Reproducible Verdict Kernel

An event study that cannot be recomputed is an anecdote with error bars. The Reproducible Verdict Kernel is the engine I built to make event-study claims adjudicable: feed it the events and the series, and it returns a verdict that anyone holding the same inputs can regenerate byte for byte.

The kernel exists — 5,037 lines of Rust, 77 test functions, zero dependencies. Zero-dependency is not an aesthetic choice: every library between your data and your conclusion is a version number that can silently change your answer. Here the entire inferential chain is in one auditable codebase, and determinism is a tested property, not a hope.

Every verdict the kernel issues is built to survive one blunt question: does that result recompute? Hand it the same inputs and the same binary and it either returns the verdict you were given, byte for byte, or it reports the divergence as a finding in its own right. A result that will not recompute has said something about itself before anyone gets to the argument about whether the effect is real.

What it measures

The kernel rules only on effects brought to it, and it generates no candidates of its own.

• Effect estimates with honest uncertainty. Event effects with confidence intervals from the stationary block bootstrap — resampling that respects the serial dependence in financial series instead of assuming it away — and HAC standard errors that survive autocorrelation and heteroskedasticity rather than pretending the residuals are polite.
• Multiplicity accounting. When you test a family of events or hypotheses, some will look significant by volume alone. The kernel applies Benjamini-Hochberg false-discovery-rate masking across the family, so the verdict sheet distinguishes effects that survive multiplicity from effects that merely got lucky in a crowd.
• The can-fail proof, built in. The kernel’s keystone self-test plants a known synthetic effect and demands the engine recover it — while simultaneously killing forty placebo effects it must reject. An engine that cannot demonstrate both powers, detection and refusal, has no business issuing verdicts. This test runs before I trust any output, including my own.
• Byte-reproducible verdicts. Same inputs, same binary, same verdict — down to the byte. A verdict you receive can be recomputed by you, by a reviewer, by a counterparty, and the recomputation either matches exactly or the discrepancy is itself a finding.

The verdicts on my own published event families — which effects survived, at what estimated magnitude, under which threshold — are published, and the answer is that none of them survived. Both entries in the results index are recorded KILLED, each against a registry denominator of eighteen at an FDR threshold of 0.05 over three walk-forward folds, and each carries the reason rather than the label alone: H-010’s zone-revisit reversion at −63.47 with an interval of [−146.40, 17.92], and S-001’s origination-level salience enrichment at −0.0423 with [−0.1269, 0.0416]. Both intervals span zero, which is why both are killed. Each ships with the sha256 of its own result document, so the recomputation is available rather than promised — those are the same two digests the kill ledger cites, and they agree.

A separate and larger family is on a resolving track and is reported here with its grade rather than its best number: fifteen registered predictions, tested exhaustively at 819 rotations, of which three are nominally significant and zero survive correction across the family. No edge has graduated. That figure is the one worth printing, because a kernel whose own candidates all fail its own correction is the only kind whose eventual survivor would mean anything.

What it does not do

The Reproducible Verdict Kernel does not find you an edge. The kernel adjudicates evidence for effects you claim; it does not generate candidates, and it happily — mechanically — returns “no effect survives” on a family you were fond of. That answer is not a malfunction. It is the most common true answer in this business.

Who it is for

Teams whose event-study claims must survive hostile recomputation: quant researchers publishing internally to a risk committee, desks evaluating third-party research, and anyone who has watched a “significant” effect evaporate the moment someone else ran the numbers.