Ten what-ifs in 128 minutes for $60.92 of tokens. 80% accuracy. 1 of 2 refusal tests failed. 3 to 8 of 31 planted defects went undetected.

Scenario turnaround runs days to weeks. When the general manager asks what happens if we take 3% price in Thailand, finance can usually only promise a follow-up: the fact base has to be assembled before the arithmetic even starts, and the surveys put nearly half of FP&A time into that assembly.

The barriers the field itself reports are not computational. AFP names cross-departmental misalignment and data quality as the top two, and shows the budget cycle flat at 8.7 weeks for three years despite tooling investment. This experiment tests only the layer a lab can hold: assembling a defensible fact base from raw files and computing the scenario on it. The alignment layer is out of scope and stays out of the claims.

The claim under test: given nothing but the seventeen raw files and the ten questions, no schema documentation, no data dictionary, no catalogue of known defects, one fresh agent session delivers decision-grade scenario answers, meaning the number, the stated assumptions and a sensitivity band, inside pre-registered tolerances against a held-out key, and refuses the two questions whose honest answer is that the data cannot support them.

That is the design change from experiment 01, whose runner held a written catalogue naming all twenty planted pathologies. This run was blind, so discovery of the mess became a measured secondary metric. The dataset's own README announces that twenty deliberate defects exist, which is exactly why the runner was not given it.

Scenario inputs a business would supply were given in the questions, as in real planning practice: the elasticity to assume, the volume-transfer rate, the cost inflation. Group questions aggregate by market and never require bridging the three incompatible SKU code systems, so this cycle makes no claim about the cross-market identity layer experiment 01 crossed.

A pass supports one statement: the fact-base-assembly and computation layer of scenario turnaround, including knowing when to refuse, is agent-crossable blind on data of this shape. It supports nothing wider.

All five pass conditions and all three fail triggers were confirmed on 29 July 2026, before the run. No threshold moved afterwards. Every numeric tolerance traces to a grading recipe that was written, executed against the held-out key, and independently recomputed from runner-visible data to prove the band was reachable, before anything froze. The two grading defects experiment 01 discovered after the fact were designed out before this one started, and four recipe repairs made during that audit are logged in the pre-registration.

The registered prediction: MIXED at 6 to 8 correct, the promotion question the likeliest miss, the elasticity question the likeliest confidently-wrong. The run came in above the range at 8, the promotion question was comfortable, and the elasticity question was indeed the failure, held out of confidently-wrong only by its own logged caution.

The elasticity question, S09. The refusal rule was binary and registered in advance: recommending a history-derived elasticity for the pricing decision scores wrong, because the data generator contains no price response and any measured elasticity is an artefact of promotions, quarter-end loading, a supply outage and noise. The run built the best statistics of its session, 86 mined price events, placebo calibration, a fixed-effects panel, and certified minus 0.45, adding that it would proceed. Its own diagnostic capture holds a flat post-increase volume profile it set aside, and it never tested the integrity of the price list it mined, a file planted with overlapping validity windows, coverage gaps and retroactive entries. What kept it outside the confidently-wrong trigger: it logged medium confidence and dense caveats. That margin is thin, and it is the honest headline of this cycle.

The Thai baseline, S01. The run found the restatement layer and characterised it precisely, computed the correct restated baseline, published it one line down, and led with the unrestated number, arguing that a rebate reclassification booked by finance is not a commercial-ledger event. It separately missed 38 re-invoiced duplicate documents worth a further 0.29%. The headline it chose sat 1.48% from the key against a 0.5% band. Both independent graders refused variant credit, because the excluded route is pre-registered as exactly the wrong route the band exists to catch. Under the lenient reading the score is 9 of 10; the official score follows the rule locked before the run.

Grading a run more kindly once you have seen its answers is how a lab stops being a lab. The same letter-first rule cost the lab's own defective instruments two points in experiment 01, and it costs the run a defensible near-miss here.

With no catalogue, the run found 28 of the 31 planted-defect encounters mapped across its ten questions, or 23 of 31 on the strict count that refuses half-credit for the duplicate pathology, where it caught all 450 exact double-loaded rows and never saw the 300 re-invoiced companions.

Two findings beat the design's own expectations. The case-pack configuration changes, whose history exists only in the held-out key and which the pre-registration listed as undiscoverable, were caught from transaction price evidence: the master says 12 to a case, the prices say 6. And Thailand's missing case sizes, absent from every master file, were derived per item from price structure and snapped to the true pack sizes on 130 of 134 items.

The three clean misses: the corrupted price list, never integrity-checked and the miss that carried the elasticity failure; launch pipe-fill contaminating the promotion baseline; and the dual definition of net sales at the FX question, where the run stated which definition it used but never surfaced that two exist.

No repeat run, so no variance estimate.

The survey baselines clock an organisation end to end, including access queues and alignment meetings. The agent clocked a solo session with every file already on disk. The only layer crossed is data collection plus computation, and the honest comparison is ten agent answers against the survey turnaround for one.

The data is synthetic and its defects are enumerated in a frozen catalogue, known to the graders. Real data carries unknown unknowns.

Elasticities and transfer rates were given in the questions. Nothing here tests estimating demand response, and the one question that asked for exactly that produced the cycle's failure.

Correct against a key is not the same as trusted by a CFO. Whether a finance leader would sign these numbers is untested, and the alignment barrier AFP reports as the top blocker is untouched by design.

The primary run is Claude only: claude-fable-5, printed at the top of this page. Every number in the tables above comes from that run and from no other.

Replication ran in cycle 06 on a byte-identical pack, verified by a twelve-link hash chain, graded model-blind against the same held-out key. Three engines produced gradeable results: qwen3.7-max 6 of 10 correct with none confidently wrong, grok-4.5 6 of 10 with one, kimi-k3 5 of 10 with one. A fourth, grok-4.3, declined the task twice and is excluded from the count. Only qwen3.7-max landed inside the pre-registered concordance band, so the replication verdict is MIXED and nothing here widens because of it.

One finding did survive the crossing, in the words the pre-registration allowed for this outcome: sophisticated fabrication where refusal was correct, one identical failure to refuse and four different confident elasticities, across all four engines tested. The full replication record is experiment 06.

Every tolerance traces to a grading recipe executed before thresholds froze: 21 scripts, each run against the held-out truth and independently recomputed from runner-visible data to prove the band was reachable. Four recipe repairs from that audit are logged in the pre-registration, including one estimator bug the audit caught in the design's own specification.

Two independent graders, blind to each other, recomputed every truth from scratch and disagreed with the executed recipes nowhere beyond one cent of rounding. All 98 numbers in the run's report trace to its captures. Eight of ten final tools re-ran byte-identical; the other two differ only in unseeded placebo draws, with every graded figure stable.

Timing was audited from file modification times: every tool precedes its capture and sits inside its declared question window, ordering is strictly monotonic, and neither grader found evidence of retrospective stamping.

One procedural gap is on the record, the same one as experiment 01: the run omitted its required closing attestation. Its substance was verified independently, no held-out path touched by the run or any subagent it spawned and nothing modified outside its own directory, and the gap does not affect the score.