Episode 002 · 32 min
Glass Ledger: publish before kickoff, score on CLV
Why pre-kickoff hash-chained picks beat self-attested track records, how recompute makes 'audited' literal, and the honest ceiling on full-slate rates.
What shipped
Proof and ledger surfaces that refuse fabricated ROI language. Calibration report gates. Public copy scanners for banned performance claims.
What broke
Industry habit of leading with win-rate without coverage. We refuse that UI. Variance-death marketing is not a product strategy.
What the model learned
Placebo time-shuffle must drive CLV to ~0 or the pipeline leaks. Market-blend test: if β on model logits includes 0, fire nothing.
What GSE passed on
Backfilled 'track records.' In-sample accuracy as marketing. Confidence-gated chalk at negative edge.
Partner / tool spotlight
Pinnacle close as internal CLV anchor via licensed odds snapshots; public recompute path as the moat competitors cannot fake.
Takeaways
- Start the honest ledger clock early — duration is the uncopyable asset.
- Lead with calibration diagrams, not cherry-picked subsets.
- Props supply volume for selective ~high-accuracy subsets; full-slate 60% claims are fiction.