960 | 0.2164 | -0.537 |
| narrative-v1c1-full-60 (shadow) | 960 | 0.1165 | +0.172 |
| Base-rate forecaster | 960 | 0.1408 | 0 |
Candidate closer to the outcome on 766 of 960 paired questions. 8 pending.
| v0 # | Question | v0 | v1 | Outcome |
|---|---|---|---|---|
| #4188 | "apple" | 25% | 29% | faded |
| #4187 | "bird" | 5% | 2% | faded |
| #4186 | "uxento" | 60% | 44% | faded |
| #4185 | "cat" | 47% | 31% | faded |
| #4172 | "rough" | 76% | 6% | faded |
| #4171 | "inu" | 62% | 29% | accelerated |
| #4170 | "hello" | 63% | 12% | faded |
| #4169 | "dog" | 52% | 19% | faded |
| #4156 | "attention" | 37% | 19% | faded |
| #4155 | "dirty" | 23% | 2% | faded |
| #4154 | "alon" | 15% | 3% | faded |
| #4153 | "muse" | 34% | 1% | faded |
narrative-v1c1-short-360
would require: learning.signal_evaluationGENESIS v1 CANDIDATE (shadow research, not GENESIS's active intelligence). 6H narrative acceleration: L2 logistic regression on 10 features from the last 60 minutes only (ablation). Coefficients learned by developers from v0's observation record.
No held-out evaluation was possible with the data available (10712 examples over too short a span for a purged split). Its only evaluation is the live record below.
| n | Brier | Skill vs base rate | |
|---|---|---|---|
| GENESIS v0 (active) | 920 | 0.2128 | -0.174 |
| narrative-v1c1-short-360 (shadow) | 920 | 0.1676 | +0.075 |
| Base-rate forecaster | 920 | 0.1812 | 0 |
Candidate closer to the outcome on 615 of 920 paired questions. 48 pending.
| v0 # | Question | v0 | v1 | Outcome |
|---|---|---|---|---|
| #4032 | "presale" | 15% | 32% | faded |
| #4031 | "tumor" | 8% | 9% | faded |
| #4030 | "timmy" | 8% | 10% | faded |
| #4029 | "worth" | 30% | 17% | faded |
| #4016 | "affiliated" | 39% | 50% | faded |
| #4015 | "team" | 43% | 42% | faded |
| #4014 | "sonny" | 50% | 15% | faded |
| #4013 | "certified" | 11% | 26% | faded |
| #4000 | "usepaid" | 41% | 26% | faded |
| #3999 | "mba" | 18% | 6% | faded |
| #3998 | "bot" | 31% | 8% | faded |
| #3997 | "shift" | 11% | 16% | faded |
narrative-v1c1-short-60
would require: learning.signal_evaluationGENESIS v1 CANDIDATE (shadow research, not GENESIS's active intelligence). 1H narrative acceleration: L2 logistic regression on 10 features from the last 60 minutes only (ablation). Coefficients learned by developers from v0's observation record.
| On unseen later data | n | Brier | Base-rate Brier | Skill | PR-AUC | ECE | Calibration |
|---|---|---|---|---|---|---|---|
| 2026-10-09 18:23 UTC → 2026-10-09 21:30 UTC | 8751 | 0.1636 | 0.1756 | +0.068 | 0.351 | 0.015 | none |
Block bootstrap over 20 held-out moments: Brier difference vs base rate 90% CI [-0.0160, -0.0081]; vs v0 [-0.0821, -0.0689]. Negative = candidate better. One market regime only.
| n | Brier | Skill vs base rate | |
|---|---|---|---|
| GENESIS v0 (active) | 960 | 0.2164 | -0.537 |
| narrative-v1c1-short-60 (shadow) | 960 | 0.1299 | +0.077 |
| Base-rate forecaster | 960 | 0.1408 | 0 |
Candidate closer to the outcome on 739 of 960 paired questions. 8 pending.
| v0 # | Question | v0 | v1 | Outcome |
|---|---|---|---|---|
| #4188 | "apple" | 25% | 18% | faded |
| #4187 | "bird" | 5% | 3% | faded |
| #4186 | "uxento" | 60% | 35% | faded |
| #4185 | "cat" | 47% | 31% | faded |
| #4172 | "rough" | 76% | 11% | faded |
| #4171 | "inu" | 62% | 25% | accelerated |
| #4170 | "hello" | 63% | 18% | faded |
| #4169 | "dog" | 52% | 19% | faded |
| #4156 | "attention" | 37% | 11% | faded |
| #4155 | "dirty" | 23% | 8% | faded |
| #4154 | "alon" | 15% | 12% | faded |
| #4153 | "muse" | 34% | 12% | faded |
Finding: token outcome labels
judge-v1 counts a token as having sustained attention if it has graduated, regardless of what happened after. About 1% of launches are already graduated when GENESIS first sees them, and their value then collapses. Every token "survivor" in v0's record so far is one of these. The frozen record is not changed; token research uses a stricter label (still trading and at least 2× launch market cap at the horizon) until a decision is made on a versioned judge-v2.