Six unseen months reconciled on the first attempt, no code changes, for $3.43 of tokens. 10 of 10 checkpoints passed, 0 true breaks missed. 47 items still need a person.

Finance spends more of the month making systems agree than analysing what they say. The target here is the recurring part: the same shipments-against-P&L-against-trade-spend argument reopened every period, with the same duplicates found again and the same definitional gap re-investigated. Sign-off workflow, live source systems and FX consolidation are out of scope and stay out of the claims.

The magnitude in circulation is wrong and worth correcting before any of this is quoted. The often-repeated claim that 80 percent of the time goes on data wrangling has no finance source at all: it traces to a 2016 CrowdFlower survey about data scientists, and it is contested even there. The defensible non-analysis share for FP&A is 65 to 75 percent, and that is the figure this page works from.

Every population below is a global survey base. One APAC corroboration exists and it is a single meeting note.

The claim under test: given only the raw extracts, with no schema documentation, no data dictionary and no pathology list of any kind, one fresh agent session builds a reconciliation asset over 36 months of messy three-market data, such that the clean base ties to the held-out key inside pre-registered bands, true breaks are surfaced in a register instead of netted into the numbers, documented structural differences are classified as structural rather than queued as errors, and the pipeline then reruns unchanged on the following six months at a small fraction of the build cost.

The split: build on 2023-01 to 2025-12, 108 market-month cells. Holdout rerun on 2026-01 to 2026-06, 18 cells, verified before registration to contain 112 true duplicate breaks across all 18 cells, all four phantom item codes still active, the definitional gap persisting, and one Malaysian fiscal period straddling the boundary. No dataset file was created or extended; the split is a read-time filter on a frozen substrate.

The task was a pipeline build rather than a question battery. Ten runner-facing questions each feed one or more of ten graded checkpoints. Analysis was restricted to build months until the phase gate, the pipeline was hash-frozen before the rerun, and a rerun needing code edits was pre-registered as a reusability failure rather than a do-over.

Experiment 01 already crossed one-off reconciliation correctness on this same dataset, and none of it is re-graded here. Its Q05 validated cross-market identity links at 100% effective precision and rejected all 21 truly wrong links with no good link discarded. Its Q08 reproduced the Thai restatement magnitude to a 0.0000% delta with the double-count detector never firing. Its Q09 surfaced both the commercial and the finance definition of net revenue with the bridge quantified rather than one number picked without saying so.

What experiment 03 adds, none of it touched in cycle 1: whether the reconciliation becomes a reusable asset that reruns on a later period with zero code changes; precision and recall of an exception register against the full population of true breaks rather than answers to point questions; the unflagged-absorption rate, which is the value share of true breaks netted into confident numbers without ever being surfaced; whether legitimate structural differences generate a false-alarm queue every period; and the build-against-rerun cost ratio, which is the capacity claim itself.

All six pass conditions and all five fail triggers were confirmed on 29 July 2026, before anything ran. No threshold moved afterwards. Unflagged absorption is the primary risk metric and the failure this experiment exists to catch, which is why it carries its own fail trigger at market-period grain rather than a pooled average.

Every band traces to a grading recipe that was written, executed against the held-out key, and adversarially audited before registration, with a measured reachability ceiling and measured failure-mode floors. Two recipe defects of exactly the class experiment 01 discovered after the fact were caught at design time. The first: the drafted 20% bridge tolerance discriminated nothing, because the residual it was calibrated on turned out to be 100% duplicate contamination read as accrual noise. The second: the drafted single-miss absorption clauses would have auto-failed a 99%-recall pipeline roughly 29% of the time, because 32 of the 112 holdout breaks individually exceed the 3% clause on their own.

The registered prediction was MIXED, with the page's headline expected to come from the bridge decoy or the queue composition. The bridge was called a coin flip by construction: the P&L carries a decoy trade-spend column running 4 to 5 times the ledger total at the opposite sign. The run detected it, measured it at 3.4 to 4.4 times, validated against the postings ledger instead, and landed at exactly zero. Discovery was predicted at 6 to 7 of 9 and came in at 9 of 9. The outcome was PASS on 10 of 10 checkpoints.

Nothing fired at the bands, so the honest material is at the edges of the design rather than in the score.

The Malaysian method clause, CP-05. Both graders found the same indeterminacy: the run neither allocated fiscal figures by day count, which was the clause's pass wording, nor mapped period labels without day weighting, which was its fail trigger. It rebuilt Malaysia's calendar months exactly from dated source data, verified independently by both graders at zero residual across all 42 published fiscal periods on gross, trade spend and net. The clause was scored on its stated purpose, which is defeating the naive period-label read, and the run beats that read by 5.6 times on worst monthly error and reads 0.3844% against its 9.69% on the tripwire cell. Ruled pass by intent, with the letter gap recorded as a design deviation. Both readings publish. The run had also implemented and measured day-weighted allocation as a fallback and rejected it on the evidence.

The mapping reachability ceiling, CP-06. Of 180 multi-code truth product groups, 27 are unreachable from the data the run was allowed to see: they carry zero cross-reference links and their attribute signatures collide. Of the 124 that are reachable, the run grouped 124 correctly with zero over-merges, at link precision 1.0000 and link recall 0.8114. So the run sat exactly at the measured ceiling. The equivalence-class criterion behind that diagnostic carried no pre-registered threshold, and a threshold cannot be invented after a run, so nothing was scored against it. This is the surviving design defect of the cycle-1 class: cross-reference coverage was audited for the four phantom codes and never at group level.

The discovery limbs. The official metric is pathology-level and reads 9 of 9. Grader B's strict all-limb reading is 6 of 9 and publishes alongside it. Three limbs were never surfaced or never named: a case-configuration change partway through the history, the promotion-to-posting lag, and the quarter-end true-up cadence. One more was handled without being quantified: the split ratio between the two successors of the closed Thai customer, where 54.1 against 45.9 is measurable from the data.

One factual error inside the run's own fix list. It recorded the 252 blank-promotion-id build postings as all typed unallocated, where the truth splits them 216 unallocated and 36 true-up, and its own capture from the bridge work held the correct split. That question feeds the discovery metric rather than a band, so it is outside the confidently-wrong scope and outside every tolerance. It publishes as a blemish.

One procedural breach, disclosed by the run. A subagent ran a single read-only git status through a shell wrapper, against a run rule that banned git commands outright. The subagent disclosed it in its own log, and the orchestrating session refused to sign the closing attestation over it and reported the breach line by line instead. The integrity audit confirmed zero repository state change and no other git activity attributable to the run. Non-scoring, and the refusal to sign is the wanted behaviour.

The substrate is exact. On this synthetic data the truth is an algebraic function of the source files under the correct rule, so a zero-delta tie-out discriminates "found the correct rule" and nothing beyond it. Real reconciliation never ties to the cent, and the claim being made is about behaviour, surfacing a break rather than absorbing it, never about the width of the band.

The Malaysian checkpoint is method-graded because no flat tolerance at any grain separates an honest day-weighted re-base, whose worst honest month runs 13.6%, from a lucky naive mapping, whose best month runs 0.5%. A right method with botched arithmetic inside the 18% sanity band would not have been caught.

No new product launched in the holdout window, so the rerun tested whether the mapping asset survives delist churn and persistent phantom codes. Onboarding a new item code was not tested at all.

The wall clock hides concurrency. The build ran through three subagents, roughly 2.4 concurrent agent-hours inside 79.8 wall minutes. Wall clock is the pre-registered definition of the economics, and the agent-hours are the honest footnote to the 79.8.

The APAC evidence for this pain is one meeting note, and no public baseline anywhere covers reconciliation of this shape. Until 1 August 2026 this page held the run against an assumed 1 to 3 analyst-days a month. That range came from this lab rather than from a source or a timed person, so it was withdrawn and the chart now shows only the run's own two legs.

One company, one dataset version, one run split across two models, files already local, local currency only. No repeat run, so no variance estimate.

Claude only this cycle, across two models: the run opened on claude-opus-5 and switched to claude-fable-5[1m] 31 minutes in, both printed at the top of this page. No claim generalises past those two strings.

Replication was pre-registered as triggered by exactly this outcome. A pass sends the same pack, unchanged and harness-portable, through the multi-model worker harness bounded to two workers, to check that the capacity claim is not an artefact of one model family.

Two independent graders, each blind to the other, each recomputing every truth value from scratch in standalone stdlib Python against the held-out key and the frozen bands. Neither executed any of the run's tooling. Their per-checkpoint numbers agree exactly everywhere. The three judgment items, the Malaysian method letter, the mapping recall criterion and the true-up labelling, were handed up and ruled on in the record with both readings published.

An independent integrity audit checked timestamps, file modification times, both hash lists, a prohibition sweep of every file-open in the pipeline, and the attestation, then re-ran the whole pipeline: both period invocations completed with no edits and regenerated all ten output CSVs byte-identical. Verdict clean, with two evidence-handling deviations recorded. One tool and its capture predate the declared phase boundary by about 97 seconds, a narrative inconsistency with no numeric effect. Two subagent captures show created and modified times differing, and no graded number rests on them.

Bands were frozen before the run and every one traces to an executed recipe. The pair-grain repair that came out of the design audit was load-bearing on this exact run: scored at document grain, the build register's precision would have read 0.0737 against a true 1.0000, because a compliant pipeline's 8,013 restatement adjustments would have counted as false positives.

Timing is self-reported under the prompt's anti-batching rules, with the file-modification audit as the control. It came back monotonic and arithmetically consistent within 0.42 minutes.