An open research book of systematic investing techniques: every idea preregistered, backtested against real costs, and either killed in public or promoted to live monitoring.https://researchbook.dev/enhttps://researchbook.dev/writing/the-common-clock-replay/https://researchbook.dev/writing/the-common-clock-replay/Two strategies developed separately each carry their own clocks — signal time, fill time, mark time. Combine their equity curves as-is and the portfolio math silently rewards the mismatch: correlation biases low, overlap risk hides, and the book looks better-diversified than it is. The fix is mechanical: freeze both decision streams, rebuild every sleeve at one shared clock against the displayed order book, and only then let the portfolio numbers speak — with the incompleteness, cost surface, and fragility bands published next to them.Sat, 25 Jul 2026 00:00:00 GMTmethodsexecutionportfolioreplayhttps://researchbook.dev/writing/anti-correlation-hedges-not-stacks/https://researchbook.dev/writing/anti-correlation-hedges-not-stacks/PR #802's vol_weighted_mean_revert composite was predicted to stack — pairwise correlation -0.42, both score-stage, no sign-vs-rank conflict. All four prior iteration conditions satisfied. Instead the composite TIED its better parent on both single-signal and dual-signal data. The mechanism: anti-correlated parents HEDGE each other; the sum captures the average, bounded by the better parent. This is the rule's 5th iteration: negative correlation predicts hedging, not stacking.Fri, 22 May 2026 00:00:00 GMTcompositionresearchpairwisehttps://researchbook.dev/writing/composite-strategies-can-interfere/https://researchbook.dev/writing/composite-strategies-can-interfere/Two top-of-leaderboard strategies — three_clock_momentum (highest min-Sharpe) and vol_regime_filter (highest mean) — got composed into one. Expectation: stack. Result: interferes. The composite's mean sits between the parents, its min goes below both, and the variance contribution from the gate is unchanged. On synthetic data with one signal source, both noise-reducers smooth the same noise; you can't double-count.Fri, 22 May 2026 00:00:00 GMTstrategyresearchcompositionhttps://researchbook.dev/writing/designing-an-event-aware-strategy/https://researchbook.dev/writing/designing-an-event-aware-strategy/A working note that pulls the event-clock research into one recipe. Six steps from 'I have an event hypothesis' to 'I have a strategy whose cost I can defend.' The four prior notes are the why; this is the how.Fri, 22 May 2026 00:00:00 GMTeventsresearchchecklisthttps://researchbook.dev/writing/disconfirmed-the-transition-gate-fix/https://researchbook.dev/writing/disconfirmed-the-transition-gate-fix/PR #717 named a candidate fix for vol_regime_filter's underperformance on clustered-vol data: gate on vol CHANGE instead of vol LEVEL. PR #721 ran the experiment. Result: the transition gate underperforms the level gate on BOTH modes by 0.2 Sharpe. The 'gate on change' fix doesn't recover the regime-filter performance. Both interpretations matter.Fri, 22 May 2026 00:00:00 GMTresearchregimenegative-resulthttps://researchbook.dev/writing/dont-pay-for-caution-you-cant-justify/https://researchbook.dev/writing/dont-pay-for-caution-you-cant-justify/A risk-reduction gate looks free until you measure what it costs. Most don't survive the measurement. The ones that do clear two specific bars — and if you can't say which bar you cleared, you've added ceremony, not safety.Fri, 22 May 2026 00:00:00 GMTriskevaluationresearchhttps://researchbook.dev/writing/drift-vs-reversal-as-counterfactual/https://researchbook.dev/writing/drift-vs-reversal-as-counterfactual/Running the drift strategy and its sign-flipped twin on the same data, in the same window, with the same lookback — that's not two strategies. It's a thermometer. The sign of `drift − reversal` is the answer to 'does this event lead to continuation or over-correction?'Fri, 22 May 2026 00:00:00 GMTeventsresearchcounterfactualhttps://researchbook.dev/writing/dual-signal-full-leaderboard-n100/https://researchbook.dev/writing/dual-signal-full-leaderboard-n100/PR #790 documented the partial dual-signal result (composite stacks slightly). This is the full 27-arm leaderboard at N=100 with `--fomc-drift-bps 50 --mean-revert-bps 100`. Five findings: ts_momentum still leads; three composites converge again; multi-factor variants are LESS negative on dual-signal; vol_penalty flips negative; the mean_revert diagnostic confirms data shape.Fri, 22 May 2026 00:00:00 GMTresearchharnessdual-signalleaderboardhttps://researchbook.dev/writing/dual-signal-makes-composites-stack/https://researchbook.dev/writing/dual-signal-makes-composites-stack/Added a second alpha source to the harness data — per-bar mean-reversion independent of the FOMC drift. The composite arm that INTERFERED on single-signal data (PR #664) now STACKS on dual-signal: mean +1.190 above three_clock_momentum (+1.181) and vol_regime_filter (+1.125). Confirms PR #666's hypothesis: 'interferes' was a property of the synthetic having one alpha source, not of the strategies themselves.Fri, 22 May 2026 00:00:00 GMTstrategyresearchcompositionhttps://researchbook.dev/writing/dual-signal-n100-composites-stack/https://researchbook.dev/writing/dual-signal-n100-composites-stack/PR #777's pairwise-rule final-form note ended with: 5/5 direction predictions correct, 0/5 magnitude — no composite cleanly stacked at N=100 on single-signal data. The hypothesis: synthetic only has one alpha source. Re-ran the harness at N=100 with `--mean-revert-bps 100` (dual-signal mode). Result: `three_clock_vol_weighted` is now ABOVE both parents (+0.584 vs +0.573 + +0.515). The stack is small but directional. Confirms the dual-signal hypothesis.Fri, 22 May 2026 00:00:00 GMTcompositionresearchdual-signalhttps://researchbook.dev/writing/event-aware-needs-signal-concentration/https://researchbook.dev/writing/event-aware-needs-signal-concentration/Adding a directional component to a synthetic FOMC shock made the baseline momentum strategy capture the alpha. The event-aware drift wrapper didn't help. The reason is geometry — if the signal extends over multiple days, a 4-hour capture window is the wrong shape.Fri, 22 May 2026 00:00:00 GMTeventsresearchevaluationhttps://researchbook.dev/writing/event-gates-cost-when-the-event-has-edge/https://researchbook.dev/writing/event-gates-cost-when-the-event-has-edge/Ran the 15-arm FOMC compare at a config where the synthetic has +50 bps drift baked into event days. Long-only wins (Sharpe 1.88), baseline trails at 1.24, the blackout and damping arms lose -0.11 and -0.14 Sharpe to baseline, the drift arm crashes to 0.64. The gates aren't broken — they're throwing away the exact exposure the synthetic data rewards.Fri, 22 May 2026 00:00:00 GMTstrategyresearcheventsevaluationhttps://researchbook.dev/writing/fifty-seeds-full-leaderboard/https://researchbook.dev/writing/fifty-seeds-full-leaderboard/The 50-seed run gave us full per-arm stats across all 24 harness arms. Three surprises: ts_momentum is the leader by mean (+0.910), not baseline (+0.767). The four-factor and three-factor variants are all NEGATIVE. The portfolio_vol_gate is mid-pack on mean — its true value is the smaller stdev, not the higher mean. This note is the full ranking and what each one teaches.Fri, 22 May 2026 00:00:00 GMTresearchharnessleaderboardhttps://researchbook.dev/writing/fifty-seeds-reveal-the-tie/https://researchbook.dev/writing/fifty-seeds-reveal-the-tie/At N=10, three_clock_vol_weighted beat baseline by +0.10 Sharpe — directionally encouraging but within the stdev (1.22). At N=50, the composite ties baseline at +0.765 — within rounding of baseline's +0.767. The directional signal didn't survive the higher seed count, exactly as the session-summary v2 note's caveat predicted.Fri, 22 May 2026 00:00:00 GMTresearchharnesscompositionhttps://researchbook.dev/writing/higher-moments-add-noise-faster-than-signal/https://researchbook.dev/writing/higher-moments-add-noise-faster-than-signal/Under the Gaussian approximation used here, the Nth moment has finite-sample standard error roughly proportional to √(N!/n). Adding a kurtosis term to a momentum + reversion + skew score buys (in theory) more explanatory power; in practice it buys mostly more variance. The strategy catalog's `kurtosis_weight = −0.5` default is the empirical compensation.Fri, 22 May 2026 00:00:00 GMTmomentsresearchevaluationhttps://researchbook.dev/writing/how-to-use-the-comparison-harness/https://researchbook.dev/writing/how-to-use-the-comparison-harness/The 19-arm comparison harness has grown through this session into a full toolchain — 19 strategies, two synthetic signal modes, five CLI flags, one analysis script. This note is the operator's how-to: a step-by-step workflow for asking 'does this new strategy work?' and reading the answer.Fri, 22 May 2026 00:00:00 GMTharnessresearchworkflowhttps://researchbook.dev/writing/hundred-seeds-confirms-and-converges/https://researchbook.dev/writing/hundred-seeds-confirms-and-converges/Re-ran the harness at N=100 to refine the 50-seed leaderboard. ts_momentum's lead over baseline holds (+0.132, 82% hit rate). The three three_clock variants converge to exactly +0.798 — what looked like distinct arms at N=10 are statistically identical at N=100. baseline and spread_filter are also indistinguishable. Two findings: ts_momentum is robust; many arms are duplicates.Fri, 22 May 2026 00:00:00 GMTresearchharnessleaderboardhttps://researchbook.dev/writing/intervention-point-rule-confirmed/https://researchbook.dev/writing/intervention-point-rule-confirmed/Three per-symbol vol interventions failed. PR #731's discipline rule said: move to a different intervention point. PR #732 shipped the cross-symbol portfolio-scale gate. Result: it recovers 75-80% of the Sharpe gap to baseline that the per-symbol variants sacrificed. Same alpha source, same data, different intervention point — the rule held.Fri, 22 May 2026 00:00:00 GMTresearchvolinterventionhttps://researchbook.dev/writing/long-only-buys-asymmetric-exposure/https://researchbook.dev/writing/long-only-buys-asymmetric-exposure/Added a long-only momentum arm to the 14-arm harness. Mean Sharpe was −0.196 vs the long-short baseline. The interesting number wasn't the mean — it was the stdev: 1.487 across 5 seeds, more than double baseline's 0.581. Long-only doesn't just give up the short leg's contribution. It gives up the dollar-neutral diversification that flattens per-seed dispersion.Fri, 22 May 2026 00:00:00 GMTstrategyresearchdesignhttps://researchbook.dev/writing/pairwise-correlation-predicts-composition/https://researchbook.dev/writing/pairwise-correlation-predicts-composition/Closes the loop on the session-summary note's `pairwise correlation across the 19 arms` follow-up. The matrix shows ts_momentum and three_clock_momentum at +0.34 correlation across 10 seeds — different signals — and three_clock_momentum and blackout at +0.93 — same signal in different framings. The empirical correlation tells you whether composing two arms will stack OR interfere, before you've built the composite arm.Fri, 22 May 2026 00:00:00 GMTcompositionresearchharnesshttps://researchbook.dev/writing/perfect-correlation-explains-the-interference/https://researchbook.dev/writing/perfect-correlation-explains-the-interference/The composite `three_clock_portfolio_vol` underperformed both parents in PR #740. The pairwise correlation analysis (10 seeds) shows why: three_clock_momentum and three_clock_portfolio_vol correlate at exactly +1.00 across seeds. The composite IS the parent — a vol-gate added on top of a momentum signal that's already vol-gated doesn't catch anything new.Fri, 22 May 2026 00:00:00 GMTcompositioncorrelationresearchhttps://researchbook.dev/writing/property-tests-catch-cross-strategy-bugs/https://researchbook.dev/writing/property-tests-catch-cross-strategy-bugs/Three event-aware strategies, each with its own cold-inner test, each shipped under code-review. Two had the citation contract right. One didn't. The bug surfaced only when a single property-based test parametrised over all three at once. Per-strategy tests are necessary but not sufficient — the contract has to be expressed at the family level.Fri, 22 May 2026 00:00:00 GMTtestingresearchcontractshttps://researchbook.dev/writing/redundant-vs-multiplicative-composition/https://researchbook.dev/writing/redundant-vs-multiplicative-composition/Two composites at the same +0.70 pairwise correlation produced opposite outcomes. mom_ma_composite (same quantity at different smoothings) stacked above baseline. vw_tc_composite (different quantities at the same granularity) interfered below both parents. The 6th iteration of the pairwise rule codifies the distinction. This note documents the side-by-side experiment that the rule rests on.Fri, 22 May 2026 00:00:00 GMTcompositionresearchpairwisehttps://researchbook.dev/writing/sharpe-is-scale-invariant/https://researchbook.dev/writing/sharpe-is-scale-invariant/Added an `equal_risk_long_only` arm to the harness — same strategy, gross_leverage scaled to 0.71. Sharpe was identical to the 1.0×-gross arm. Dollar PnL and dollar drawdown scaled 0.71×. The arm earned its place by making the invariant visible: gross-leverage rescaling moves dollars, not ratios.Fri, 22 May 2026 00:00:00 GMTsharperesearchmeasurementhttps://researchbook.dev/writing/sign-vs-rank-composition-conflict/https://researchbook.dev/writing/sign-vs-rank-composition-conflict/PR #768's ts_filtered composite was predicted to stack (low pairwise correlation, different mechanisms). Instead it interferes — composite +0.527 vs leader ts_momentum's +0.995. The mechanism: ts_momentum's edge is sign-based (trade each active symbol in its momentum direction); cross-sectional ranking's edge is rank-based (short the lower-ranked even if positive). Composing them OPPOSES the parents' edges, not aligns them. This is a third refinement to the pairwise rule.Fri, 22 May 2026 00:00:00 GMTcompositionresearchpairwisehttps://researchbook.dev/writing/six-regression-tests-pin-the-findings/https://researchbook.dev/writing/six-regression-tests-pin-the-findings/The session's 100+ PRs produced 12+ discipline rules and 60+ research notes. Six of those findings now have CI-enforced regression tests pinning their empirical predictions against future drift. This note documents what each test catches and why the tests-vs-research-notes ratio matters.Fri, 22 May 2026 00:00:00 GMTtestingdisciplineregressionhttps://researchbook.dev/writing/strategy-shape-beats-factor-count/https://researchbook.dev/writing/strategy-shape-beats-factor-count/The 11-arm harness has a four-factor strategy that combines momentum + reversion + skew + kurtosis. It also has a one-factor strategy that just adds a 2% threshold on absolute return. The threshold-gated one-factor wins on mean AND on min Sharpe across seeds. The threshold matters more than three extra factors.Fri, 22 May 2026 00:00:00 GMTstrategyresearchdesignhttps://researchbook.dev/writing/synthetic-data-shows-what-you-were-solving-for/https://researchbook.dev/writing/synthetic-data-shows-what-you-were-solving-for/Shipped a chop-filter strategy designed to gate cross-sectional momentum during noise periods. Ran it on the synthetic harness. Got numbers identical to baseline — the threshold never gated because the synthetic data has no chop. That's not a failure of the strategy. It's a feature of synthetic data: when a wrapper has no effect, you've learned which problem the wrapper solves and which problem your data has.Fri, 22 May 2026 00:00:00 GMTsyntheticresearchevaluationhttps://researchbook.dev/writing/the-baseline-arm-you-forgot/https://researchbook.dev/writing/the-baseline-arm-you-forgot/An A/B with two arms tells you whether your gate moves the number. A run with the gate's inverse as a third arm tells you whether the input data has anything for the gate to defend against. Most A/Bs are missing the third arm and don't know it.Fri, 22 May 2026 00:00:00 GMTevaluationresearcheventshttps://researchbook.dev/writing/the-clustered-vol-finding-was-also-small-n/https://researchbook.dev/writing/the-clustered-vol-finding-was-also-small-n/PR #717 claimed vol_regime_filter underperforms baseline by 1.3 Sharpe on clustered-vol data. At N=3 seeds that was right. At N=30 seeds with seeds 1-30, the gap collapses to +0.001 — within noise. The vol-regime-filter doesn't break on clustered vol the way the PR #717 note claimed; the original finding was N=3 luck. Add another instance to the pattern: directional claims at N less than 20 don't survive higher seed counts.Fri, 22 May 2026 00:00:00 GMTresearchharnessregimesmall-nhttps://researchbook.dev/writing/the-event-clock-isnt-the-panel-clock/https://researchbook.dev/writing/the-event-clock-isnt-the-panel-clock/A research platform calibrated for day bars and minute bars handles event-driven trades awkwardly. Adding event-clock research doesn't need a new runtime — it needs one new column on the feature DAG and one new rule in the risk gate.Fri, 22 May 2026 00:00:00 GMTresearchplatformeventshttps://researchbook.dev/writing/the-first-robust-single-signal-stack/https://researchbook.dev/writing/the-first-robust-single-signal-stack/PR #815's mom_ma_composite (baseline raw return + MA-crossover) stacks above baseline by +0.071 Sharpe at N=100 single-signal. This is the first composite this session to stack ROBUSTLY at high seed count on single-signal data — earlier 'first stack' claims (PR #746) collapsed at N=50. The mechanism: same underlying momentum signal at two different smoothing scales, additive in the +0.70 'marginal' band of ADR-0062's rule.Fri, 22 May 2026 00:00:00 GMTcompositionresearchstackhttps://researchbook.dev/writing/the-full-pairwise-matrix/https://researchbook.dev/writing/the-full-pairwise-matrix/The pairwise rule has been the session's most-iterated discipline. This is the full correlation matrix across all 28 arms at N=100 single-signal — the empirical surface the rule operates on. Surfaces six observations: complete dormancy clusters, sign-flipped diagnostics, the lowest-correlation pair, and three composite candidates the matrix predicts.Fri, 22 May 2026 00:00:00 GMTresearchpairwiseharnessmatrixhttps://researchbook.dev/writing/the-pairwise-rule-final-form/https://researchbook.dev/writing/the-pairwise-rule-final-form/The pairwise-correlation discipline rule started simple (PR #710: low correlation predicts stack) and iterated four times across the session as composites were built and measured. This is the consolidated final form: four conditions all required for composite stacking. Includes the pre-build pre-test checklist, the decision tree for when each iteration of the rule applies, and the load-bearing failure modes.Fri, 22 May 2026 00:00:00 GMTcompositionresearchdisciplinehttps://researchbook.dev/writing/the-pairwise-rule-predicted-this-one/https://researchbook.dev/writing/the-pairwise-rule-predicted-this-one/Picked the most-decorrelated top-arm pair (+0.71 correlation), composed them at the score stage, and shipped the result. The composite beat both parents AND baseline — the first composite-arm result this session to do all three. The empirical confirmation closes the loop the matrix opened: low pairwise correlation predicts composite stacking.Fri, 22 May 2026 00:00:00 GMTcompositionresearchharnesshttps://researchbook.dev/writing/the-single-seed-lead-was-a-fluke/https://researchbook.dev/writing/the-single-seed-lead-was-a-fluke/PR-shipping the single-seed FOMC compare result said long_only was the leader at Sharpe 1.88. Re-running 5 seeds at the same config: long_only's mean Sharpe is 0.81 — below the 1.00 baseline — with stdev 1.49 and a range of −0.93 to 2.45. ts_momentum is the actual consistent leader at mean 1.27 across the sweep. The note's previous reading was the result you get when the noise lines up.Fri, 22 May 2026 00:00:00 GMTstrategyresearchevaluationmethodologyhttps://researchbook.dev/writing/the-strategy-catalog/https://researchbook.dev/writing/the-strategy-catalog/As of May 2026, alphakernel shipped twenty-nine registered live-trading strategies. Twenty-eight of them shared a single 17-line ranking helper, then differed only in their score function. The catalog is the working surface of the platform — what an operator can point a runner at, what an A/B harness can measure, what citation graphs read.Fri, 22 May 2026 00:00:00 GMTcatalogstrategiesplatformhttps://researchbook.dev/writing/three-clock-momentum-tops-the-harness/https://researchbook.dev/writing/three-clock-momentum-tops-the-harness/Of the 16 arms now in the comparison harness, the best mean Sharpe across 5 seeds (+1.429) came from XsThreeClockMomentumStrategy — a linear combination of 5-, 20-, and 60-bar momentum with default weights (-0.5, +1.0, +0.5). That beats single-window momentum (+1.001) by 0.43 Sharpe and single-window TSMom (+1.269) by 0.16. The result is consistent with the literature: composite-horizon momentum extracts more information than any single window cut.Fri, 22 May 2026 00:00:00 GMTstrategyresearchmomentumhttps://researchbook.dev/writing/three-vol-experiments-zero-wins/https://researchbook.dev/writing/three-vol-experiments-zero-wins/Across three different shapes — level filter, transition filter, score-stage continuous penalty — no per-symbol vol intervention beats baseline cross-sectional momentum on the harness. The score-stage penalty was the best of the three on clustered-vol data, but still below baseline. The meta-finding: cross-sectional momentum is robust enough that per-symbol vol interventions consistently cost Sharpe regardless of shape.Fri, 22 May 2026 00:00:00 GMTresearchvolnegative-resulthttps://researchbook.dev/writing/vol-cluster-breaks-the-regime-filter/https://researchbook.dev/writing/vol-cluster-breaks-the-regime-filter/vol_regime_filter was designed for vol-regime detection. Add vol clustering to the synthetic — the regime structure it's allegedly designed to detect — and its Sharpe drops by 1.3 below baseline. The reason isn't that the gate fails to fire; the gate fires CONSTANTLY. When vol clusters, the vol_5/vol_60 ratio is structurally elevated most of the time, so the gate drops most of the universe most of the time, leaving 2-3 names rotating through the ranking on a cluster-driven schedule that has nothing to do with momentum.Fri, 22 May 2026 00:00:00 GMTvolregimeresearchhttps://researchbook.dev/writing/vol-of-vol-distinguishes-regime-from-level/https://researchbook.dev/writing/vol-of-vol-distinguishes-regime-from-level/Two symbols sitting at 1% daily vol look identical to vol_20. One has been at 1% for sixty bars. The other just arrived at 1% from a 0.3% regime, with thirty bars of transition. Vol-of-vol — the 20-bar std of vol_20 — is what tells them apart, and it changes how a sizing rule should treat them.Fri, 22 May 2026 00:00:00 GMTvolatilityfeatureresearchhttps://researchbook.dev/writing/vol-ratio-as-cross-symbol-regime-gate/https://researchbook.dev/writing/vol-ratio-as-cross-symbol-regime-gate/Realised vol is a level. Vol-of-vol is its second derivative. The vol_5/vol_60 ratio is neither — it's a dimensionless regime classifier that works the same across a 1%-daily ETF and a 5%-daily small-cap. Why scale-invariance matters when classifying regime across heterogeneous universes.Fri, 22 May 2026 00:00:00 GMTvolatilityfeatureregimehttps://researchbook.dev/writing/vol-regime-filter-mean-vs-variance/https://researchbook.dev/writing/vol-regime-filter-mean-vs-variance/The vol-regime gate dropped the harness's top spot from three_clock_momentum (+1.429) to vol_regime_filter (+1.493) — a 4.5% mean Sharpe improvement at the cost of doubling the per-seed dispersion (0.672 → 1.383). When a gate concentrates exposure on a survivors-only universe, the mean might tick up but the worst case gets meaningfully worse.Fri, 22 May 2026 00:00:00 GMTstrategyresearchregimehttps://researchbook.dev/writing/what-19-arms-told-us/https://researchbook.dev/writing/what-19-arms-told-us/A session of incremental work grew the harness from 6 to 19 arms across two synthetic signal modes. The findings cluster into four claims about composition, regime gates, and what synthetic harnesses can and can't measure. This note is the index.Fri, 22 May 2026 00:00:00 GMTsummaryresearchharnesshttps://researchbook.dev/writing/what-24-arms-told-us/https://researchbook.dev/writing/what-24-arms-told-us/Five more arms shipped after the original `what-19-arms-told-us` summary. The new findings cluster into four additional claims about intervention points, composition stages, and pairwise correlation as a pre-test. This note supersedes the 19-arm summary as the top-of-stack index.Fri, 22 May 2026 00:00:00 GMTsummaryresearchharnesshttps://researchbook.dev/writing/what-27-arms-told-us/https://researchbook.dev/writing/what-27-arms-told-us/The v2 summary at 24 arms / 8 claims is now stale. The v3 covers 27 arms / 11 claims, including the small-N discipline corpus (two findings disconfirmed at higher N), the pairwise rule's four iterations and final form, and the dual-signal stacking confirmation. This note replaces v2 as the top-of-stack index for the session's research arc.Fri, 22 May 2026 00:00:00 GMTsummaryresearchharnesshttps://researchbook.dev/writing/what-29-arms-told-us/https://researchbook.dev/writing/what-29-arms-told-us/The v3 summary at 27 arms / 11 claims is now stale. The v4 covers 29 arms / 25 strategies (now 27) / 12 claims, including the pairwise rule's 5th iteration (anti-correlation hedges, doesn't stack), ADR-0062 codifying the rule, and four CI-enforced regression tests pinning load-bearing findings. This note replaces v3.Fri, 22 May 2026 00:00:00 GMTsummaryresearchharnesshttps://researchbook.dev/writing/why-ts-momentum-leads-at-n100/https://researchbook.dev/writing/why-ts-momentum-leads-at-n100/At N=100 seeds, ts_momentum (+0.995) beats baseline (+0.863) by +0.132 Sharpe. Hit rate 82% vs baseline's 79%. Not a small margin — 1.2 standard errors above baseline. The mechanism: ts_momentum's |return| > threshold entry rule closes time-series chop that cross-sectional ranking can't filter. Deepens PR #557's 'strategy shape beats factor count' rule into a sharper claim about the entry-rule axis.Fri, 22 May 2026 00:00:00 GMTresearchharnessts-momentumhttps://researchbook.dev/writing/ai-agents-as-platform-citizens/https://researchbook.dev/writing/ai-agents-as-platform-citizens/The platform decisions that turn a hallucinating LLM into something a research firm can safely point at a backtest engine. None of them are about the model.Thu, 21 May 2026 00:00:00 GMTaiplatformagentshttps://researchbook.dev/writing/audit-isnt-a-feature/https://researchbook.dev/writing/audit-isnt-a-feature/The moment compliance asks where a trade came from, you find out whether your platform was designed around audit or had it bolted on. The difference is small, structural, and decided years before the conversation.Thu, 21 May 2026 00:00:00 GMTplatformdesignaudithttps://researchbook.dev/writing/refusal-as-planning-hint/https://researchbook.dev/writing/refusal-as-planning-hint/A 500 tells an agent something went wrong. A typed refusal tells it what to do next. The cost of getting this right is a vocabulary; the payoff is an agent that doesn't bluff.Thu, 21 May 2026 00:00:00 GMTaiplatformagentshttps://researchbook.dev/writing/search-isnt-research/https://researchbook.dev/writing/search-isnt-research/Most discovery systems quietly become unbounded search loops the platform can't cost. The trick is to let the operator name the bound, not the algorithm — and to refuse runs whose spec exceeds it.Thu, 21 May 2026 00:00:00 GMTplatformresearchdiscoveryhttps://researchbook.dev/writing/the-citation-graph-as-substrate/https://researchbook.dev/writing/the-citation-graph-as-substrate/Once every decision is a row that names what it cites, four otherwise-separate features become one shape — and the platform you can build on top of that shape is meaningfully different from one you can't.Thu, 21 May 2026 00:00:00 GMTplatformdesigncitationhttps://researchbook.dev/writing/right-sizing-the-research-data-platform/https://researchbook.dev/writing/right-sizing-the-research-data-platform/Most platform debates skip the right question. It isn't 'what's best practice?' — it's 'which threshold have we crossed today?' Build the smallest architecture that covers the answer; upgrade exactly when the next threshold lands.Sun, 17 May 2026 00:00:00 GMTplatformjudgmentresearchhttps://researchbook.dev/writing/the-promotion-gate/https://researchbook.dev/writing/the-promotion-gate/Validation that runs as a dashboard tells you what went wrong after the model trained. Validation that runs as a gate makes it structurally impossible for a strategy to read what the platform hasn't vouched for. The difference shows up in PnL.Sun, 17 May 2026 00:00:00 GMTplatformvalidationresearchhttps://researchbook.dev/writing/the-deploy-contract-isnt-a-yaml-file/https://researchbook.dev/writing/the-deploy-contract-isnt-a-yaml-file/The most important artifact a research engineer ships is also the one that doesn't look like code. It looks like the conversation that happened the week before.Sun, 12 Apr 2026 00:00:00 GMTplatformprocessresearchhttps://researchbook.dev/writing/walk-forward-without-leakage/https://researchbook.dev/writing/walk-forward-without-leakage/Most leakage bugs don't look like leakage. They look like a model that's just good. Here's the small set of checks I run before I'll trust any backtest number.Wed, 18 Mar 2026 00:00:00 GMTevaluationmlresearchhttps://researchbook.dev/writing/mlops-for-quant-research/https://researchbook.dev/writing/mlops-for-quant-research/Web-MLOps wants to retrain on yesterday's data and ship to A/B. Quant-MLOps wants to defend bit-exact reproducibility of a model that traded a year ago. Same vocabulary, different platform.Sun, 22 Feb 2026 00:00:00 GMTmlopsplatformresearchhttps://researchbook.dev/writing/observability-for-alpha-pipelines/https://researchbook.dev/writing/observability-for-alpha-pipelines/If a model can be in production, you should be able to answer three questions about it without writing a query. Each question gets its own dashboard, and they don't share an owner.Fri, 30 Jan 2026 00:00:00 GMTobservabilitymlopsplatform