Independent Verification: How CPP Closes the “Same-Author Crossval” Daubert Gap

Which test applies, and where. Daubert is the United States federal standard, applied under Federal Rule of Evidence 702. It does not govern admissibility in a Canadian court. In Canada the framework is R v Mohan, [1994] 2 SCR 9, as restated in White Burgess Langille Inman v Abbott and Haliburton Co, 2015 SCC 23: four threshold requirements (relevance, necessity, absence of an exclusionary rule, and a properly qualified expert), with the expert’s duty to give fair, objective and non-partisan assistance treated at that threshold, followed by a discretionary cost–benefit weighing at the gatekeeping stage. This article is framed on Daubert because it is the clearest published articulation of the specific question at issue: whether a method has been tested, and by whom; and because that question lands the same way under either framework. Reliability is what the Canadian gatekeeping stage weighs, and “who else has tested this?” is asked in Toronto as readily as in Texas. Nothing here should be read as a claim that a Canadian tribunal applies Daubert. For the Canadian test as it bears on instructing a schedule expert, see White Burgess and the schedule expert.

Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993), asks first whether a method “can be (and has been) tested” (at 593), and separately whether it has been subjected to peer review. The Court does not say who must do the testing. But a method tested only by its own author answers the first question thinly and the second not at all. For software-driven forensic schedule analysis, that is a real exposure. If the expert wrote the engine, and wrote the reference implementation, and wrote the test fixtures, the expert has tested their own work. Nobody else has.

It is not a fatal exposure. Software-based forensic methods clear Daubert every day. But when opposing counsel is competent, “who else has tested this?” is the first question. The answer cannot be “nobody yet.”

Critical Path Partners has shipped infrastructure that closes this gap in three layers. None of it is theoretical. All three are live, public, and verifiable today by anyone with a laptop. This post walks through what each layer does, what it proves, and what it still does not prove.

Layer 1: Public Continuous Integration

Layer 2: Sigstore Attestation on the Public Transparency Log

Layer 3: One-Command Local Reproduction

Cross-Examination Walk-Through

Picture the exhibit on the screen: the engine version, the topology hash, the CI run URL. Opposing counsel rises.

The point of the layered posture is that the expert does not have to ask the court to take their word. The expert points to the public log and steps back.

What This Does Not Close

Mechanical reproduction is not peer review. Daubert Prong 2 wants “subjected to peer review and publication.” That is a different bar, and the three layers above do not meet it on their own.

What the layers establish:

• Reproducibility. A third party can independently confirm that the same input yields the same output. Prong 1.
• Tamper-evidence. A third party can confirm the build that produced the disclosure is the build that is public.
• Auditability. A third party can read the engine, read the tests, read the workflow, and read the verification log without our cooperation.

What the layers do not supply:

• External peer review. No peer-reviewed publication of the engine’s methods exists.
• Cross-validation against a second, independently authored engine. None has been published.
• Independent academic study. None exists.

Where to Read the Detail

The canonical write-up lives in the engine repository at DAUBERT.md §3.1. It records the round-7 hardening that produced the public CI, the Sigstore attestation, and the local-reproduction path. The engine itself is MIT-licensed and lives at github.com/danafitkowski/cpp-cpm-engine.