BL-IS-016: Verdict Reproduction

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

1. What it measures

Whether a prior measurement's published figures can be regenerated from its own specification, to the basis point, before that measurement is examined for anything else.

The check is deliberately small. It re-runs the exact pre-registered specification under the exact conditions the published verdict was drawn under, and compares the output against the published numbers, one quantity at a time. Nothing about the strategy or the market is evaluated at this step. The only question is whether the artifact and the code that produced it still agree.

2. Why it exists

An audit instrument that cannot first reproduce what it is auditing has no standing to disagree with it. If a re-run cannot recover the published gross return, cost, and net figure from a verdict that has already been acted on, any subsequent disagreement is ambiguous between a defect in the audit and a defect in the original measurement, and the audit has produced doubt rather than a finding.

This is the instrument most programmes skip, because it produces no new number and answers no new question. It only checks that the old number is still the old number.

3. Interface

No standalone, reusable module implements this pattern yet. It exists today as the first phase of the anchor-robustness audit in the anchor audit, BL-IS-004, ahead of the twenty-four-phase sweep that instrument runs for its own purpose [BL-IS-004]. The pattern is simple enough to state without a function signature:

run the locked specification at the exact anchor phase the published
  verdict used
compare, quantity by quantity, against the artifact that carried the
  original verdict
proceed to any further audit only if every compared quantity matches
  within reporting precision

The specification re-run is bound by the rule that bound the original: no parameter in it may be adjusted at this step. A reproduction that required a parameter change would not be a reproduction.

4. A worked result

The anchor audit into the G5 gate reproduced the published primary result at the gate's anchor hour before sweeping the other twenty-three. All four reported quantities matched to the basis point:

Only after that match did the audit proceed to the question it was actually built for: was the gate's kill verdict robust to the rebalance hour, a parameter that had never been chosen but had been fixed since the grid was first built. It was. Net ranged from -15.48 to +10.19 across the twenty-four phases, with a median of -3.54 and zero phases significant in the positive direction, against the published phase's own net of +4.28 ranking third of twenty-four [E-0029]. The gate's verdict stood, but it stood on a re-verified foundation rather than an assumed one.

5. What it does not do

It does not evaluate whether the original verdict was correct. Reproduction confirms that the code and the artifact agree with each other; it says nothing about whether the specification they agree on was the right one to run. That question belongs to the audit step that follows reproduction, such as the anchor sweep in section 4, not to this one.

It does not tolerate approximate agreement as a substitute for exact agreement. A reproduction that is close but not exact points to a defect somewhere in the pipeline, in the harness, the data snapshot, or a silent parameter drift, and is reported as a failed reproduction rather than rounded into a pass.

It produces no new verdict and consumes no fresh sample quota of its own. It is bookkeeping in the most literal sense: confirming that a ledger entry still says what it said when it was written, before anyone is allowed to argue with the entry.

It does not exist yet as a general, reusable interface. The pattern above has been run once, inside a larger audit, and generalising it into a standalone module that takes any prior artifact and any specification as arguments is future work rather than a claim made here.

6. Cross-references

BL-IS-000 section 3.4 for the principle stated at catalog level. BL-IS-004 for the anchor-robustness sweep this reproduction step precedes. BL-IN-006 for the incident the full audit belongs to. the anchor audit, BL-IS-004. Evidence E-0029, E-0037.

Copyright (c) 2026 Bijective Labs. All rights reserved.