Third-Party Technical Replication¶

Verification method — instantiates Capture-Resistant Institutional Design

Has an independent party reproduce the regulated actor's key technical claims from scratch, so the institution's decisions rest on evidence it can verify rather than on figures only the actor can produce.

Third-Party Technical Replication breaks the institution's dependence on the regulated actor's own evidence by commissioning a disinterested party to reproduce the critical technical results — the safety test, the model output, the emissions figure, the risk number — from the underlying data and methods. The distinguishing idea is that it attacks capture through the evidence base specifically. When the only source of the facts a decision turns on is the actor being judged, the actor effectively controls the decision, however independent the decider. Independent replication supplies a second, disinterested source of truth, so the institution can stand on evidence it has verified rather than on numbers only the regulated party can compute.

Example¶

A bank tells its supervisor that its internal model puts required capital at a comfortable level. If the supervisor simply accepts the figure, the bank has, in effect, set its own capital requirement — capture straight through the evidence base. Instead the supervisor has an independent team reproduce the result: re-running the bank's exposures through a supervisory benchmark model, or re-estimating from the raw data with independent methods, to see whether the ≈comfortable number survives. When the independent replication lands materially higher, the burden flips — the bank must now explain the gap, rather than the supervisor having to disprove a figure only the bank could produce. The decision no longer rests on the regulated party's word about its own math.

How it works¶

Reproduce from primary inputs. The replicator rebuilds the result from the underlying data and independent methods — not a re-reading of the actor's write-up, which would only re-check its presentation.
A disinterested, separately-resourced replicator. Screened for ties to the actor and funded independently, so the second source is genuinely a second source.
Divergence as the signal. The result is a comparison: a material gap between the actor's number and the independent one shifts the burden of proof onto the actor rather than resolving quietly in its favour.
It diversifies facts, not voices or votes. Its object is whose evidence the decision rests on — a different lever from broadening who speaks or who decides.

Tuning parameters¶

Independence of the replicator — a fully external lab versus an in-house second team. More independent is more credible but slower and costlier.
Reproduction depth — re-running the actor's own pipeline versus rebuilding from raw data with independent methods. Deeper catches capture baked into the method, not just the arithmetic, but costs more.
Coverage — replicate everything versus only the decision-critical claims; targeting scarce capacity keeps the method affordable.
Divergence threshold — how large a gap counts as material enough to flip the burden of proof.
Data-access terms — how much of the actor's data and methods the replicator receives. More access improves fidelity but risks quietly importing the actor's assumptions along with its data.

When it helps, and when it misleads¶

Its strength is that it removes the actor's monopoly on the facts — often the deepest capture channel of all, since selective data and self-serving models bend a decision before any decider sees it — and it does so on the actor's own turf of technical merit, where voice-based and procedural checks cannot reach.

Its failure mode is captured replication: a "replication" that quietly reuses the actor's data, code, or assumptions is not independent and merely rubber-stamps the original with a second letterhead. A replicator with undisclosed ties, or one scoped only to the safe, non-decisive claims, gives false assurance — worse than none, because it looks like verification.^[1] The disciplines that keep it honest are to rebuild from primary inputs with independent methods, screen the replicator for ties to the actor, and treat divergence — not agreement — as the informative result worth chasing.

How it implements the components¶

This method fills a single component, and pointedly so:

evidence_diversification_layer — it is the evidentiary diversification: an independent reproduction of the actor's key technical claims, so the evidence base underpinning a decision no longer rests solely on the actor's own figures.

It touches only the evidence base. It does not open lay or stakeholder voice (that is the Public Comment Docket), convene a countervailing expert body (Countervailing Stakeholder Panel), map or measure the capture channels (Capture Risk Audit), or decide anything (Independent Oversight Board) — those handle voice, deliberation, diagnosis, and the call; this handles whose facts.

Instantiates: Capture-Resistant Institutional Design — it denies the regulated actor a monopoly over the evidence the institution relies on.
Sibling mechanisms: Public Comment Docket · Capture Risk Audit · Conflict-of-Interest Disclosure and Recusal · Countervailing Stakeholder Panel · Enforcement Pattern Dashboard · Ex Parte Contact Disclosure Rule · Funding Firewall Rule · Independent Oversight Board · Mandatory Reason-Giving for Exceptions · Privileged Access Log · Revolving-Door Cooling-Off Period · Sunset and Reauthorization Review

References¶

[1] Independent replication — reproducing a result from the underlying data and methods by a party with no stake in the outcome. Its value lies in whether the result survives an independent redo, not in the actor's own attestation that it holds; a replication that reuses the actor's pipeline forfeits exactly that value. ↩