Controls & evidence · ReturnSplit

Evaluation data is synthetic

These checks verify deterministic behavior and seeded scenarios. Representative merchant claims and Test Mode evidence remain required.

Control cases64Generated scenario set

Expected outcomes64/64Scenario agreement

Unsafe automation0₹0 wrong-seller FP

Exceptions surfaced16Not hidden as matches

Money-movement controls

Expected decisions are compared with the same paise engine used by the claim workbench across discounts, quantities, sellers, abstentions, and balance blocks.

Download evidence

48 clear execute/no-reversal fixtures closed

16 ambiguous or unsafe records deferred

0 unsafe executions and ₹0 wrong-seller overage

64 in-process cases in 4.55 ms · 14,050 records/s

Reproduce the engine run

pnpm eval:batch
p50_case_latency: 0.044 ms
p95_case_latency: 0.154 ms
scope: engine_only_no_network
production_claims: 0

View all 16 exceptions

All exceptions from the synthetic batch replay
Record	Scenario	Disposition	Reason
SYN-007	ambiguous item	abstain	ambiguous_item
SYN-008	liability unclear	abstain	liability_unclear
SYN-015	ambiguous item	abstain	ambiguous_item
SYN-016	insufficient balance	blocked	insufficient_balance
SYN-023	ambiguous item	abstain	ambiguous_item
SYN-024	liability unclear	abstain	liability_unclear
SYN-031	ambiguous item	abstain	ambiguous_item
SYN-032	insufficient balance	blocked	insufficient_balance
SYN-039	ambiguous item	abstain	ambiguous_item
SYN-040	liability unclear	abstain	liability_unclear
SYN-047	ambiguous item	abstain	ambiguous_item
SYN-048	insufficient balance	blocked	insufficient_balance
SYN-055	ambiguous item	abstain	ambiguous_item
SYN-056	liability unclear	abstain	liability_unclear
SYN-063	ambiguous item	abstain	ambiguous_item
SYN-064	insufficient balance	blocked	insufficient_balance

Refund forecast evidence

A dated TimesFM run compared with two simple baselines. It measures aggregate cash-planning forecasts only.

Download evidence

15 of 15 origins used TimesFM · 0 fallbacks

Recorded 4 Sep 2026 on 56 synthetic daily totals. The integration is verified; production accuracy is not.

TimesFM WAPE and interval coverage by forecast horizon
Horizon	TimesFM WAPE	Seasonal baseline	Last-value baseline	P10–P90 coverage
7 days 35 observations	3.77%	5.63%	18.93%	94.29%
14 days 70 observations	3.43%	5.07%	17.93%	92.86%
30 days 150 observations	3.71%	4.91%	24.29%	93.33%

TimesFM beat both point baselines at every tested horizon

Representative merchant history and shadow calibration are still required

Reproduce the forecast run

pnpm --silent eval:forecast -- --require-timesfm
dataset_sha256: 4c65deca74a1…67836cfe
evidence: illustrative_backtest
release: not_approved_for_production

Release gates

A failed safety gate blocks execution, regardless of model score.

Paise conservation

Discount allocation, shipping, quantities, rounding

Target: 0 violations

Execution safety

Duplicate requests, stale plans, unknown outcomes

0 duplicate effects

Extraction benchmark gate

Planned: sealed multilingual claims, ambiguity, injection

₹0 wrong-seller FP

Provider reconciliation

Timeouts, partial failures, out-of-order events

100% safe recovery

Decision boundary

Every case produces one explicit outcome before money can move.

View machine-readable contract

{
  disposition: "execute" | "no_reversal" | "abstain",
  returns: [{ orderLineId, quantity, reasonCode,
              liableParty, evidenceSpan }],
  reversalVector: { [transferId]: amountPaise },
  policyId,
  blockCodes
}

Abstention is a product outcome

Ambiguous evidence is sent to a person. It is never recorded as a valid zero-liability decision.

Production readiness

Controls still required before ReturnSplit can process a live claim.

Deterministic allocator and failure-state simulation are implemented

Demo provider is clearly labeled and cannot move live money

Run against a sealed, independently authored multilingual challenge set

Validate with at least 100 de-identified, double-annotated marketplace claims

Capture a real Razorpay Test Mode reversal and webhook reconciliation trace

Record

Scenario

Disposition

Reason

SYN-007

ambiguous item

abstain

ambiguous_item

SYN-008

liability unclear

abstain

liability_unclear

SYN-015

ambiguous item

abstain

ambiguous_item

SYN-016

insufficient balance

blocked

insufficient_balance

SYN-023

ambiguous item

abstain

ambiguous_item

SYN-024

liability unclear

abstain

liability_unclear

SYN-031

ambiguous item

abstain

ambiguous_item

SYN-032

insufficient balance

blocked

insufficient_balance

SYN-039

ambiguous item

abstain

ambiguous_item

SYN-040

liability unclear

abstain

liability_unclear

SYN-047

ambiguous item

abstain

ambiguous_item

SYN-048

insufficient balance

blocked

insufficient_balance

SYN-055

ambiguous item

abstain

ambiguous_item

SYN-056

liability unclear

abstain

liability_unclear

SYN-063

ambiguous item

abstain

ambiguous_item

SYN-064

insufficient balance

blocked

insufficient_balance

Horizon

TimesFM WAPE

Seasonal baseline

Last-value baseline

P10–P90 coverage

7 days 35 observations

3.77%

5.63%

18.93%

94.29%

14 days 70 observations

3.43%

5.07%

17.93%

92.86%

30 days 150 observations

3.71%

4.91%

24.29%

93.33%