Where the Work Piles Up · Working Paper No. 28

1. The constraint

Eliyahu Goldratt was a physicist who fixed factories. His claim is that a production system has one binding constraint, and the throughput of the whole system is the throughput of that step. Asked to compress the theory, Goldratt offered one word, focus. Jeff Bezos made his executives read the book.

Two exhibits for the same rule

Exhibit	What it shows
Goldratt, The Goal (1984)	One binding constraint sets output. Effort elsewhere is inventory, not throughput.
Sheila Taormina, swimmer	Missed the 1992 United States Olympic swimming trials with an elite aerobic engine that took most of her coaching effort. Her constraint was power, at five foot two. Four years of neglected strength work bought three seconds at the 1996 trials and a relay gold.

2. The audit

This pipeline records a cause of death for every hypothesis it files.

The filing base, as of 11 August 2026. The 308 sweep rows are machine-generated candidates adjudicated in bulk against a vendor simulator’s rulebook, where one row is one expression rather than one idea.

Rows	Count
All rows in the database	393
Combinatorial expression sweep, held out	308
Ideas from a named source, through this engine’s own gates	85

Cause of death, every hypothesis filed, as of 11 August 2026

Stage	Hypotheses
Filed	85
Never ran: blocked before any statistic	6
Died at the gates	72
Still queued	5
Survived	2

3. What the constraint is not

Goldratt’s field test is that a bottleneck has work piled up in front of it while a non-bottleneck sits idle. The gates here are idle in industrial quantities.

Idle capacity at the gates, and the lane they have already exhausted. Several variants of one idea mean several chances to get lucky, so the bar raises itself.

Measure	Value
Pairs tests filed and closed, 11 August, one afternoon	5
Round trips in the pair-level simulation	8,051
Overlapping books it covered	247
Required t-statistic, first test to fifth	2.50 to 3.03
Nightly runtime, deterministic statistics code	Minutes, then idle
Deaths at the standard validation stage	35
Deaths at study-specific bars, same arithmetic	37
Price-derived candidates now dead	129

Sharpening the gates further is repaving the road behind the bridge. They are the aerobic engine of this operation, already elite, already over-coached, and not the reason so little survives.

4. Where the work piles up

Four of the deaths that never touched a statistic read as one queue in front of one door.

Blocked before any statistic

Idea	What the record says
Biomedical event study	7,682,445 prediction market rows checked, 43 relevant, 23 resolved. Died of data coverage.
Tail risk	Needs historical index options chains, a paid purchase. The spec itself declares the idea blocked.
Fundamentals source	Needs an API (application programming interface) key whose free signup is a human-only action. No key, no run.
Event drift	Needs 8-K filing bodies. The filings cache holds 14,158 documents, all 10-K and 10-Q. Zero 8-Ks.

Price is the one input this factory has in unlimited supply, and that lane is strip mined to zero; both survivors were plumbing rather than price. The lane the gates keep pointing toward, positioning, flow and forced trades, has no data on the shelf. The slowest step is the data source, not the statistical test.

5. Five focusing steps, applied

Goldratt’s procedure is identify, exploit, subordinate, elevate, repeat.

The five steps, applied to this pipeline. Orthogonal here means statistically unrelated to the price data already mined.

Step	What it means here
Identify	Orthogonal event and positioning data supply is the constraint.
Exploit	Done. The free lanes are strip mined to zero survivors and the database carries the receipts.
Subordinate	The statistical gates stay exactly as they are. No further sharpening.
Elevate	The 8-K source first: two recorded deaths point at it, and it opens the one class of ideas this database has never been able to test, risk arbitrage. Then delisted prices, which make every convergence number honest. Then options and positioning data as budget allows.
Repeat	Once a source opens, the constraint moves, probably to idea quality inside the new lane. The audit gets rerun.

6. What this does not establish

This is a small factory and these ratios will move. Only filed ideas are visible, so ideas never filed because the data was known to be missing do not appear and the queue in front of the data step is understated. Idle capacity and the narrow data lane are coupled, since cheap data made cheap tests abundant, and elevating supply will load the gates. An audit picks a direction, not a result. The 8-K source may open and every idea in it die. The gates will say so, and that is the product.

Provenance, bs-prov/1.0

Method: SQL (Structured Query Language) queries over the database of every hypothesis tested here, all rows as of 11 August 2026, deaths grouped by recorded stage; error texts quoted from the database verbatim. No statistic in this paper is anything but a row count.
Counts: 85 rows: 2 passed, 72 rejected, 6 error, 5 pending.
Framework: Goldratt, The Goal (1984); five focusing steps. The Taormina account follows the source video, watched 11 August 2026.
Related: Working Paper No. 13 (the factor zoo bar), No. 22 (who has to trade), No. 27 (a fitted value is not a measurement).