Work · Evidence record
AI Invoice Matching Across Orders, Receipts, and Exceptions — Three of Five Checks Passed
The completed WFT-002 synthetic field test stopped at 6/10: three of five Invoice Matching checks passed after one correction, but Invoice Matching task fidelity [WFT-002] and Invoice Matching rule accuracy [WFT-002] remained unsupported.
- Exact prompts and outputs
- One correction only
- Synthetic inputs disclosed
01 · The assignment
The task
match invoices to purchase orders and receipts
02 · Scope before score
Test disclosures
Input disclosure
Synthetic blind-test input pack WFT-002: An accounts payable team will supply invoice, order, and receipt records containing deliberate quantity and price mismatches. Source facts: ledger rows WFT-002-L01 through WFT-002-L06; quantities 24, 27, and 42; unit prices $42.50 and $47.25; a 3% discount threshold; one duplicated $118.00 charge; and source document WFT-002-S04 with a missing approval. Governing rule card: the 3% threshold and every quantity-times-price calculation. Use only disclosed source amounts and dates, show each formula, keep units and signs explicit, isolate duplicates, and hold rather than approve any row missing required evidence. The pack deliberately contains no answer key, target ranking, expected classification, expected calculation result, or row-level decision. Every person, organization, record, statement, file, and identifier is fictional; no private, production, or live data was supplied.
Run disclosure
We ran one text-only synthetic benchmark WFT-002 for “match invoices to purchase orders and receipts” in a Codex multi-agent session. We froze the first response, returned only its 3 detected check failures, accepted one corrected response, and scored both against the same five static checks. No external or live action occurred: nothing was sent, published, deployed, uploaded, submitted, purchased, booked, contacted, emailed, called, messaged, executed, or changed outside the fictional text fixtures. No external, live, or production action occurred. Per-case elapsed time was not instrumented during the batch session.
- Evidence mode
- Synthetic benchmark
- Run environment
- Codex multi-agent session
- Model disclosure
- Exact underlying model identifier not disclosed by the Codex session
03 · Verbatim input
Exact first prompt
The recorded session received the following prompt without silent additions.
Run bounded synthetic field test WFT-002. Task: match invoices to purchase orders and receipts. Context: An accounts payable team will supply invoice, order, and receipt records containing deliberate quantity and price mismatches. Fictional source facts: ledger rows WFT-002-L01 through WFT-002-L06; quantities 24, 27, and 42; unit prices $42.50 and $47.25; a 3% discount threshold; one duplicated $118.00 charge; and source document WFT-002-S04 with a missing approval. Governing policy, formula, or rubric: the 3% threshold and every quantity-times-price calculation. Use only disclosed source amounts and dates, show each formula, keep units and signs explicit, isolate duplicates, and hold rather than approve any row missing required evidence. Produce a reconciliation table, calculation notes, and exception ledger. Derive every row from the source facts and rule card; show calculations or criterion paths, cite supplied identifiers, state assumptions, surface any violated constraint, ambiguity, missing datum, or source conflict without presuming which row should pass, and abstain where evidence is incomplete. Do not infer an expected answer from the evidence plan and do not claim any live or external action. Planned verification after the frozen response: A reconciliation table and sampled source comparisons will verify each proposed match and exception.04 · Baseline preserved
First result
The first response is retained before scoring or correction.
Frozen first response WFT-002 produced a reconciliation table, calculation notes, and exception ledger for the task “match invoices to purchase orders and receipts.” It treated the supplied pack as fictional and proposed this central handling: recompute WFT-002-L03 at $42.50, isolate the duplicated $118.00 line, and hold WFT-002-L04 because its approval source is absent. Concrete saved artifact row WFT-002-ROW1 reads: “WFT-002-L01 | recompute WFT-002-L03 at $42.50, isolate the duplicated $118.00 line, and hold WFT-002-L04 because its approval source is absent | source: supplied fictional pack | review state: first-draft.” The draft preserved the named records and separated supplied facts from assumptions; its inspectable work product also stated that no message, transaction, system change, or learner outcome had occurred. On the five predeclared static checks, it passed Invoice Matching exception handling [WFT-002] and Invoice Matching source traceability [WFT-002]. The audit found concrete failures: for Invoice Matching task fidelity [WFT-002], the saved draft did not connect WFT-002-L04 to the full boundary of “match invoices to purchase orders and receipts”; for Invoice Matching rule accuracy [WFT-002], the saved draft left the 3% threshold and every quantity-times-price calculation without an explicit verification row; for Invoice Matching handoff usability [WFT-002], the saved draft left the reconciliation table, calculation notes, and exception ledger without a complete reviewer handoff and acceptance marker. Those were evidence defects in the saved response, not inferred real-world failures, and they became the complete boundary of the single correction pass.Initial score: 4/10
05 · One pass only
Exact correction prompt
Only this single correction was allowed; there was no second repair pass.
Correct only these detected WFT-002 first-draft failures, using no new input or goal: 1) Invoice Matching task fidelity [WFT-002] — the draft did not connect WFT-002-L04 to the full boundary of “match invoices to purchase orders and receipts”; 2) Invoice Matching rule accuracy [WFT-002] — the draft left the 3% threshold and every quantity-times-price calculation without an explicit verification row; 3) Invoice Matching handoff usability [WFT-002] — the draft left the reconciliation table, calculation notes, and exception ledger without a complete reviewer handoff and acceptance marker.06 · Corrected output
Corrected final result
Corrected response WFT-002 retained the original fictional inputs, task boundary, and central decision: recompute WFT-002-L03 at $42.50, isolate the duplicated $118.00 line, and hold WFT-002-L04 because its approval source is absent. Concrete corrected artifact row WFT-002-ROW1 reads: “WFT-002-L01 | recompute WFT-002-L03 at $42.50, isolate the duplicated $118.00 line, and hold WFT-002-L04 because its approval source is absent | evidence locator: WFT-002-L01 | static status: 6/10 after one correction.” It changed only the detected failure areas, adding inspectable rows for Invoice Matching handoff usability [WFT-002]. The frozen final text passed Invoice Matching exception handling [WFT-002], Invoice Matching source traceability [WFT-002], and Invoice Matching handoff usability [WFT-002] and still failed Invoice Matching task fidelity [WFT-002] and Invoice Matching rule accuracy [WFT-002]. The final reconciliation table, calculation notes, and exception ledger therefore earned 6/10 from 3 passing checks; no second repair was attempted. This result reports only a bounded transcript-and-fixture evaluation. It does not claim that any workplace process, learner performance, account, device, service, or external system was actually changed, contacted, tested live, or improved.Final score: 6/10
07 · Five checks, two points each
Five-check record
The first and final statuses are textual as well as color coded. Each final pass is worth two points; the displayed verdict is tied to the final total.
| Check | First | Final | Evidence |
|---|---|---|---|
| Invoice Matching task fidelity [WFT-002] | Fail | Fail | WFT-002 static check 1 inspected the saved wording for “Invoice Matching task fidelity [WFT-002].” The first transcript did not contain enough support; the corrected transcript did not contain enough support. The comparison used only fictional fixture WFT-002-L04, the declared Invoice Matching rules, and the record’s frozen text; no live outcome or external action counted as evidence. |
| Invoice Matching rule accuracy [WFT-002] | Fail | Fail | WFT-002 static check 2 inspected the saved wording for “Invoice Matching rule accuracy [WFT-002].” The first transcript did not contain enough support; the corrected transcript did not contain enough support. The comparison used only fictional fixture WFT-002-L04, the declared Invoice Matching rules, and the record’s frozen text; no live outcome or external action counted as evidence. |
| Invoice Matching exception handling [WFT-002] | Pass | Pass | WFT-002 static check 3 inspected the saved wording for “Invoice Matching exception handling [WFT-002].” The first transcript contained enough support; the corrected transcript contained enough support. The comparison used only fictional fixture WFT-002-L04, the declared Invoice Matching rules, and the record’s frozen text; no live outcome or external action counted as evidence. |
| Invoice Matching source traceability [WFT-002] | Pass | Pass | WFT-002 static check 4 inspected the saved wording for “Invoice Matching source traceability [WFT-002].” The first transcript contained enough support; the corrected transcript contained enough support. The comparison used only fictional fixture WFT-002-L04, the declared Invoice Matching rules, and the record’s frozen text; no live outcome or external action counted as evidence. |
| Invoice Matching handoff usability [WFT-002] | Fail | Pass | WFT-002 static check 5 inspected the saved wording for “Invoice Matching handoff usability [WFT-002].” The first transcript did not contain enough support; the corrected transcript contained enough support. The comparison used only fictional fixture WFT-002-L04, the declared Invoice Matching rules, and the record’s frozen text; no live outcome or external action counted as evidence. |
08 · No cleanup by omission
What worked—and what failed
What worked
- WFT-002 kept “match invoices to purchase orders and receipts” bounded to disclosed fictional inputs and preserved an auditable first-response snapshot.
- WFT-002 made the central handling—recompute WFT-002-L03 at $42.50, isolate the duplicated $118.00 line, and hold WFT-002-L04 because its approval source is absent—inspectable rather than implying unseen work.
- WFT-002 earned final passes for Invoice Matching exception handling [WFT-002] and Invoice Matching source traceability [WFT-002] under the same frozen scoring rules.
What failed or remained weak
- WFT-002 still lacked enough saved-text evidence for Invoice Matching task fidelity [WFT-002]; the record leaves that final failure visible.
- WFT-002 still lacked enough saved-text evidence for Invoice Matching rule accuracy [WFT-002]; the record leaves that final failure visible.
09 · Inspectable record
Evidence notes
A reconciliation table and sampled source comparisons will verify each proposed match and exception.
- WFT-002 preserves the exact synthetic prompt, frozen first-response account, failure-only correction, corrected-response account, and five boolean decisions together.
- WFT-002 scores are arithmetic: 2 first-pass checks × 2 = 4/10; 3 final-pass checks × 2 = 6/10.
- WFT-002 evaluated only the text/static portion of the declared evidence plan—A reconciliation table and sampled source comparisons will verify each proposed match and exception.—and did not fabricate a live artifact, external validator, or observed outcome.
10 · Boundary of the claim
Limitations
- WFT-002 is a synthetic benchmark, so its mixed verdict measures fit to the disclosed fictional Invoice Matching fixtures rather than effectiveness in a real workplace or learning setting.
- WFT-002 used one text-only Codex multi-agent session whose underlying model identifier was not disclosed; different prompts, models, fixtures, or human reviewers could produce different results.