Non-Propagating Refutation

A celebrated finding is published. Hundreds of studies build on it. Products ship, policies are written. Then it fails to replicate and is retracted — or a model is deprecated, a policy revoked, an API marked unsafe. The correction exists somewhere. Downstream work still treats the claim as live. Citations continue. Decisions justified by the refuted claim are never revisited. The claim was demoted from established to refuted — and the demotion reached almost none of the things that depended on it.

That opening story is vignette The Retraction in Where Everything Breaks. The failure mode is non-propagating refutation: systems treat refutation as deletion or local notice instead of a connected demotion that travels the dependency graph.

In software systems

Deprecated API still in production paths. A service marks v1/score deprecated after a calibration bug; microservices keep calling it because nothing links runtime traffic to the deprecation commit. Feature flags flip off a bad model in one region; other regions run stale weights until an incident surfaces.

Silent tombstones. Events are soft-deleted from consumer views; projections never receive ClaimRefuted or ModelSuperseded. The append-only log may contain the refutation; operational read paths still serve pre-refutation truth.

Metrics without lineage. A dashboard shows “94.2% accuracy” with no edge to the run, dataset version, or refutation event that invalidated it. When replication fails, the number is edited or the paper retracted — but BI cubes, slide decks, and automated alerts built on the metric are not flagged for review.

What good looks like: refutation as L0 with explicit refs to the refuted artifact and registered dependents; tier-13 treatment (The Epistemic Ladder) — refuted claims stay connected so demotion triggers dependent review.

In human organisations

Journal retraction without cascade. A paper is retracted on the publisher site. Course syllabi, regulatory submissions, and meta-analyses that cited it are not systematically notified. The retraction is an embarrassment in one venue, not an organisational event in every system that relied on the claim.

Policy supersession without impact analysis. Legal replaces a safety standard; operations teams discover the gap during an audit years later. The old policy was “withdrawn”; nothing tracked which decisions and runbooks still assumed it.

Expert reversal without record. A senior analyst publicly reverses a forecast in a meeting; the budget spreadsheet and board minutes still embed the old number because reversal never crossed the commitment boundary.

In socio-technical systems

ML replication failure. The ML reproducibility crisis combines ruptured provenance chain (cannot replay how the result was produced) with non-propagating refutation when replication does disprove a claim: retracted or contested results still anchor product features, clinical guidelines, and follow-on papers. See Event Sourced Science.

Model card updates that never reach binders. A vendor publishes a limitation; downstream automations keep auto-approving because integration stores only the score, not model version or refutation lineage.

Structural causes

Epistemic tier collapse

Refuted claims are removed from view or left at their old effective tier in downstream systems. Tier is not carried along dependency edges; weakest-link rules never run on refutation.

CRUD as domain language

“Delete bad row” or “update status to inactive” replaces refutation as a new fact referencing the old. Without supersession, correction events, and explicit dependent linkage, refutation cannot propagate — only disappear from one table.

Remedy (via patterns)

Honest correction is append-only: refutation is itself a recorded fact, able to reach everything that depended on the overturned claim — the discipline When Facts Are Wrong summarises. Positive capability: refutation propagation.