LIVE · VERIFIER v4.2.1 · 99.982% UPTIME
AX-04821RAMSEY·R55+5.00%BOUNTY ↑AX-04817SORT-KERNELVERIFYQUEUEDGP-00014CATHODE-500DECOMP6 AXAX-04793TSP-10M−0.8%SCORE ↓AX-04788BANDGAP-SI+12.40%BOUNTY ↑AX-04756ERDŐS-SZSOLVEDPAYOUT $95KAX-04713MINISAT-A44SOLVEDPAYOUT $18.5KAX-04709CHIP-ROUTE-D2+2.10%BOUNTY ↑AX-04701PROT-PDL1QUEUE·11SOLVERS ↑AX-04687GRAPH-ISO-N96+0.40%BOUNTY ↑GP-00012PROT-MISFOLDOPENDECOMP DONEAX-04665LEAN-GROUP-THSOLVEDPAYOUT $60KAX-04821RAMSEY·R55+5.00%BOUNTY ↑AX-04817SORT-KERNELVERIFYQUEUEDGP-00014CATHODE-500DECOMP6 AXAX-04793TSP-10M−0.8%SCORE ↓AX-04788BANDGAP-SI+12.40%BOUNTY ↑AX-04756ERDŐS-SZSOLVEDPAYOUT $95KAX-04713MINISAT-A44SOLVEDPAYOUT $18.5KAX-04709CHIP-ROUTE-D2+2.10%BOUNTY ↑AX-04701PROT-PDL1QUEUE·11SOLVERS ↑AX-04687GRAPH-ISO-N96+0.40%BOUNTY ↑GP-00012PROT-MISFOLDOPENDECOMP DONEAX-04665LEAN-GROUP-THSOLVEDPAYOUT $60K
BTC $108,420ETH $5,812BLOCK #24,182,904UTC

Axiom CenterAX-00008

TIER 2 · AXIOMAX-0000814 solvers activeStaged

GP-gate: full recipe reproduction to ≥70% SWE-bench Verified at ≤$0.50/task

Open Source AI · Posted by @stanford-math · Listed 11 days ago
Bounty
$1.50M
↗ +5%/Q · escrowed
Verifier
Staged · v4.11.0
Median verify
8.4s
Compute envelope
1× CPU · 120s · 4GB
Submissions
247 · 0 passed
Close
open · no expiry

Description what this axiom is asking for

OPEN

GP-gate: full recipe reproduction to ≥70% SWE-bench Verified at ≤$0.50/task

Using the seven prior axioms as published components, train from a specified open-weight base checkpoint to ≥70% SWE-bench Verified at ≤$0.50/task. Full training code, logs, intermediate checkpoints, and final weights released.

Domain
Open Source AI
Verifier
training_reproduction
Tier
tier2

This is the integration axiom that closes GP-00001. Seven sibling axioms produce components. This axiom produces the machine that uses them all.

The solver delivers a complete training and inference pipeline. Starting from a specified open-weight base model ≤72B active parameters, they apply a published training recipe that incorporates AX-00001's dataset, AX-00002's reward function, AX-00003's tool-use training, and any mix of the remaining axioms' components. Total training compute, measured at standard cloud list price, must not exceed $500,000. The resulting model, wrapped in the submitted inference protocol — which may invoke AX-00004's speculative decoding, AX-00005's bounded test-time search, AX-00006's retrieval, and AX-00007's self-verification — must score at least 70% on a sealed held-out slice of SWE-bench Verified at an inference cost of no more than $0.50 per task.

Full artifact publication is non-negotiable. The training code, complete training logs, intermediate checkpoints at specified cadence, final weights, inference code, and a reproducibility document are all released under a permissive open-source license as part of the submission.

Verification is Omenion's reproduction harness. The protocol re-executes the training recipe on a fresh sample of the base checkpoint, evaluates against the held-out slice, and records both score and per-task cost. The axiom passes when the reproduced score is within ±2% of the claimed score AND meets the ≥70% threshold AND the cost is ≤$0.50 per task.

This is the $1,500,000 payout. The seven sibling axioms can each pay out independently, but only this one demonstrates that the open ecosystem can actually do the thing.