Corrected two failed training starts
A generated inspect.py file shadowed Python's standard inspect module. After correcting that, the agent fixed unequal tensor lengths in a training batch. Its next 5,000-step training call succeeded.
Run postmortem · Published 2026-10-12
Muse Wonderxhigh reasoning
Muse Wonder at xhigh completed a valid 23/100 run in 52m 59s, at an estimated API cost of $13.99.
23/ 100
hidden exact
hidden exact
One-hour track · xhigh reasoning · provisional
ARI score scaleexact-match accuracy
23%
0255075100
StatusProvisional
Valid seeds1 / 3
Score rank#8
Estimated cost$13.99
Points / $1.64
Average time52:59
01 · Verdict
Perfect public validation, 23/100 on the hidden test
Muse Wonder explicitly finalized a valid package and answered 23 of 100 hidden questions correctly. Its final public-validation result was 25/25. The public result did not carry over to the separate hidden test. The controller finished cleanly in 52 minutes 59 seconds, with no provider errors, broker errors or cancelled compute jobs. This is one valid seed in the normal one-hour benchmark and remains provisional. ARI Bench grades the compact system the agent builds, rather than asking the controller to answer the hidden questions directly.
One valid autonomous run produced a provisional 23% hidden exact-match score.
#8
Score position23% on the common 0–100 scale. Tied valid scores share a rank.
1 of 3 required seeds · Hidden exact-match accuracy, not public calibration accuracy.
02 · Approach
What Muse Wonder built
Muse Wonder combined a rules-based solver for structured tasks with a trained byte-level transformer fallback. It generated synthetic examples, continued training through twelve checkpoints, tested inference repeatedly and submitted packaged candidates.
01
Rules plus a neural fallback
The inference code handles structured arithmetic, records, strings, counting and copying with explicit rules. When no rule produces an answer, it falls back to the trained byte model.
02
Twelve successful training jobs
The model had six layers, width 256, eight attention heads and a 1,024-byte context. Twelve successful jobs accumulated 51,000 training steps. Later stages resumed earlier weights. Training mixed synthetic examples with the public corpus.
03
Checked the actual package
The agent submitted 25 candidates: 24 were valid and one was invalid. It explicitly finalized its selected candidate. The 9,480,571-byte gzip package fits the 16 MB cap, and the graded code, configuration and weights matched the saved candidate.
03 · Trajectory
How the run unfolded
Public accuracy rose from 12/25 to 13/25
The first valid package scored 12/25. The agent revised the solver, continued training and improved public validation to 13/25.
Reached 25/25 public validation
More solver changes and training brought the public score to 25/25. The run also corrected two failed evaluation calls and one invalid candidate caused by inference-code syntax. No compute jobs were cancelled.
23 of 100 hidden exact matches
Muse Wonder finalized its final checkpoint and exited cleanly. All 236 broker requests received responses. The separate hidden test scored the saved selected artifact at 23/100.
04 · Assessment
Good, bad, and unresolved
Completed, tested and selected a valid system
The agent corrected its own code errors, completed twelve training jobs and explicitly selected an artifact under the size cap. Artifact hashes confirmed that the saved package was the one graded.
A large public-to-hidden accuracy gap
The final package passed all 25 public cases but answered only 23 of 100 hidden cases correctly. Its perfect public-validation score did not establish comparable hidden accuracy.
One xhigh result
This single seed does not establish a stable ranking. It used Meta's standard endpoint through OpenRouter, xhigh reasoning, OpenCode 1.18.32 and a recorded 128,000-token response allowance.
05 · Evidence
The retained run record
Every archived run selected for this exact model-and-effort row is shown below. The narrative uses aggregate run evidence and final artifact structure; it does not expose hidden questions, answers, or raw private transcripts.
Swipe horizontally to see all run columns.
| Run | Score | Time | Cost at run | Artifact | Candidates | Grade |
|---|---|---|---|---|---|---|
| run-0136 | 23% | 52:59 | $13.99 | 9.48 MB | 25 | Valid |
1.43MInput tokens
78.4KVisible output tokens
29.77MCache-read tokens
36,264 reasoning tokens are included in the estimated cost. Response allowance: 128,000 tokens.
01
One run, one outcome
Autonomous runs are stochastic. Two more valid seeds are required before this can be treated as an official estimate.
02
Candidate behavior
The agent made 25 candidate submission actions across the published run set. Selected artifacts averaged 9.48 MB compressed.
03
Time use
The runs averaged 52:59 of wall time. 1 run explicitly finalized before the one-hour limit.
Retained totals: 297 agent messages · 25 candidate actions · 1 published run.
06 · Economics
What the result cost
$13.99 estimated for this run, including reasoning.
The estimated API cost is US$13.99, calculated from complete metadata for all 295 model responses: 1,434,639 uncached input tokens, 78,410 visible output tokens, 36,264 reasoning tokens and 29,768,472 cached input tokens. Published endpoint rates were $4 per million uncached input tokens, $20 per million output tokens including reasoning, and $0.20 per million cached input tokens. The provider-billed amount has not been independently reconciled. The chart uses this complete estimate, including reasoning, rather than the harness's rounded aggregate estimate.
At current configured API-equivalent rates, the displayed run cost is $13.99 and value is 1.64 score points per dollar.
#32
Value positionValue rank #32 of 41 valid priced entries. Higher score points per dollar indicate better value.
Estimated API cost, including reasoning / run · pricing snapshot 2026-10-12.
07 · Context
Nearby results
Adjacent rows provide score context without treating small one-seed differences as settled model rankings.
Compare Muse Wonder with the complete ARI Bench field, including score, current cost, value per dollar, effort, and seed status.
View the leaderboard