Proof Report — Cash Recovery Engine

Strictness: PROOF_STANDARD = IRS_AUDITOR Status: PROOF_INCOMPLETE

CERTIFIED is not used. Under IRS_AUDITOR rules, CERTIFIED requires that all disclosed seams do not contradict the customer outcome. The customer outcome is a cash-recovery worklist for a real receivables ledger; no real ledger is present in this workspace, so that outcome is a DISCLOSED_SEAM. Status is therefore PROOF_INCOMPLETE. The authoritative machine record is proof/PROOF_DECISION.json (assigned by the Proof Layer, not by hand).

What is claimed, and the evidence for each

All numbers below have lineage in proof/CLAIM_EVIDENCE.json and verification-report.json. Reproduce with node verify.mjs (see proof/REPRODUCE.md).

Claim	Evidence	Status
Engine + optimizer pass all structural + benchmark checks	verification-report.json (13/13), proof/evidence/verify.log	SUPPORTED
Self-cure propensity AUC 0.817; calibration ECE 0.018	verification-report.json#/syntheticBenchmark	SUPPORTED (synthetic input)
Engine recovers +127% cash vs. the strongest simple baseline; 2.94× FIFO	verification-report.json#/syntheticBenchmark	SUPPORTED (synthetic input)
Worklist is always within the collector-hours budget	verification-report.json#/checks (knapsack feasibility)	SUPPORTED
Recovery figure for a specific company	none	DISCLOSED_SEAM — NOT PRODUCED
Individual per-invoice uplift on real data	n/a (only one potential outcome observable)	DISCLOSED_SEAM — fundamentally unobservable

Downgraded statements (per IRS_AUDITOR rule)

• "Recovers $X for your company" → "On a synthetic benchmark, the worklist

recovers +127% more cash than the strongest simple baseline at equal hours; a company-specific figure requires that company's ledger and a holdout test."

• "Production-validated" → "Not deployed and not validated in production; the

engine runs and verifies locally on synthetic data only."

• "Knows which invoice will pay" → "Estimates calibrated probabilities and an

uplift; outcomes are probabilistic, not certainties."

Disclosed seams (DISCLOSED_SEAM)

1. No real ledger — all metrics are measured on src/synth.mjs, not on any

company's receivables.

1. Uplift identifiability — individual treatment uplift is measurable here

only because the synthetic world exposes both potential outcomes. On real data, uplift is an estimate validated by holdout/A-B test.

1. Modelled effort & timing — collector effort-hours and days-to-pay are

modelled parameters, not timed field observations.

1. Projections, not realized cash — "cash accelerated" and

"collection-days reduction" are model projections over the ledger.

See proof/LIMITATIONS.md and the verification report's "Disclosed seams" section for the authoritative list.

Required proof artifacts

Present under proof/: EXECUTIVE_EVIDENCE.md, CLAIM_EVIDENCE.json, EXECUTION_TRACE.json, ARTIFACT_MANIFEST.json, CHECKSUMS.json, REPRODUCE.md, VERIFY.md, LIMITATIONS.md, AUDITOR_OBJECTIONS.md, AUDITOR_CHALLENGE.md, EVIDENCE_GRADE.md, TRUST_SCORE.json, PROOF_SCORECARD.json, PROOF_DECISION.json, and raw runs under proof/evidence/. Regenerate from the repo root with node tools/forge-proof.mjs --outcome delivery-package/cash-recovery-engine --verify "node verify.mjs". Verify integrity with node tools/forge-proof-verify.mjs --outcome delivery-package/cash-recovery-engine.

Path to closing the disclosed seam

Provide data/official/history.csv + data/official/open.csv, run node verify.mjs (real control-AUC) and the Proof Layer generator. To attribute recovered cash, run a holdout or A/B test (work the engine's queue for one cohort and business-as-usual for another) and compare realized cash. If checks pass on real data with evidence preserved and the seam closed, the status may be re-evaluated against the certification criteria.