Offline payment-review reference¶
This synthetic reference profiles a structured review pipeline:
transaction/request fixture
-> suspicious signal
-> fixture refund eligibility
-> manual-review requirement
-> review route label
Use a fresh or empty output directory. This is an evaluation workload, not financial authorization logic. It contains no payment processor, fraud model, refund policy engine, money movement or production approval service. Real financial eligibility, fraud, compliance and approval policies remain the deploying organization's responsibility.
Frozen inputs and independent outcomes¶
Eight fixtures contain only qualitative amount bands and fictional evidence/
policy flags. No card, account, bank, payment, transaction or personal identifiers
are included. within_fixture_window is an invented fixture signal, not a real
refund time limit or legal rule.
Independent labels cover suspicious, eligible, manual_review and route.
Cases include ordinary eligibility, legitimate high-value activity, a small
duplicate signal, missing/conflicting evidence, a failed fixture condition,
required review and missing proof. The application functions receive only the
input record and intermediate state; expected outcomes remain separate.
Missing evidence produces null business decisions and a manual-review label.
That is a deliberately correct abstention under this fixture policy, distinct
from a timeout that fails execution and produces no output at all. Route values
such as refund_review and standard_review are labels; no refund or denial is
executed.
Executed configurations¶
| Configuration | Behavior |
|---|---|
baseline |
Four deterministic local model-style fixture calls |
hybrid |
A fixture call for the suspicious signal, then explicit local rules |
cheap |
An amount-only proxy and simplified rules that intentionally miss important evidence/policy constraints |
failing |
Fixture calls with a declared eligibility timeout for one case |
The cheap proxy creates a false positive on a legitimate high amount and a false negative on a small duplicate signal. It also mishandles missing evidence/proof and required review. Lower fixture cost or agreement with a baseline on some records cannot establish safety. Independent labelled checks reject its mistakes; the example never selects or activates a configuration.
Profiling and evidence¶
All four decision stages use the generic workflow contract with stage/configuration versions, explicit dependencies, actual local callback durations, outcomes and input/output references. The workflow and every backend measurement are synthetic. A failed stage retains its error category and failed workflow, without fabricating an output or an absent billing response.
The shared local backend uses the documented json-whitespace-fixture-v1
tokenizer and a synthetic rate card. Counts are exact for those fixture units,
not real LLM token counts. Reported fees come from the synthetic backend, not an
actual processor or model provider. Span metadata identifies both sources.
These values exercise cost/token accounting; they are not real-world savings.
Every candidate runs on the same frozen inputs and fields-scorer version. Native quality, replay, paired-study and analysis artifacts retain false positives, false negatives, labelled abstentions, manual-review outcomes and actual failed steps. Timings are measured, so normal latency gates can vary for short callbacks; the reference does not manufacture or claim a general latency improvement.
Outputs include the corpus/summary, per-case quality and replay JSON/Markdown, native study manifests/reports, and trace/diagnosis/HTML analysis for every run. Unknown quality and billing remain unavailable. The timeout remains in study denominators rather than becoming an artificially cheap successful run.
To adapt this reference, supply permission-cleared data, an independently reviewed organization-owned rubric, explicit implementation/scorer versions and honest usage provenance. Keep evaluation separate from authorization and action systems. AgentLoop records and compares the application's decisions; it does not decide real financial eligibility or grant permission to move money.