Independent Recomputation or Replication¶
Reproduction method — instantiates Independent Verification Oversight
A separate calculation, experiment, retest, or reanalysis used to check whether the claimed result can be reproduced.
Independent Recomputation or Replication verifies a claim by re-deriving the result: a separate party takes the inputs and re-runs the calculation, re-executes the experiment, retests the specimen, or re-analyzes the data to see whether the producer's answer comes back out. Its defining move is re-doing the work, not tracing it or re-reading it. Agreement between two independent derivations is the evidence; a divergence is not a note in the margin but a contradiction — a second, equally-derived answer that puts the burden of explanation back on the producer. What makes it bite is the combination of genuine independence and enough competence to actually reproduce the result: without independence, it merely re-executes the producer's own work; without competence, a "failure to reproduce" is just the reviewer's mistake.
Example¶
A corporate pension plan reports a $1.2B benefit liability, computed by the plan's own actuary — who works for the sponsor that benefits from a lower figure. Because the number drives the funding contribution, an independent actuary is engaged to recompute it. Not to sample the file, and not to re-read the report, but to re-run the valuation from the raw participant census against a stated set of assumptions. Working from the same census but deriving independently, the reviewer arrives at $1.34B. Tracing the gap, they find it sits almost entirely in a mortality-table choice and a discount rate set at the optimistic edge of the acceptable range. Because the reviewer reproduced the whole calculation rather than merely spot-checking it, they can say precisely which assumptions moved the number — and, holding an equally-derived counter-figure, they have the standing to challenge the reported liability and require it be revisited before the funding decision relies on it.
How it works¶
- Obtain the inputs and the claimed method — not just the producer's output, but the data and the procedure needed to re-derive it.
- Re-derive independently — a competent party separate from the producer computes the result afresh, ideally with its own tools and hands.
- Compare against an agreement tolerance — decide in advance how close counts as "reproduced" versus "diverged," so noise isn't mistaken for error.
- On divergence, localize and challenge — trace the discrepancy to the input, assumption, or step responsible, and put the finding back to the producer as a contradiction to be resolved.
Its distinguishing trait is that it produces a second independent answer and compares — it does not check whether claims are backed by records, only whether the claimed answer is right by getting it again.
Tuning parameters¶
- Independence degree — a fully independent re-derivation (own team, own tools, own code) versus a re-run of the producer's own pipeline; more independence catches shared-method errors but costs far more.
- Recompute versus replicate — re-computing from the same inputs tests the arithmetic and method; replicating with fresh data or a new experiment tests whether the result is robust, not merely repeatable.
- Agreement tolerance — how close the two results must land to count as reproduced; tight tolerances catch small errors but flag benign numerical noise.
- Competence bar — who is qualified to reproduce the derivation; below it, disagreement is uninformative.
- Scope — one headline figure versus the whole computational pipeline; wider scope catches more but multiplies effort.
When it helps, and when it misleads¶
Its strength is that it is the most decisive form of check: it can contradict the producer with an equally-derived counter-result, catching errors that tracing and re-reading miss entirely — the wrong-but-well-documented calculation whose paperwork is immaculate and whose answer is simply incorrect.[n1]
It misleads in two ways. It is expensive, so it is reserved for load-bearing claims and can't cover everything. And same-inputs recomputation faithfully reproduces the producer's shared errors: if both parties apply the same flawed model to the same data, they will agree — and agree wrongly. Reproducibility is not correctness. The classic misuse is the "replication" that reuses the original's exact code and dataset and so merely re-executes it, confirming that the machine is deterministic rather than that the result is true. The discipline is to maximize the independence of inputs and tools wherever the feared error is systematic, and to keep clear whether a given check re-computes (same inputs) or genuinely replicates (fresh ones).
How it implements the components¶
independent_verifier_role— the re-derivation is performed by a party structurally separate from the producer; that separation is what makes a matching answer meaningful and a diverging one credible.reviewer_competence_threshold— reproducing a non-trivial derivation demands real domain skill; the competence bar is what distinguishes a genuine failure-to-reproduce from the reviewer's own error.challenge_authority— a non-reproduction is an empowered contradiction: armed with a second derived result, the reviewer can withhold acceptance and require the producer to explain the gap before the claim is relied on.
It re-derives the answer itself; it does not trace a sample of the producer's assertions to existing records and log each as a traceable_finding_record — that sampling-not-recomputing move is Audit-Trail Sampling, its nearest twin. Recomputation reproduces the whole result; sampling only checks that a subset of claims is backed by the file.
Related¶
- Instantiates: Independent Verification Oversight — supplies the strongest check, an independently derived second answer that can confirm or contradict the producer's result outright.
- Sibling mechanisms: Audit-Trail Sampling · Blind Revalidation · Certification Signoff with Scope Limits · Third-Party Audit · Chain-of-Custody Evidence Review · Red-Team Verification Review · Verification Hold Point
Editorial Notes¶
Form Classification¶
Form family: Experiment, Test & Rehearsal
Rationale: Independent Recomputation or Replication operates as a bounded trial, probe, simulation, or rehearsal that generates evidence from performance because it a separate calculation, experiment, retest, or reanalysis used to check whether the claimed result can be reproduced
Independent corroboration: The frozen evidence defines Independent Recomputation or Replication as 'A separate calculation, experiment, retest, or reanalysis used to check whether the claimed result can be reproduced', so its operative form is Experiment, Test & Rehearsal.
Nearest alternative: Assessment, Review & Assurance — Its form includes a separately executed replication, retest, or experiment rather than only desk review of evidence.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Universal
Rationale: Reproducibility through independent reanalysis or repeat testing is anchored in experimental-method and statistical validation traditions.
Related originating lineages:
- Accounting & Auditing — Independent recalculation and reperformance contributed the parallel verification tradition for documentary and computational claims.
- Computer Science & Software Engineering — Deterministic recomputation of code and calculations supplies a distinct computational verification lineage.
Review resolution: Both independent reviews place the primary lineage in statistics_experimental_design. The queued differences (reported_ambiguity, alternate_origin_disagreement, origin_mode_disagreement, encyclopedia_synthesis_disagreement) concern secondary metadata rather than primary provenance. The final retains computer_science, accounting_auditing only where a reviewer supplied a formative-lineage rationale; this does not convert downstream applicability into origin. origin_mode=cross_disciplinary_synthesis because the entry's present form deliberately composes methods from the documented lineages. domain_reach=universal records application breadth separately from provenance.
Attribution caveat: The entry deliberately joins computational reproducibility with experimental replication, practices with different evidential force.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; medium confidence.
Notes¶
Reproducibility and correctness are not the same thing. Two independent runs that both rest on the same wrong assumption will agree — the check confirms only that the answer is derivable again, not that it is right. This is why the highest-assurance replications vary the inputs, tools, and even the method: shared-error agreement is the one failure that identical re-derivation cannot detect.
[n1] Reproducibility — obtaining consistent results from the same inputs and method — and its stronger cousin replicability — obtaining consistent results from a fresh study — are treated across the sciences as a precondition for trusting a reported result; a finding that cannot be independently reproduced is conventionally held as provisional rather than established. ↩