Frozen first
The ilXyr registration merged before promotion bytes opened. All models were hashed before blinding and prediction.
EXP-008 · completed 2026-08-31
The shortcut was real.
The harder bridge did not pass.
The unchanged sparse integer class head fit every training seed exactly, then failed to transfer across held-out sentence templates with balanced distractors.
Shared-task no-go
Minimum exact rate425,000 ppm
Worst seed-class0 ppm
Seed agreement531 / 1,000
Decision
Close the fixed sparse class-head path at this transfer boundary.
Every input used the same prefix. Classes and wrong-operation distractors were balanced, literal labels were absent, and all 20 promotion template families were held out from the 30 training families.
The data hashes, source commits, model form, four epochs, three seeds, three gates, forecasts, and model-first firewall merged before promotion opened.
Measured result
Some signal, no stable transfer.
| Seed | Exact selections | Exact rate | Overall gate |
|---|---|---|---|
| 1 | 425 / 1,000 | 425,000 ppm | miss |
| 2 | 430 / 1,000 | 430,000 ppm | miss |
| 3 | 533 / 1,000 | 533,000 ppm | miss |
The minimum result remained 225,000 ppm above the exactly balanced prefix-only baseline. That is evidence of partial learning, but it is far below the 900,000 ppm overall gate. The per-class and agreement gates also failed.
Discriminating task
The successor changed the information surface, not the threshold: common prefix, balanced decoys, and held-out templates.
Public negative evidence
Models, traces, blinded inputs, predictions, exact checks, and a complete manifest are published in NSRL.
Next boundary
Change abstraction; do not tune the opened test.
- The no-go covers this fixed hashed feature map, perceptron objective, and four-epoch rule.
- It does not reject all sparse or integer models, arithmetic generation, or the wider Solomon lineage.
- A successor needs fresh evaluation templates and a representation- or objective-level change.
- A read-only representation-sufficiency diagnostic should come before more training.
Public evidence
The complete no-go is hash-bound and independently checkable.
The result SHA-256 is c3b826685fc004fbc9b69a3a8ee539987522c1d3bf61e6e8d7c251147b394900. The scientific run used NSRL commit fa60312f93ec859e42a57f383f5972b8df4cf807; the published evidence is at merge commit df244eb414e68ad97c7bcce228b6129522e9a969.
Read the machine-readable evidence →
Inspect the frozen NSRL bundle →