Snapshot integrity and coverage
12,043valid JSONL result rows
1,002complete matched 12-policy blocks
2partial source blocks at copy time
0correctness / solution / given-consistency failures
The journal contained 14 more committed rows than the copied PROGRESS.json, consistent with a live extract taken while the run continued. The append-only result journal is used as the row authority for this snapshot.
Live extract SHA3-256
ef4aa0db52043270e77747146393cc629f7c3d5cc9614d8a366877f2858f1c3c
Frozen arena SHA3-256
2f94d8e13763cf1a950504a1dc1e605e3019e5a83279b5b45bc811a0497ec1fa
ef4aa0db52043270e77747146393cc629f7c3d5cc9614d8a366877f2858f1c3c
Frozen arena SHA3-256
2f94d8e13763cf1a950504a1dc1e605e3019e5a83279b5b45bc811a0497ec1fa
First paired signal on the common root-stuck cohort
The comparisons below use only the 421 completed puzzles whose shared root logic status was STUCK_LOGIC. They are descriptive first-snapshot totals—not the preregistered final two-split statistical verdict.
| Comparison | Backtracks | First-snapshot result | Interpretation now |
|---|---|---|---|
| AI8 depth-2 vs MRV ascending | 356 vs 851 | 58.17% fewer · 249/156/16 W/T/L | Strong provisional system-level search signal. |
| Fixed residual vs matched sham | 409 vs 826 | 50.48% fewer · 252/141/28 W/T/L | Informed lookahead currently beats the sham control. |
| Fixed MDL vs fixed residual | 445 vs 409 | MDL currently 8.80% worse · path differs 9/421 | MDL-specific value is not supported by this partial snapshot. |
| Adaptive MDL vs adaptive residual | 378 vs 333 | MDL currently 13.51% worse · path differs 12/421 | The residual twin currently leads; final verdict remains open. |
| Depth-3 vs depth-2 | 276 vs 356 | 22.47% fewer | More depth currently improves search; total-cost accounting still matters. |
The arena is doing real science. It is already capable of producing an uncomfortable component result: the current residual controls outperform their MDL twins. That is evidence that the experiment can reject a favored mechanism rather than simply manufacture confirmation.
What cannot yet be concluded
No final claim is made here about pooled performance, split-B replication, bootstrap lower bounds, leave-one-out stability, MDL-specific promotion, DCC-specific promotion, arbitrary Sudoku, cross-domain transfer, AGI, consciousness or P=NP. The campaign must finish and be evaluated under its frozen contract.