Capture-Recapture Defect Estimation¶
Estimation model — instantiates Parallel Independent Inspection Design
Estimates how many defects remain unfound by treating the overlap between two independent inspection passes as a mark-recapture sample.
Capture-Recapture Defect Estimation does not find defects — it estimates how many were missed. Run two independent inspection passes over the same artifact, then sort every real defect into found-by-A-only, found-by-B-only, or found-by-both. The size of the both-found overlap relative to the two singleton sets reveals how large the still-hidden population probably is. The defining idea is borrowed from ecology's mark-recapture: if two independent searchers overlap heavily, few defects remain outside their combined reach; if they barely overlap, each is sampling from a large unseen pool. It converts "we found forty problems — are we done?" into a defensible residual number.
Example¶
An audit team is testing a company's accounts-payable ledger for misstatements before signing off. Two senior auditors work from independent samples of the same locked transaction set, blind to each other, and each flags exceptions. Auditor A flags 30 misstatements; Auditor B flags 26; 12 of those are the same transactions — the overlap. The Lincoln–Petersen estimator gives a rough total of (30 × 26) / 12 ≈ 65 misstatements, so about 65 − 44 ≈ 21 remain unfound.[n1]
That residual number, not the 44 already caught, is what tells the engagement partner whether one more pass is warranted or the remaining risk is tolerable. The estimate lives or dies on its assumptions, so the team first checks that neither auditor's sampling steered the other's — because any leakage between the two passes would inflate the overlap and quietly make the ledger look cleaner than it is.
How it works¶
Two (or more) independent, near-complete passes run over the same artifact. Their findings are matched to identify the "recaptures" — the same defect seen twice. A mark-recapture estimator (Lincoln–Petersen for two passes, a richer model for more) turns the counts into an estimated total, a residual, and a confidence band. Independence is not a nicety here but a mathematical precondition: if the passes are correlated, the overlap is inflated and the residual is biased downward.
Tuning parameters¶
- Number of passes — two supports a simple estimator; three or more supports heterogeneity-aware models at higher cost.
- Estimator and heterogeneity correction — defects vary in how catchable they are; a hard defect is unlikely to be caught by either pass, which biases naive estimators, so a heterogeneity-robust model matters when detectability is uneven.
- Match strictness — how close two flags must be to count as the same defect; loose matching over-counts recaptures, strict matching under-counts them.
- Independence enforcement — how hard the two passes are isolated; the tighter, the less biased the estimate.
- Band carried forward — how wide an uncertainty interval the decision inherits.
When it helps, and when it misleads¶
Its strength is giving a principled stop-or-continue signal and a residual-risk figure where teams usually have only a gut sense of "probably enough."
Its failure modes come straight from its assumptions. The estimator assumes the passes are independent and that defects are roughly equally detectable; correlated reviewers — say, two auditors trained on the same checklist — overlap for the wrong reason, understate the residual, and manufacture false confidence. And it cannot count what neither pass could ever see: a defect invisible to both searchers contributes nothing to the overlap and is silently excluded from the estimate.[n1] The guarding discipline is to verify independence, prefer heterogeneity-robust estimators, and treat the residual as a lower bound on what remains.
How it implements the components¶
overlap_coverage_plan— deliberate overlap between two full passes is the data the estimator runs on; with no planned overlap there is no recapture to count.independence_protocol— statistical independence of the two passes is a precondition of the math, so the isolation is enforced as a requirement rather than a courtesy.marginal_yield_and_saturation_metric— the residual estimate is a saturation reading: a small hidden pool means inspection has saturated; a large one means it has not.
It counts whether two passes agree; it never decides which flag is correct or routes a repair — that reconciliation_and_adjudication_loop belongs to Dual or Triple Diagnostic Read and Finding Reconciliation Board. It also plants nothing: measuring catch rate with known defects is calibration_set, owned by Seeded Defect Calibration Exercise.
Related¶
- Instantiates: Parallel Independent Inspection Design — the estimation layer that reads residual risk off overlapping passes.
- Consumes: two independent passes such as those produced by Parallel Code Review Round or Dual or Triple Diagnostic Read.
- Sibling mechanisms: Blind Document Proofing Passes · Dual or Triple Diagnostic Read · Finding Reconciliation Board · Independent Checklist Variant Rounds · Independent Security Review Lenses · Multi-Inspector Manufacturing Sort · Overlap Heatmap · Parallel Code Review Round · Seeded Defect Calibration Exercise
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Estimates how many defects remain unfound by treating the overlap between two independent inspection passes as a mark-recapture sample, making its operative form a computation, comparison, model, or analytic representation used to infer, estimate, or choose.
Independent corroboration: The frozen evidence defines Capture-Recapture Defect Estimation as 'Estimates how many defects remain unfound by treating the overlap between two independent inspection passes as a mark-recapture sample', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Biology & Ecology
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Ecology originated mark-recapture estimation: overlap between repeated captures estimates a population including individuals never observed, which is the exact inferential mechanism later transferred to defect discovery.
Related originating lineages:
- Accounting & Auditing — Audit and inspection settings apply the same overlap logic to estimate defects or events missed by multiple review passes.
- Computer Science & Software Engineering — Software-quality research transferred capture-recapture models to overlapping defect discoveries from independent inspections and tests.
- Statistics & Experimental Design — Statistical sampling formalizes the overlap estimator, independence and catchability assumptions, confidence intervals, and model-bias diagnostics.
Review resolution: Neither reviewer made ecology primary, but both recognized it as formative. Software-engineering literature explicitly describes capture-recapture defect estimation as an adaptation of ecological population estimation. Primary origin therefore follows the originating mechanism rather than the application domain, with statistics, computing, and auditing retained as material lineages.
Attribution caveat: Statistics formalized the estimator and software engineering established defect-estimation applications; ecology is primary because the named and structurally identical source mechanism is mark-recapture.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Researched adjudication after independent review; high confidence.
Sources consulted:
- Capture-Recapture Models in Software Engineering
- A Comprehensive Evaluation of Capture-Recapture Models for Estimating Software Defect Content
- A Capture-Recapture Model for Estimating Software Defects
Notes¶
The overlap heatmap and this estimator both read "saturation" from overlapping passes, but they answer different questions: the heatmap shows where coverage piled up, region by region; this model computes how many defects the piles imply remain. And where Seeded Defect Calibration Exercise plants known defects to measure catch rate directly, capture-recapture plants nothing — it infers the hidden count from two real passes.
[n1] The Lincoln–Petersen estimator, from ecological mark-recapture, estimates a population's size as (captures in sample 1 × captures in sample 2) / recaptures, assuming the two samples are independent and each member is equally catchable. Both assumptions map directly onto the two inspection passes and are the first things to check. ↩a ↩b