The Research Sandbox: A Reference Implementation for Auditing a Corpus's Revision History, Why the Code Cannot Verify Its Own Honest Use, and What a Genuinely Independent Recoding of One Real Object Found

Describing a research corpus as an "operating sandbox" — a bounded, inspectable environment in which concepts, formalizations, and objections can be instantiated, recombined, and revised — is a metaphor until something is actually built to inspect it. This paper builds that something, states plainly what it can and cannot show, and runs it on real data rather than illustrative fiction. The reference implementation is a small Python program, presented here alongside the paper and executable independently of it, that loads a corpus's papers as structured research objects, checks each against two of this corpus's own already-established standards — whether a stated falsification condition exists (Order 1 of Huynh 2026, The Third-Order Audit) and whether that condition points at something outside the corpus's own vocabulary rather than only at itself (Order 2) — and computes, across the whole loaded set, a simple, honest ratio of revision events that narrowed or withdrew a claim against events that merely reaffirmed one unchanged after challenge (an operational, non-mathematical stand-in for Order 3's Lakatosian progressive/degenerating question). Run on seven real papers from this corpus's own recent production, the tool returns a genuine, non-trivial finding: six of seven pass both structural checks, one — The Load-Bearing Test — is correctly flagged for lacking an externally checkable falsification condition, and the programme-level constraint ratio across eleven real revision events is 0.636, with the tool independently surfacing the one documented case in this corpus's own record of a claim being reaffirmed unchanged after a substantive error rather than genuinely revised (Huynh, 2026, The Audit That Missed Itself). This revision adds the specific test Section V of the original edition named as the condition under which its own worked example should be treated as confirmed or overturned: an independent recoding, performed without reference to the original coder's decisions, of the one object in the dataset a separate reviewing process could verify entirely from primary knowledge rather than from the corpus's own report of itself. That recoding reproduces the original's Belsky et al. (2022) reaffirmed-unchanged finding exactly, and surfaces one additional narrowing event the original dataset did not contain — the same paper's own abstract having been corrected, during its drafting, from a claim about "leading models" to a claim about one specifically named, mid-tier model, once verification found the cited statistic characterized the latter and not the former. Extending the programme-level count to include this newly surfaced event moves the constraint ratio from 0.636 to 0.667. This paper's central caution, stated as plainly as its central result, is that neither number demonstrates the underlying philosophical claims are correct, that the tool cannot verify the honesty of its own input coding for the six objects this revision did not have primary access to recode, and that a specific instance of Goodhart's Law (Strathern, 1997) applies to exactly the gap between what was verified and what was reconstructed — a gap this paper marks explicitly rather than lets the reader assume away.