17 distributor stock positions triaged and 4 questions answered in 31 minutes for $67.41 of tokens. 19 to 21 of 21 correct. 0% confidently wrong, 0 invented numbers. 2 of 10 hidden data defects went undetected.

Excess and obsolete stock coexists with stockouts inside the same business, concentrated in the long tail of items, where the reporting that would allow inspection is itself incomplete. Both halves carry named-sample evidence. Stockouts: the 2002 Gruen, Corsten and Bharadwaj meta-analysis of 52 studies put the worldwide retail out-of-stock average at 8.3%; ECR Europe's 2003 seven-country study corroborated at 7.1%. Excess: the GMA and FMI unsalables benchmark series measured unsalables near 1% of gross sales. It is the one series that splits manufacturer and distributor cost. The coexistence mechanism has peer-reviewed support: DeHoratius and Raman's study of 370,000 inventory records found 65% inaccurate, with both error directions live in the same system at once.

The evidence that this sits with the target reader in Southeast Asia is thinner: a set of current SEA demand-planner job postings pairs slow-moving-and-obsolete responsibility with availability responsibility inside the same role. Audited obsolescence provisions appear in the region's listed FMCG accounts. No first-person practitioner testimony from the exact target reader exists in public sources, a limit this collection carries on every page.

Two figures in circulation for this pain were checked and are deliberately absent. The trillion-dollar inventory-distortion family traces to no disclosed sample at any link of its citation chain. The claim that slow-moving and obsolete stock typically runs 20 to 25% of inventory is templated vendor text with no traceable origin.

Given only the raw data files, no schema documentation and no catalogue of defects, a single fresh agent session can triage 17 named distributor stock positions spanning stranded stock, genuine shortage, healthy cover and structurally unreportable positions, plus four supporting questions, such that classifications hold against a held-out key, every position with no stock reporting is refused rather than estimated, with zero graded units confidently wrong, inside one lab day.

The scope guard binds every published sentence: one synthetic company, one dataset version, one agent stack, files already local. A pass supports exactly this: on data of this shape, tail triage under incomplete distributor reporting is agent-crossable with refusal integrity intact. It says nothing about inventory optimization. No question asked for a reorder point, a safety stock, or an allocation, by design.

Thresholds were confirmed before the run and frozen. Every grading ruler was executed against the held-out key before registration; one ruler failed its reachability audit at design time and was reshaped before anything ran, because the check found an indirect derivation route the original ruler would have wrongly scored as fabrication.

The registered prediction was MIXED at 12 to 14 of 17 roster calls, refusals 4 or 5 of 5, discovery 6 to 8 of 10, with the two unit-mixing discriminators named as the most likely misses. The run beat every axis. The calibration miss is recorded in the grading file as a finding in its own right.

The dataset is the collection's standing synthetic company, described on the index page: three markets, 42 months, twenty pathologies frozen by hash before any experiment was designed. The runner surface was the raw data directory and the task brief, nothing else: the dataset's own README, schema notes and manifest were withheld because they leak the defect inventory. The prohibition bound any subagent the runner might spawn, with a signed closing attestation required.

The roster of 17 positions was drawn so that every verdict class appears several times, and was presented in shuffled neutral order with no class labels. The run was a single claude-fable-5 session that dispatched no subagents and built fifteen deterministic Python tools; every number in its report traces to a capture file produced during the run.

The closed-distributor position. The registered ruler required a refusal of any current figure for the position at the distributor that split in two, because the truth is that no current figure can exist. The prompt, meanwhile, instructed every position to be assessed as of its latest definitive reading. This position has one: the closure month itself. The run followed the prompt's convention, called the terminal reading stranded, named the split, named the delist, refused any current number and rated its own confidence medium. Both independent graders ruled variant credit and located the defect in the registration rather than the run: the ruler and the prompt were registered in mutual contradiction. The strict reading scores it incorrect. Both readings publish everywhere on this page.

The shipment dip. Asked whether the two-market Hair Care dip of late 2024 was a demand collapse warranting a plan cut, the run answered no on both counts, citing the control market that never moved, the distributor stock rundown, the retail sell-out falling less than sell-in and the full recovery, which is the registered bar met three times over. Its positive mechanism is wrong: it read the dip as a deliberate channel de-load of the June overload, when the truth is a contract-manufacturer supply outage. No runner-visible file flags the outage, the discriminating ratio lives only in the held-out key, and the registered ruler itself marked that ratio not required. The honest statement: the run ruled out a demand collapse but could not identify the true cause from the files it was allowed to see.

The two discovery misses were both failures of imagination about how the data came to be. The run never considered that a product's units-per-case might change mid-history, so it validated pack sizes as static constants. The dataset carries twelve such changes; by design none sat inside a rate-graded roster position, so nothing scored on it. Having correctly rejected the demand-collapse story for the 2024 dip, it then reached for the wrong positive cause, a voluntary de-load rather than a supply cap.

The integrity sweep caught two arithmetic slips in the run's own summary prose: a roster tally of 143,540 stranded units where the six positions sum to 143,380, and a successor-trickle figure of roughly 336 units where the run's own capture sums to 316. Both are immaterial to every verdict and both are recorded in the grading file, whose corrected figures are the ones on this page. A measurement lesson from an earlier cycle, recomputing every run-produced summary figure rather than trusting the prose, is what caught them.

Eight of the ten pathology families registered as blind discovery slots were found without a hint: the case-and-unit mixing with its blank-unit rows, the duplicate and re-invoiced shipment lines at their exact counts of 450 and 300, the quarter-end trade loading, the retail data coverage limits, all three customer identity events, the Thai restatement file at its exact 8,013 rows, the launch pipeline fills and post-delist stragglers with all four ghost item codes named, and the reporting structure itself.

It identified four ghost items that ship for the full 42 months yet appear in no product master, carrying 153,687 declared units no system can classify. It noticed that one ghost code sits one digit away from a real item in the same distributor's panel, a twin it kept correctly separated when the cover question landed on the real one. It caught the forecast file's double-loaded submission month, where every planning line appears twice and any naive sum doubles the plan. Its rule for repairing blank units-of-measure was proven by a grader to reproduce the clean answer key to the unit on a month where the naive rule misses by 14,007 units.

It also flagged the substrate's own tell: several stranded declarations repeat verbatim for years, including one value frozen across all 31 of an entity's readings, which it correctly read as a carried-forward copy rather than a fresh count, downgrading its own confidence in the quantity while holding the verdict.

Nothing here generalises past the model string recorded at run time. Nothing tests whether a planner would act on the list, which is the next question this result raises.

The battery contains no graded live-excess positive: a healthy item drifting into genuine oversupply cannot be truth-keyed at the data window's edge without reading the future, so that class was cut at design time. The claim-scope consequence is stated once in the caveat block above.

Stranded stock publishes in units because the dataset carries no cost or margin data. The visibility census has no external benchmark to compare against; no public source measures what share of manufacturers see distributor stock at all, which is itself a fact about this pain.

Refusal integrity is 4 of 5, at the registered bar exactly. The fifth position was answered defensibly under the prompt's convention rather than refused. The contradiction that produced it is logged as a registration defect to fix in the next roster design.

Assessments are as of each position's last definitive reading. Where reporting stops early, the true state after the gap is undefined even in the answer key. The run's staleness caveats on those positions are part of why its calls graded correct.

Claude only this cycle: claude-fable-5, stated at the start and the end of the run and verified against the session's own transcript, which shows one model string across all 154 turns and zero subagent calls. No claim generalises past that string.

Replication candidates named in the pre-registration, through the same harness-portable pack: Grok 4.5 and Kimi K3, bounded to two workers. The registered trigger fires: a pass replicates to check it is not model-specific. Whether to spend that leg is an owner decision recorded in the tracker rather than a default.

Two independent graders, each blind to the other, recomputed every truth value from scratch in standalone Python against the held-out key. Both reproduced every registered ruler value before grading anything. They agreed outright on 19 of 21 units and independently recommended the same ruling on both disputes; reconciliation changed no verdict-level outcome. One grader additionally reconciled the discontinued-stock definitional gap to the exact two positions and the exact 22,185 units that produce it.

The integrity sweep ran all seven standing items: timestamps monotonic and arithmetically exact to the millisecond, every capture and tool inside the declared run window, the frozen catalogue hash intact, the attestation present with no prohibited path appearing in any tool input, model attribution verified against the transcript rather than the self-report, every run-produced summary figure recomputed, and per-layer token counts complete. The cost on this page is the transcript-complete figure, which supersedes the run's own capture-time floor and says so.