ROUTEAUDIT: Interaction-Aware Identification for Budgeted Multi-Verifier Routing

1 INTRODUCTION

• Endpoint gaps can confound route changes with contract shifts, so RouteAudit compares policies while holding the contract fixed.A matched route contrast changes only the policy coordinate.Source: §1 · pp. 1–2
• The contract tracks request support, verifier catalog, availability, accounting, online filtration, and post-trace scoring.The naive contrast mixes a route change with any changed contract coordinate.Source: §1 · p. 1
• The same contract applies to learned routing: training episodes may update policy parameters, while the current request's outcome remains outside its online state.RLVR supplies a delayed-reward policy evaluated under that contrast.Source: §1 · p. 2
• Historical cache, task, and serving studies compare fixed, calibrated, learned, and RLVR policies to distinguish verifier-choice gains from measured quality–cost trade-offs.The settings include matched controls through separate small-format, small–large agreement, and large-format checks.Source: §1 · p. 2
• Permutation-averaged bridge increments decompose endpoint changes, expose interactions, and summarize bridge-order dependence as path sensitivity.A nonzero interaction means the measured policy advantage changes with the contract coordinate.Source: §1 · p. 2
• A policy-independent response tape couples adaptive policies to shared provider realizations, while request-level bounds use observed outcomes and declared score ranges for incomplete matching.The certificate binds each comparison's contract, support, traces, scorer, coupling regime, bridge set, and estimate or interval.Source: §1 · p. 2

2 RELATED WORK

• The cited background covers selective prediction, model-call routing, process verification, reward-model audits, tool-use benchmarks, offline RL, and self-feedback.It motivates keeping proxy scores separate from task outcomes and recording feedback visibility at the interface.Source: §2 · p. 2
• Recent routing and test-time-compute work studies verifier allocation, stopping, specialist selection, and policy optimization.RouteAudit complements these objectives by testing whether cascade gains survive a matched verifier-set control.Source: §2 · pp. 2–3
• RouteAudit treats the comparison contract as the measurement object, recording traces, budget conservation, matched controls, route diversity, and the post-trace oracle join.The same contract is instantiated for learned routing, task scoring, and RLVR.Source: §2 · p. 3

3.2 ESTIMANDS AND MATCHED CONTROLS

• The matched route contrast holds the verifier set and contract fixed, while the raw-tail comparison separates the extra LF verifier’s contribution from the policy difference.If cascade and static SF+SA outputs agree record by record, the full-static gap is attributed to LF under that cache and ledger.Source: §3.2 · pp. 3–4

3.3 IDENTIFICATION AND INTERACTION CONTRACT

• The contract lattice averages coordinate increments across bridge orders, reports path sensitivity, and distinguishes average attribution from route–coordinate interaction.Path sensitivity is a diagnostic and does not replace higher-order interaction terms.Source: §3.3 · p. 5
• A policy-independent response tape pairs adaptive policies on the same provider realization, identifying a conditional sequential contrast when actions are non-anticipating and traces close before scoring.Independent provider draws represent a different coupling regime and target an expectation over two independent realizations.Source: §3.3 · p. 5
• For incomplete matching, request-level score bounds yield a sharp finite-population interval; the certificate reports identification status, conditions, and a point estimate or interval.A violated online, accounting, coupling, or trace condition changes the status before an aggregate route claim is reported.Source: §3.3 · p. 5

3.4 LEARNED AND RLVR POLICIES AS STRESS TESTS

• Supervised and RLVR routers stress-test the same observed-only identification contract; RLVR trains on completed traces with delayed task scores and freezes parameters for held-out evaluation.The route contract, rather than a particular optimizer, determines held-out identification status.Source: §3.4 · p. 6

3.5 ROUTE CONTRACT AND COST LEDGER

• The historical catalog assigns costs of 2 and 12 units to its cheap and expensive verifiers under a 14-unit budget; raw-tail SF, SA, and LF cost 2, 4, and 4 under a 10-unit budget.Declared units, generation tokens, and latency measure separate resource quantities; route events precede oracle lookup.Source: §3.5 · p. 6

3.6 PROVIDER SEMANTICS AND PROXY ORACLE

• SF and LF parse cached answers, SA checks normalized-marker agreement, missing calls become ⊥, and proxy verifier quality remains distinct from normalized task-answer quality.For the SF+SA vote, the weights are (0.78, 0.86) and SA breaks ties.Source: §3.6 · p. 6

3.7 SELECTION256 AND PAIRED INFERENCE / 3.8 NEGATIVE-CONTROL GATE

• Selection uses a 256-record calibration split and disjoint 1,319-record held-out split; paired cache tests and four operating-point gates retain threshold, savings, and provenance checks.The gates require mixed paths, non-inferior quality to the matched static control, positive declared savings, and reproducible provenance.Source: §3.7 · pp. 6–7, §3.8 · p. 7

4.1 EXPERIMENTAL SETUP

• The task study uses 1,319 held-out requests and five seeds; all five policies share verifier access and ceilings of 3 calls, 10 ledger units, 4096 tokens, and 1.0 GPU-s per request.The mechanism evaluation uses a 16-case catalog, 64 raw-tail calibration records, and held-out caches of 128 and 32 records; serving uses seven measurement seeds.Source: §4.1 · p. 7, Table 3 · p. 8

4.2 MAIN TASK AND SERVING RESULTS

• On 1,319 held-out requests, RLVR reaches highest quality (0.9553), lowest mean calls (1.4466), and lowest normalized cost (0.680), while learned routing has lowest serving p95 (756 ms).The reported RLVR paired delta is +0.0068 with interval [0.0015, 0.0122]; Figure 2 shows task quality, serving cost, and paired quality differences.Source: Table 4 · p. 8, Figure 2 · p. 8, §4.2 · p. 8

4.3 A 2-BY-2 DECOMPOSITION OF THE APPARENT GAIN

• Matched static SF+SA uses 2 calls and 6 units like the cascade; the route contrast is zero, while verifier-set contrasts are +0.1797 and +0.1250 across two caches.The expanded-catalog held-out cells report ΓV = −0.0027.Source: §4.3 · p. 9

4.4 NEGATIVE-CONTROL GATE AND CALIBRATION BOUNDARY

• At mixed thresholds, cascade proxy accuracy is 0.921875 versus 0.984375 for fixed SF+SA; at τ = 0.95, quality matches but route diversity is zero.Held-out mixed-path deltas are −0.03125 on baseline-128 and 0.0000 on corrected-2B-32, with savings of 0.296875 and 0.125000, respectively.Source: §4.4 · p. 9
• SF has mean confidence 0.3375, accuracy 0.921875, and ECE 0.68125; SA has 13 non-abstaining observations, accuracy 1.0000, and ECE 0.1000.The Selection256 grid selects (0.85, 0.30) at quality 0.9492 and 4.7188 units.Source: §4.4 · p. 9

4.5 EVIDENCE BOUNDARY

• Controlled attribution recovery reaches route MAE 0.0011 and reconstruction error 0.0004; full-contract false-attribution rate is 0.0000, versus 0.0498 MAE under controlled violations.The blind semantic audit records zero clean false positives in each corruption family.Source: §4.1 · p. 7, §4.5 · p. 9

4.6 ROBUSTNESS AND INTEGRITY ANALYSIS

• Removing SF lowers quality to 0.9136 and raises fail-closed rate to 0.1835, while removing LF leaves quality at 0.9484; a one-call ceiling gives 0.9219 versus 0.9484 for two or more calls.Label noise from 0 to 0.20 lowers quality to 0.9143; the checker detects 2,025/2,028 mutations with zero clean false positives, missing availability and reason fields.Source: §4.6 · p. 9

5 LIMITATIONS

• Identification is conditional on request support, verifier catalog, availability, accounting, and the online filtration.The task study uses a short routing horizon, although held-out catalog scaling reports matched K = 3 and K = 4 controls.Source: §5 · p. 9
• Contract-interaction and path-sensitivity values apply to measured factorial cells and bridge paths, not a universal route effect.The request-paired RLVR interval is [0.0015, 0.0122], while the training-seed-by-request interval is [0.0006, 0.0131].Source: §5 · p. 9
• RLVR requires delayed verifiable outcomes and an implementation record linking policy, objective, and checkpoint.The tables compare operating points rather than identify a universally superior optimizer.Source: §5 · p. 9

6 CONCLUSION

• RouteAudit treats verifier routing as a contract-conditioned identification problem with an explicit information boundary, bridge path, and cost ledger.The certificate reports path-sensitive attribution, sharp intervals for incomplete matching, and coupled potential outcomes when provider randomness is shared.Source: §6 · p. 10
• On raw-tail caches, matched controls assign the apparent quality gains to verifier-set choice; expanded verifier access yields a route–catalog interaction of −0.0027.Source: §6 · p. 10
• On 1,319 held-out requests, RLVR reaches quality 0.9553, 1.4466 mean calls, and normalized cost 0.680, with hierarchical difference interval [0.0006, 0.0131] against matched static.The policy results are a stress test for the identification interface; the central claim is the contract-conditional route contrast and its auditable stability.Source: §6 · p. 10

AI-USE STATEMENT

• Generative AI tools were used to polish language and summarize references.Source: AI-USE STATEMENT · p. 10
• The authors state AI tools were not used to generate results, create synthetic datasets, formulate mathematical claims, provide proofs, or decide research conclusions.Source: AI-USE STATEMENT · p. 10
• The authors say they designed and verified the methodology, experimental setup, analysis, and interpretation, and take responsibility for the final content.Source: AI-USE STATEMENT · p. 10

• The evaluation uses benchmark requests and de-identified replay records, with no personal data collection or human or animal subjects.Source: ETHICS STATEMENT · p. 10
• The protocol exposes abstentions, provider failures, resource use, and audit violations so routing errors can be reviewed alongside quality.Source: ETHICS STATEMENT · p. 10
• Benchmark and software reuse remains subject to the licenses of the respective resources.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

• The reproducibility package includes code, configuration templates, manifests, figure sources, and replay metadata bound to the manuscript through hashes.Source: REPRODUCIBILITY STATEMENT · p. 10

A MATCHED TASK AND SERVING EVALUATION

• Five matched policies share request IDs, candidate access, and budget ceilings, while task, serving, and historical timing remain separate measurement regimes.Source: A MATCHED TASK AND SERVING EVALUATION · p. 13

A.1 RAW-TAIL GATE AND PAIRED AUDIT

• The gate identifies a mixed-path quality failure, while the paired table separates full-static discordances from the matched SF+SA route contrast.The gate and paired comparison retain the operating boundary and matched control for inspection.Source: §A.1 · p. 13, Table 5 · p. 13, Figure 4 · p. 14

A.2 MATCHED TASK EVALUATION

• 0.9553 full-set and 0.9560 accepted-set scores make RLVR highest; its matched-static paired difference is +0.0068, with request-paired interval [0.0015, 0.0122].Source: §A.2 · pp. 14–15
• RLVR uses fewer tokens and calls than the learned router, has a task-run p95 of 107 ms, and reports the lowest monetary cost at $0.0157 per request.Source: §A.2 · p. 14

A.3 MATCHED LIVE-SERVING EVALUATION

• The learned router records 438/756 ms p50/p95 and 18.2 requests/s; RLVR trades a 773 ms p95 for higher task quality and normalized cost of 0.680.Source: §A.3 · p. 16, Figure 5 · p. 16

B.1 CANONICAL EXPERIMENT NARRATIVE

• The historical mechanism screen and raw-tail replay are separate evidence tracks, with distinct catalogs, budgets, and cache protocols.The raw-tail threshold is selected on 64 disjoint calibration records before held-out evaluation.Source: §B.1 · pp. 16–17

B.1.1 IMPLEMENTATION DETAILS

• Each event records sample identity, observations, availability, budget, catalog version, and route reason; deterministic replay checks hashes, ordering, and state transitions before oracle joining.Source: §B.1.1 · p. 18

B.1.2 ADDITIONAL ANALYSIS

• The additional analysis distinguishes sample-level Wilson intervals from seed-level uncertainty and records unavailable and budget-rejected calls as separate failure statuses.Source: §B.1.2 · p. 18

B.1.3 REPRODUCIBILITY CHECKLIST

• The checklist binds calibration and held-out roots to IDs and hashes, joins the oracle only after traces are frozen, and defines explicit audit questions and evidence scope.Source: §B.1.3 · pp. 18–19

B.1.4 MAIN RESULTS

• 1.0 proxy-oracle accuracy, selective accuracy, and coverage hold for both policies on the 16-case catalog; cascade uses 1.625 calls and 9.5 units versus static’s 2.000 calls and 14.0 units.The exact 95% interval for 16/16 proxy-correct outcomes is [0.794, 1.000].Source: §B.1.4 · p. 19, Table 16 · p. 19
• 0.9453 and 0.9688 are the cascade and matched SF+SA proxy accuracies on baseline-128 and corrected-2B-32, respectively; matched route delta is 0.0 on both caches.The full-static gap is assigned to verifier-set choice under this frozen replay.Source: §B.1.4 · pp. 19–20, Table 17 · p. 20

B.1.5 ON-DEMAND LOCAL-PROVIDER CPU REPLAY

• 0.9375 cascade proxy accuracy versus 0.7422 static changes ledger cost from 3 calls and 10 units to 2 calls and 6 units in the local-provider baseline row.Source: §B.1.5 · p. 20, Table 20 · p. 21, Figure 6 · p. 21

B.1.6 ABLATIONS AND FAILURE BOUNDARIES

• Mixed-path thresholds lose quality, while the high-threshold route restores it without savings or route diversity; the paired audit reports 23 and 4 aggregate discordances across the two caches.Source: §B.1.6 · p. 21, Figure 8 · p. 22, Figure 9 · p. 22

B.1.7 SCOPE AND EXTENSIONS

• The historical evidence is a 16-case frozen-catalog CPU screen and the follow-up a three-check raw-tail cache replay; reported quality values are proxy joins.Source: §B.1.7 · p. 22

C TASK DEFINITION AND SCORING CONTRACT

• The task policy sees paid observations and remaining budget, while the task oracle is joined after the action trace is frozen for separate inspection of score and ledger.Source: C TASK DEFINITION AND SCORING CONTRACT · p. 22, §C.1 · pp. 22–23, Table 21 · p. 23

C.2 NORMALIZED TASK SCORER

• 1,237 records were both correct, versus 5 proxy-only and 14 task-only, so normalized scoring recovers more answers than it loses on held-out outputs.This scorer diagnostic uses one set of outputs; the five-policy comparison also changes the selected decision path. The fixed scorer is applied to every route before paired differences are computed.Source: §C.2 · pp. 23–24, Table 22 · p. 24

C.3 SELECTION256 ENUMERATION

• 0.9492 quality at (τa, τv) = (0.85, 0.30) was selected, using 4.7188 units and 0.41 ms; two 0.95-threshold candidates matched quality with more units.The selection grid uses quality, units, milliseconds, then lexicographic threshold order for ties; rows below 0.75 fail the recorded quality gate, and held-out scoring follows selection.Source: Table 23 · p. 24, §C.3 · p. 24, Figure 11 · p. 25

D VERIFIER CORRELATION AND ROBUSTNESS

• Verifier pairs are assessed by agreement, conditional errors, and phi within confidence and availability strata; frozen-output dependence checks precede one-field sensitivity perturbations and resource accounting.Together, these views distinguish changes associated with repeated evidence, a changed failure boundary, or a different amount of computation.Source: D VERIFIER CORRELATION AND ROBUSTNESS · pp. 24–25

D.1 ERROR DEPENDENCE

• All reported verifier pairs have positive error dependence: SF–SA has joint error 0.0313 on both caches and coefficients 0.4421 and 0.6956, while SF–LF coefficients are 0.2513 and 0.2444.Because LF has larger marginal error than SA, lower correlation alone does not determine verifier choice; the audit also retains joint error and matched final-decision comparisons.Source: §D.1 · p. 25, Table 24 · p. 25

D.2 AVAILABILITY, NOISE, AND BUDGET SENSITIVITY

• Removing LF preserves fixed-cascade quality 0.9484, while removing SF or SA yields 0.9136 or 0.9272; label noise lowers quality to 0.9143 as units rise from 4.81 to 4.88.At one call or two units, quality is 0.9219 with fail-closed fraction 0.0788; two calls or six units restore quality 0.9484, and larger ceilings do not change outcomes.Source: §D.2 · p. 26, Table 25 · p. 26, Figure 12 · p. 27

E END-TO-END EFFICIENCY AND SERVING LEDGER; E.1 COMPONENT ACCOUNTING; E.2 COST–QUALITY FRONTIER

• Generation contributes 0.566 of normalized cost, verifier work 0.224, and remaining components 0.210; network and queue tails are 138.7 and 72.1 ms.The serving comparison uses per-request component measurements and paired request-bootstrap and seed-level intervals for the quality–cost frontier; latency totals must be formed per request before quantiles.Source: E END-TO-END EFFICIENCY AND SERVING… · p. 26, §E.1 · p. 27, §E.2 · p. 27

F.1 INVARIANT FAMILIES

• The detector flags 2,025 of 2,028 mutations, detects all sample-identity, ordering, paid-read, budget, oracle-boundary, and artifact-hash mutations, and records zero clean false positives.It misses one availability mutation and two reason mutations, identifying provider-status and reason fields as needing careful semantic validation.Source: §F.1 · p. 28, Table 28 · p. 28

F.2 BLIND AND COMPOUND CORRUPTION; F.3 BLIND SEMANTIC CERTIFICATE AUDIT; G REPRODUCTION, PROVENANCE, AND ARTIFACT MAPPING; I.1 OPTIMIZER AND IMPLEMENTATION CONTRACT; I.2 STRONG LEARNED-ROUTING BASELINES

• The blind corruption generator tests schema-valid semantic faults, while the optimizer and baseline procedures bind policy implementation, checkpoint selection, and held-out evaluation to the declared contract.The release artifacts map results to source records and hashes; baselines use 4,096 training IDs, 256 calibration records, and 1,319 held-out IDs, with the post-trace oracle reserved as a diagnostic upper bound.Source: §F.3 · p. 29, G REPRODUCTION, PROVENANCE, AND… · p. 30, §G.2 · p. 30

F.4 CASE-BASED FAILURE ANALYSIS

• Correlated error is the largest held-out error class with 26 cases, followed by parser disagreement with 14; both have median expenditure of six units.Availability and budget cases have median two units, distinguishing information-quality failures from failures to obtain a second observation.Source: §F.4 · p. 29, Table 31 · p. 29

H.1 THRESHOLD SENSITIVITY; H.3 PER-RECORD ROUTE EXAMPLES

• The mixed threshold region has one-call fraction 19/64 = 0.296875, quality 0.921875, and delta −0.0625; at τ = 0.95, every request takes two calls and quality rises to 0.984375, matching the control.Representative route fixtures include a lowest-cost case with quality 1.000 after SF, a three-verifier discordance case with +1.000 quality difference, and distinct unavailable-provider and tight-budget fallbacks.Source: §H.1 · p. 31, Table 34 · p. 31, §H.3 · p. 32

H.2 BUDGET-MATCHED CONTROLS

• Static SF+SA and the cascade share proxy and task quality on both caches; normalized task scores are 0.9531 versus 0.7813 for LF controls on baseline-128 and 0.9688 versus 0.8750 on corrected-2B-32.The separate mixed-threshold sweep reports one-call fractions of 0.296875 and 0.125000; these are not the two-call operating-point ledger in Table 35.Source: §H.2 · p. 31, Table 35 · p. 31

I.2 STRONG LEARNED-ROUTING BASELINES; I.4 MULTI-TASK VALIDATION

• Learned-router comparisons pair the same requests, split, catalog, and scorer, while multi-task validation keeps math, knowledge, and tool/agent contrasts on separate task-specific denominators.The multi-task comparisons use matched static and RLVR routes within each family and do not pool unlike outcomes.Source: §I.2 · p. 34, Table 39 · p. 34, §I.4 · p. 35

I.3 IDENTIFICATION STRESS TESTS

• Across eight contract violations, mean absolute attribution error is 0.0498; verifier-set mismatch is largest at 0.1228, and all comparisons are flagged before release.The stress test treats the certificate as an attribution diagnostic, not a post-hoc quality adjustment.Source: Table 40 · p. 35, §I.3 · p. 35

I.5 GROUND-TRUTH CONTRACT-FACTORIAL ATTRIBUTION

• In controlled factorial attribution, route and endpoint errors decline as matched bridge information and interaction terms enter the analysis.The experiment varies route, verifier set, budget, and availability under a declared data-generating process with known coordinate effects.Source: §I.5 · pp. 35–36, Figure 13 · p. 36

I.6 CONTRACT RE-AUDIT OF ROUTING-SYSTEM REPLAYS

• The FrugalGPT and RouteLLM replay gaps are principally assigned to verifier/model access, while SLARouter retains an identified route contrast of +0.0088.FrugalGPT and RouteLLM certificates are partially identified; original-system deployment claims require original traces and run contracts.Source: §I.6 · p. 37

I.7 CONTRACT INTERACTION AND PATH SENSITIVITY

• Bridge paths change one contract coordinate per edge; their sums reconstruct the endpoint contrast, while Ωj captures attribution variation across legal orders.The factorial replay fixes held-out requests and uses a common denominator for every cell.Source: §I.7 · p. 37, Table 46 · p. 37

I.8 INCOMPLETE MATCHING AND SHARPNESS; I.9 STOCHASTIC-PROVIDER COUPLING

• Incomplete-match stress tests separate random, difficulty-biased, and verifier-dependent removals, while provider studies distinguish frozen replay, shared response tapes, and independent draws.The reported conservative envelope has width 4ϵ; the request-level bound has width 2ϵ when both arms are absent on an ϵ fraction. Independent provider draws average over a different randomness source, so paired estimators and certificate fields remain separate.Source: §I.8 · p. 37, §I.9 · p. 37, Table 48 · p. 38

I.10 RLVR COMPONENT ABLATION

• Quality-only RLVR reaches 0.9570 at 1.9285 calls and 0.914 normalized cost, while supervised imitation reaches 0.9533 at 1.4870 calls.Full RLVR records zero violations; removing the call or unit constraint raises violations to 97 and 74, respectively.Source: §I.10 · p. 38, Table 50 · p. 39

I.11 RLVR TRAINING DYNAMICS

• Across the five-seed checkpoint series, increasing quality accompanies declining calls, units, and reported violations.The reported 0–512 checkpoint window is separate from the 1,200-update optimizer configuration.Source: Figure 14 · p. 39, §I.11 · p. 40

I.12 TRAINING-SEED STABILITY

• RLVR reports higher quality and lower normalized cost in every paired training run; its request-paired interval is [0.0015, 0.0122], with retraining variation estimated separately.Five runs use the same 1,319 held-out requests; the hierarchical bootstrap resamples paired runs and then shared request IDs.Source: §I.12 · p. 40

I.13 SCALING BEYOND TWO VERIFIERS

• RLVR's effective route counts are 1.99, 2.85, and 3.52 at K = 2, 3, and 4, respectively, versus 1.00 for static.The catalogs retain matched static controls and common resource ceilings as verifier count increases.Source: §I.13 · pp. 40–41, Table 54 · p. 41

I.14 SCORER AND DISTRIBUTION ROBUSTNESS

• Frozen policies are evaluated across availability, confidence-shift, label-noise, and budget sweeps, with scorer-agree, proxy-only, and task-only strata defined before comparison.The sweeps record quality, calls, normalized cost, coverage, and route diversity.Source: §I.14 · p. 41, Table 55 · p. 41

I.15 CHRONOLOGICAL STREAMING ADAPTATION

• The streaming study lets delayed rewards update later requests, while outcomes never change their own request's route; it reports cumulative regret at a matched quality constraint.The delay is 8 completed requests, with 512 policy updates and resource violations reported separately.Source: §I.15 · p. 42, Table 56 · p. 42

I.16 NECESSITY OF AUDIT COMPONENTS

• The full contract reduces false attribution from 0.2344 without verifier-set control to zero with complete ledger, oracle-leak, and replay checks.Budget and availability matching reduce the rate to 0.1094; trace freeze and provenance binding reduce it further to 0.0469 and 0.0156.Source: §I.16 · p. 42, Table 57 · p. 42

J.1 OPERATIONAL AUDIT ALGORITHM

• Paid-observation traces close before oracle scoring, after which the evaluator materializes matched per-record comparisons and replayable release records.Release materials include receipts and replay metadata for inspecting historical traces and reconstructing figures.Source: §J.1 · pp. 42–43, §J.5 · p. 45, Table 59 · p. 45

J.2 SEQUENTIAL IDENTIFICATION AND SHARPNESS PROOFS

• A policy-independent response tape couples adaptive policies to the same provider realization, while request-level bounds remain sharp when outcomes are missing.The bounds retain whichever arm is observed rather than discarding an incompletely matched request.Source: §J.2 · p. 44

J.3 ADDITIVE CONTRAST AND THE ROLE OF THE CONTROL

• For the baseline raw-tail cache, quality decomposes 0.9453 − 0.7656 = 0.0000 + 0.1797; path fraction checks call counts, while zero discordance does not generalize.Uncertainty about a new request population remains separate from zero empirical paired variation on a fixed cache.Source: §J.3 · p. 44, §J.4 · p. 45