BY REQUEST
EVALS · BENCHMARKS · CERTIFIED RUNS

Prove the climb.

We measure what an intervention did to a model. The protocol is fixed before the run starts, and every contribution is ablated after it.

PRICED PER RUN · OPEN TO INDIVIDUALS AND COMPANIES
BASELINEFRONTIER
01 / PROTOCOL

Frozen before the run. Ablated after it.

01FREEZE
Model revision, eval suite, held-out split and seeds are fixed and hashed.
02BASELINE
The suite runs on the frozen model. The score is the reference point.
03INTERVENE
Your contribution is applied. Method and hyperparameters go on the record.
04RE-RUN
Same suite, same split, same seeds. The gap between the two scores is the delta.
05ABLATE
Each contribution is withheld in turn, isolating the share of the delta it carries.
06CERTIFY
Delta, confidence interval and a hash over everything above.
FIG. 1 — HELD-OUT EVALUATION SPECIMEN · ILLUSTRATIVE FIGURES
45 50 55 60 65 corecraft-v2 score baseline · frozen · 52.1 intervention Δ +9.6 ±0.4 (95% CI) held-out · n = 1,240 · seeds 0–4 0 1k 2k 3k 4k training steps 45 50 55 60 65 baseline · 52.1 intervention Δ +9.6 held-out · n = 1,240 0 2k 4k training steps · corecraft-v2
RUN RECORD
modelqwen3-32b-base
revision8f21c0d
methodGRPO · 4k steps
suitecorecraft-v2
splitheld-out · 1,240
seeds0–4
baseline52.1
final61.7
delta+9.6 ±0.4
02 / ATTRIBUTION

Which input moved it, and by how much.

ENVIRONMENTS & TASKS
The setting the model acts in and the problems it is scored on.
EXPERT TRAJECTORIES
Step-by-step rollouts of a task done correctly by someone who does it for a living — the training signal, and the reference the grader scores against.
PROPRIETARY CORPORA
Workflow records from inside a company, measured in a run that company controls.
FIG. 2 — LEAVE-ONE-OUT DECOMPOSITION total Δ +9.6
+4.1ENVIRONMENTS & TASKS
+3.2EXPERT TRAJECTORIES
+2.3PROPRIETARY CORPORA
03 / RECORD

Reproducible from the record alone.

Model, revision, method, suite, split, seeds, delta and interval — under one hash, verifiable by anyone you hand it to.

CERTIFICATExc-2026-0417
RUNS4 (1 baseline, 3 ablation)
DIGEST4a9f2e71…c12b
DELTA+9.6 ±0.4
SCOPE — A certificate states what one run did, on one model, under one protocol. It carries no claim about what the same data will do in a different run.

Request access.

Tell us the model you want to move and the benchmark you want to move it on.