Forward prediction of early hash outcomes
This experiment tested whether early SHA-256-derived features could rank candidates before completing the registered evaluation. Validation AUC was approximately 0.5025 and held-out test AUC approximately 0.4998, with lift near 0.978. The hypothesis was closed as negative.
- Published
- 2026-08-14
- Updated
- 2026-08-14
- Authors
- BTC PoW Lab
- Replication
- Validation and held-out test completed
Local structure, statistical signal and operational advantage are evaluated separately.
Question
Can early-computation features distinguish later favorable candidates better than chance?
Hypothesis
A stable early feature may provide out-of-sample ranking power.
Why this matters
A real early discriminator could justify conditional work; a failed model prevents wasted engineering.
Method
Train on the discovery partition, freeze the model, then evaluate validation and untouched test partitions.
Experimental setup
- Baseline
- Random discrimination: AUC 0.5 and lift 1.0.
- Setup
- Separated discovery, validation and held-out test cohorts.
- Hardware
- Classical analysis pipeline.
- Dataset
- Frozen feature matrices and split manifests; release review pending.
Results
- Validation AUC ≈ 0.5025.
- Test AUC ≈ 0.4998.
- Lift ≈ 0.978.
Scientific analysis
PRED-01 asks whether information available before the registered full evaluation can rank later outcomes. The model was trained only on discovery data, frozen, and then evaluated on validation and untouched test partitions to prevent adaptive leakage.
Area under the ROC curve has a probabilistic interpretation: for a randomly drawn positive and negative example, AUC is the probability that the model ranks the positive higher, with half credit for ties. Random ranking has expectation 0.5.
Validation AUC 0.5025 did not reproduce as useful discrimination: held-out AUC was 0.4998 and lift was 0.978. These values are operationally indistinguishable from chance for the registered decision rule, so no conditional-work engine was opened.
Mathematical formulation
AUC = P(s⁺ > s⁻) + ½P(s⁺ = s⁻)AUC measures ranking, not calibration and not mining throughput.
AUCtest − 0.5 ≈ −0.0002The untouched test result is essentially on the random-discrimination null.
lift = precisionselected / prevalence = 0.978The selected region contained slightly less target incidence than a random selection of equal size.
Statistical reading
Discovery performance is not evidence after feature and model selection; the untouched test partition is the decision-bearing result.
AUC near 0.5 can coexist with small local fluctuations. Promoting a post-hoc threshold would require a new preregistration and a new held-out cohort.
The negative conclusion is scoped: it rejects the registered feature/model family, not every conceivable early SHA-256 feature.
Validity and scope
- Split integrity is the main defense against leakage.
- The model did not meet a predictive or operational opening gate.
- No statement is made about unregistered nonlinear models or different labels.
Methodological references
Interpretation
The results are compatible with random discrimination. No predictive advantage was demonstrated.
Limitations
- The conclusion applies to the registered feature family and model.
- It does not prove that every possible early feature is uninformative.
Reproduction
Use the same split discipline and report untouched test performance; discovery-set performance alone is insufficient.
Artifacts
Model card and split manifest are queued for publication review.
Experimental content is provided for research and educational use, without warranty. Validate independently before relying on it. Read the full disclaimer.