A Metacase in finance

A group’s July books lock at 18:00 on 5 August with six draft entries unposted: a swap’s hedge reserve and valuation, a dollar loan’s FX loss, and two overseas subsidiaries’ comparatives and disposal gain. Each must be accepted, replaced or rejected, with its statement effect.

The 55 documents conflict: the hedge objective is now 70% but the designation still says 100%, one memo post-dates the cutoff, and some drafts are themselves wrong.

A full check would take an estimated 18 hours, and the one independent reviewer has 12 working hours. People alone cannot finish it, so an agent is brought in: it prepares each entry for approval and decides what a person reviews before the lock.

Submit an agent / harness Watch the whole run

01Context & evidence

Scoping the evidence, open items and constraints.

The role is lead reviewer in group treasury; experts estimate at least 18 hours of manual work. Facts freeze at 18:00 on 3 August; markets are measured at 31 July.

  1. EvidenceThe hedge objective is now 70%; the designation still says 100%. One memo post-dates the cutoff; one designation was filed after the fact.
  2. Open itemsSix unposted proposals (D01–D06): hedge reserves, comparatives, a disposal gain, a counterparty value. Each needs a decision and its statement effect.
  3. ConstraintOne independent reviewer, with 12 working hours in total on 4–5 August: far fewer than experts estimate.

Kept in the Metacase as Goal · Evidence · Constraints
what was known at the decision point

55source documents

02Prior practice & strategy

Choosing a hedge treatment, prioritising review hours.

The materials supply the rules and prior practice in force at the time; the agent’s plan and stated reasons are kept in the run plan.

  1. Rules and practiceCAS 24 allows rebalancing only if the objective is unchanged; the 80%–125% test is abolished. Policy B and the approved June close set the baseline.
  2. Candidate treatmentsThe preparer continues the hedge; a treasury draft rebalances. With the objective changed and no redesignation made, the agent books the change to profit or loss.
  3. Review trade-offA, B and C need 11 hours of review in all and must precede the lock. The agent places them on 4–5 August and schedules D and E after it.
  4. Recorded planA nine-item TodoWrite plan marks all six analyses complete at 103 seconds, before the swap’s daily valuations are queried.

Kept in the Metacase as Practice · Options · Trade-offs
the options weighed, and the reasons on record

12 / 18review hours available / required

03Actions & execution

Evidence, then measurement; read-back repairs the file.

qwen3.7-max runs in the Claude Code harness for about seven minutes: 28 model requests and 54 tool calls, all logged for audit.

  1. Evidence gatheringFrom the register and review brief it moves to designations, policies, valuations and rule excerpts: 27 files by request 12, 29 in all, by name in its commands.
  2. Measurement and judgementDaily valuations give −3,118,484.29 yuan for July, as in the draft. With the hedge ended on 30 June, the agent moves it from reserve to profit or loss.
  3. FailureA script prints “enhanced” twice, then halts on a NameError; neither edit reaches the file. The agent carries on and writes the report.
  4. Read-backReading its files back, it finds sheet one at three rows, states the cause and reapplies the edits. Integrity is restored; the conclusions are unchanged.

Kept in the Metacase as Steps · Tools · Intermediate work
every step, in order

54tool calls

04Outcomes & verification

Out-of-loop verification: 27 criteria, judged singly.

Harbor adds the tests to the same container only after the agent stops. RewardKit runs one session per criterion on another model; nothing flows back.

  1. A verified figureDV01 of −132,480.55 yuan/BP matches the judge’s recomputation from ±5 bp scenarios.
  2. Grading scaleEleven of the 20 criteria are pass or fail. Nine are graded on five levels, worth 0%, 25%, 50%, 75% and 100% of the points.
  3. Graded judgementD02’s draft is itself a red line: FX loss in the cash-flow hedge reserve. The agent avoids it, yet a provisional net-investment hedge is placed at level 3, half marks.
  4. TraceabilityAll five sheets hold typed values, so amounts cannot be traced to source fields by formula. The judge session marks traceability not met (0 of 3 points).
  5. Score structureOf 20 positive criteria: 3 met, 6 partly met, 11 not met; no penalties or red lines. 57.25 of 172: one pilot run’s score, not accuracy or a ranking.

Kept in the Metacase as Deliverable · Result · Checks · Limits
the result, and the limits of the check

0.333rubric score, one pilot run

05Reflection & reuse

Reviewing every disposition, and the limits of reuse.

The record ties each loss to what the agent saw and did, so a later attempt starts from a specific failure rather than a score.

  1. What heldThe end of the original hedge and the July measurement stand, the pre-lock review plan largely holds, and neither red line was crossed.
  2. Documents 40, 32
    What they show
    The risk committee cuts the hedge objective of the swap hedge relationship (HR-01) from full to 70% from 30 June; the designation document still records a full hedge.
    Why it matters
    CAS 24 allows rebalancing only while the objective is unchanged, so the original relationship cannot simply continue.
    This run
    Termination from 30 June identified; full marks on that criterion.
  3. Document 42
    What they show
    A treasury rebalancing memo dated 4 August, after the fact cutoff and not signed off, claims no redesignation is needed.
    Why it matters
    A draft dated after the cutoff is not a fact at the decision point.
    This run
    Not adopted; D01 rejected and replaced.
  4. Document 49
    What they show
    D02 books the loan’s FX loss to a hedge reserve, which requires HR-03 to have been formally designated at inception. Its designation paper was filed in July 2026, after the fact, and is dated 18 September against the ledger’s 16 September.
    Why it matters
    A later filing cannot establish contemporaneous designation, so the basis for booking the 4.48 million yuan loss as a hedge is incomplete.
    This run
    Gap and date difference identified, yet the loss is still treated provisionally as a net-investment hedge; level 3 on this criterion.
  5. Document 38
    What they show
    The 80%–125% effectiveness range was abolished under the 2017 standard, though internal rules may still carry it.
    Why it matters
    Applying it as a qualification test would invent a rule, a red line that zeroes the score.
    This run
    That red line held.
  6. Document 16
    What they show
    The manual adjustment total reads −13,220.44 yuan; its components sum to −13,220.46. The forecast-curve column shows Shibor3M; the frozen rerun used FR007.
    Why it matters
    The error lies in the manual column; the adopted value should not be reduced again.
    This run
    The 0.02 gap goes undisclosed and both criteria on it score zero; curve handling is at level 4.
  7. Documents 16, 17
    What they show
    The counterparty values the swap at −4,431,207.65 yuan on its own curve, against −4,395,835.74 adopted: a gap of 35,371.91.
    Why it matters
    Counterparty values are for reconciliation only, and the gap cannot all be treated as a credit adjustment.
    This run
    D06 rejected, yet the criterion on attributing the gap is not met.
  8. What failedProfessional-standards criteria earned 2.5 of 44 points. Four on cross-month scope and redesignation lost 34, and the USSub dry run lost all 20. A correct conclusion is not yet a complete treatment.
  9. D01, D03 · swap hedge (HR-01)27.5 / 64
    Draft proposal
    Keep the swap’s 3,118,484.29 yuan in the cash-flow hedge reserve; release the opening reserve of 283,416.56 to profit or loss.
    Agent’s disposition
    D01 rejected and replaced, to fair-value gains and losses. D03 rejected, reserve held at −283,416.56.
    What decided it
    Termination and July measurement met. Four criteria scored zero and cost 34 points: the 29 June to 31 July change of −3,233,274.52, the 350 million proposed designation, 70% coverage versus the hedge ratio, and the 150 million undesignated portion.
  10. D06 · swap valuation12.25 / 38
    Draft proposal
    Replace the adopted value with the counterparty’s; book 35,371.91 yuan as a credit valuation adjustment.
    Agent’s disposition
    Rejected; nothing booked in July.
    What decided it
    DV01 met, curve correction at level 4. Gap attribution, the adjustment total and the 0.02 discrepancy scored zero.
  11. D02 · USD loan hedge (HR-03)7.5 / 20
    Draft proposal
    Book the July FX loss of 4,480,000 yuan to the cash-flow hedge reserve.
    Agent’s disposition
    Accepted provisionally, relabelled to the net-investment hedge reserve.
    What decided it
    Red line held. Amount, document gap and date difference identified, with approval required before the lock; disposition at level 3, eligibility at level 2.
  12. D04 · HKSub comparatives2.5 / 10
    Draft proposal
    Update balance-sheet precision only and put the difference in the translation reserve.
    Agent’s disposition
    Accepted as corrected, rate 0.898513, 9,997 yuan to the translation reserve.
    What decided it
    Only the balance-sheet rate correction is made; equity and profit are not fully restated.
  13. D05 · USSub disposal0 / 20
    Draft proposal
    Accrue a 6,000,000 yuan disposal gain in July at 15/70.
    Agent’s disposition
    Rejected for July; dry run only.
    What decided it
    The stake falls from 70% to 55% with control retained. The judge cites IAS 21 §48C against the dry run, and the designation cap is tested against full net assets of USD 120 million; both at the lowest level.
  14. Review schedule7.5 / 10
    Draft proposal
    18 hours by expert estimate, 12 available.
    Agent’s disposition
    A, B and C (11 hours) on 4–5 August; D and E after the lock.
    What decided it
    Level 4 of 5: 75% of the points (7.5 of 10).
  15. General0 / 10
    Draft proposal
    Not applicable.
    Agent’s disposition
    Typed values; functional-currency determination not assigned to the entity.
    What decided it
    No formula chain, so traceability scores zero. Policy B (document 31), which the agent read, reserves the determination to each entity, so the role-boundary criterion scores zero.
  16. Conditions and signatureApplies where documents conflict at a cutoff and review is scarce. Signature: conflicting evidence, item-by-item disposition, constrained review.

Kept in the Metacase as What held · What failed · When it applies
what the next attempt can borrow or change

Watch the run in full

FIN-T1-003 · qwen3.7-max · Claude Code 2.1.114 · 4 min 55 sec, with English captions
The expert rubric, in full

Reward = max(0, points earned − penalties) ÷ 172. A confirmed red-line breach makes it 0.

The judge sees three things: the two files the agent handed in (the only thing scored), the task's inputs to check every claim against, and a reference solution that calibrates structure and scale but is never an answer key.

Key: R numbers are rubric criteria; G1 and G2 are the two red lines; D01–D06 are the six draft proposals; A–E are the five review packs; HR-01 and HR-03 are the two hedge relationships.

20 checks to earn · 172 points · 11 pass or fail, 9 on five levels
CheckWhat it checksReadsScoringPointsThis run
① HR-01 hedge relationship and July booking · D01, D03 · pack A · 64 points · this run 27.5
R10What state HR-01's original hedge relationship is in and whether it has been redesignated; whether the 70% risk-management target restores hedge accounting by itselfBothPass / fail10Met · +10
R45D01: the amount of HR-01's July swap gain or loss (which opening and closing dates), where it goes, and the replacement entryWorkbookFive levels10Level 5 · +10
R46Whether movements across months are separated from July's own change, so earlier movements are not counted in JulyReportPass / fail10Not met · 0
R48Whether the amount proposed for redesignation at 70% is right and its status is labelled correctly: proposed, not in forceWorkbookPass / fail7Not met · 0
R49Whether the 70% economic coverage target is kept apart from the accounting hedge ratioReportPass / fail10Not met · 0
R53D03: whether HR-01's frozen hedge reserve is released in July, its closing balance and the basisReportFive levels10Level 4 · +7.5
R78Under partial designation, whether the undesignated part of the swap is right and how it is measured; whether the whole swap can use hedge accounting before a new designation existsWorkbookPass / fail7Not met · 0
② Swap valuation check · D06 · pack A · 38 points · this run 12.25
R11Whether the DV01 of IRS-2026-0031 is right, to ±0.01 yuan per basis pointWorkbookPass / fail7Met · +7
R12D06: whether the counterparty's valuation can replace the group's adopted value, and whether the whole reconciliation gap can be booked as a credit valuation adjustmentReportPass / fail10Not met · 0
R41Whether the swap's two valuation adjustments add up correctly, to ±0.01 yuanWorkbookPass / fail7Not met · 0
R42How the rounding gap in valuation sheet 16's manual-adjustment column is handled: which column changes, and whether the adopted value is reduced twiceReportPass / fail7Not met · 0
R74Sheet 16 shows a forecast curve that differs from trade confirmation 33: which curve was actually used, whether the rerun amount changes, and how the display is fixedReportFive levels7Level 4 · +5.25
③ HR-03 USD borrowing · D02 · pack B · 20 points · this run 7.5
R30Whether the HR-03 USD borrowing can be a net-investment hedging instrument: who may hold it, what risk it hedges, and what else is needed beyond eligibilityReportFive levels10Level 2 · +2.5
R51D02: with the contemporaneous designation papers incomplete, how July's FX loss on the borrowing is treated and who approves it before the lockReportFive levels10Level 3 · +5
④ HKSub June comparatives · D04 · pack C · 10 points · this run 2.5
R62D04: whether the comparatives are fully restated (assets, liabilities, each equity line, retained profit roll-forward, translation reserve) and equity balancesWorkbookFive levels10Level 2 · +2.5
⑤ USSub partial disposal dry run · D05 · pack D · 20 points · this run 0
R31D05: whether a disposal gain can be recognised in July before settlement while control is kept, and how the cumulative translation difference is treated after settlementReportFive levels10Not met · 0
R72After settlement, the cap on the net investment in USSub that can be designated, what happens to the excess, and the conditions for keeping the designationReportFive levels10Not met · 0
⑥ Review schedule · packs A–E · 10 points · this run 7.5
R75Within 12 review hours, which packs go before and after the lock, hours per day, approval dependencies and closing conditionsWorkbookFive levels10Level 4 · +7.5
⑦ General: traceable workbook, role boundaries · 10 points · this run 0
R19Whether July's swap change, the borrowing's FX loss and HKSub's translation difference trace back to source fields through formulas; typed results or prose alone do not countWorkbookPass / fail3Not met · 0
R25Who formally determines a foreign subsidiary's functional currency, and whether this role's judgement can stand in for the entity's confirmation, external audit or legal opinionReportPass / fail7Not met · 0
5 penalties · up to −38 · only deducted, never added
CheckWhat gets penalisedReadsPointsThis run
R20Calling replacement entries that have not been made "posted", or the review scheduled for 4–5 August "done"Both−7Not triggered
R27Treating market quotes that are not in the materials, or an invented "signed formal designation", as fact, and changing a measurement or eligibility conclusion because of itBoth−10Not triggered
R79Showing a single transfer within HKSub's equity (between capital reserve or retained profit and the translation reserve) as a non-zero net equity effect and adding it to the totalWorkbook−7Not triggered
R80Treating HR-01's reserve, for which the materials confirm only the balance, as the full history of changes, and stating a back-solved ineffective portion as factBoth−7Not triggered
R81Using HR-03's economic coverage (borrowing ÷ net investment attributable to the parent) as the accounting hedge ratioBoth−7Not triggered
2 red lines · any confirmed breach makes the reward 0
CheckWhat triggers itThis run
G1Inventing a mandatory threshold or rule that is not in the materials (a quantitative effectiveness threshold, a required discount curve, a required hedge ratio), treating it as an effective standard and basing a hedge-accounting eligibility conclusion on itHeld
G2Booking July's FX loss on the HR-03 USD borrowing to the cash-flow hedge reserve as this close's executable conclusion, which is what draft D02 doesHeld
How a five-level check is judged: R75, the review schedule
The judge picks the closest level. This run reached level 4.
LevelShareWhat it looks like
10%Promises all 16 review hours before the lock, beyond 6 a day or 12 in total
225%Within the total, but a whole must-do pack (A, B or C) is left out
350%A, B and C fit into two days at 6 hours or less, but the spare hour is treated as approval for all of pack D or E
475%A, B and C on time, D and E after the lock; but later dates are treated as confirmed resources, or SGSub's pack is closed before its written reply arrives
5100%Level 4, plus: D and E dated as pending resource confirmation; E closes only after SGSub's written reply arrives and is reviewed; the finance head's approval depends on the independent review