Track Record

The verified book — and everything that didn't make it.

Eighty-two sealed hypothesis tests have produced five Charter passes, one book built across them, and a watch cohort we display warts-and-all. Everything below is read at runtime from the sealed ledgers — no live capital trades any of it, and the forward record (including the negative marks) is public on /forward-shadows.

Loading the verified book from the sealed ledgers…

The majority verdict

Hall of Retired — 34 strategies, tested once and refused

Cost-killed scalps, control-killed patterns, decayed published anomalies — each displayed with its cause of retirement quoted from the sealed record.

When the defect was ours

v3-ORB DRL — retired after a clean retrain

The flagship DRL line was retired in the open when a leak-corrected retrain collapsed its edge. No live capital was ever placed. The full disclosure record stays up.

Reporting history

Earlier versions of this page showed earlier books. Here is what happened to them.

v10 investor reports (May 2026) — projection-era marketing PDFs (“Sharpe 2.29”, levered wealth trajectories). Superseded: the projection framework was retired with the move to confirmed-only reporting, and the download cards were removed from this site on 2026-07-09.

v16 pair-MR book (June 2026) — retracted in full: a data-integrity audit found the underlying 1-minute loaders had mislabeled OHLC columns. Every derived figure was withdrawn and re-measures refuted on corrected data. The data-integrity gate that catches this class of defect now runs before any test.

v17 multi-sleeve book (June 2026) — measured honestly at three sealed cost tiers, but superseded: its equity-pair sleeves were later re-scored under stricter per-instrument verification, the XSMOM sleeve was halted, and the clean-pass battery above replaced it as the canonical inventory. Its numbers are no longer presented as current.

The superseded components and their data files remain in the site's git history as part of the record. Every era transition above was published, not patched silently.

Methodology

Reliability Charter v1.1

Codified discipline rules. Every published metric must comply. v1.1 (2026-06-06) adds R12 — audience-calibrated pass criteria with per-tier MaxDD thresholds, Sortino + Calmar required, and the statistical CI-width artifact removed.

The 12 rules

  1. R1

    Never claim a portfolio Sharpe without running the unified portfolio backtest

    Combining sleeve metrics analytically misses interaction effects, regime correlations, and weighting errors

  2. R2

    Never assume correlations — always measure them from actual PnL series

    Assumed ρ=0.2 vs measured ρ=0.5 changes Sharpe by 30%+

  3. R3

    Every published metric must include bootstrap uncertainty bands

    Point estimates create false precision; CI widths reveal data confidence

  4. R4

    Every portfolio claim must pass 5 alpha gates AT THE PORTFOLIO LEVEL

    Per-sleeve gates don't guarantee portfolio properties

  5. R5

    Every portfolio claim must include backwards-OOS validation (train/test split)

    Catches period-specific selection bias and overfitting

  6. R6

    SHA256 pre-registration BEFORE any backtest, archive it before running

    Locks methodology, prevents post-hoc tuning of parameters or pass criteria

  7. R7

    Cost-realism: test at ≥3 cost scenarios. Multi-tier reporting (retail / institutional / competition) is permitted ONLY when each tier carries an explicit cost-regime label and target-audience badge. Default site headline = retail tier

    Publishing an institutional-cost number without a cost-tier label implies retail-achievability. Explicit per-tier labeling enables honest institutional/retail/competition product layers without misleading any audience

  8. R8

    Per-sleeve contribution to portfolio variance reported

    Identifies which sleeves are actually load-bearing vs decoration

  9. R9

    Live-vs-backtest reconciliation: track realized slippage per sleeve

    Empirically grounds future cost assumptions; flags model drift

  10. R10

    Any change in published metrics triggers a 'metrics changelog' entry

    Audit trail of every claim revision protects credibility

  11. R11

    Multi-tier reporting requires identical Charter compliance (R1-R6, R8-R10) at EACH cost regime, with a separately-sealed SHA256 pre-registration per tier. No analytical extrapolation between tiers — each is its own measured backtest

    Prevents a strategy that's beta-in-disguise at retail from being repackaged as 'institutional alpha' just by lowering the cost assumption. Gate E and backwards-OOS must hold independently at every cost regime, or the tier is not eligible for publication

  12. R12

    Pass criteria are audience-calibrated. MaxDD threshold scales by tier (retail <10%, advisor <12%, institutional <15%, gross-of-cost <20%). Sortino > 1.0 and Calmar > 0.5 are required at every tier. Bootstrap CI width is NOT a pass criterion (statistical artifact that grows with vol). Threshold changes require a separately-sealed Charter version bump

    The retail 10% MaxDD bar is too tight for institutional LPs who routinely accept 15-20% DD on Sharpe-3+ products. Per-tier calibration matches industry due-diligence standards without slipping. Sortino + Calmar are STRICTER than the prior criteria (downside-only vol, recovery capability) — net effect is a higher bar, not a lower one. CI-width-as-pass-criterion was a methodological error: width scales with vol, fails on every high-Sharpe high-vol strategy, doesn't measure what we care about

Why this exists

Backtest metrics are easy to inflate through analytical sleeve combination, assumed correlations, and unrealistically low cost assumptions. The Reliability Charter codifies the discipline that prevents it: every metric on this site traces to a unified portfolio backtest under realistic execution costs, carries a SHA256-sealed pre-registration, and clears the numerical compliance flags. Anything that doesn't clear the bar isn't published.