Instruments · A Trial Balance for Agent Omissions
Trial Balance budget curve
Set a reading budget and a diary length and read how many skipped steps each of four audits finds, then switch fault classes on and off to see what a full read of the diary finds and misses.
What it computes
The share of injected omissions each audit named correctly, at each of seven reading budgets and seven diary lengths, for one auditor model. A second chart gives detection at a full read by fault class for three of the audits.
The formula in plain words
Strict detection is the share of injected faults where the audit flagged a fault, named the right class, the right requisition and the right step. Omissions pool the silent skip and the skip the diary claims was done. Each point is 24 audits on the two primary synthetic seeds; each full-read bar is 84. The budget is the nominal reading allowance; the measured prompt tokens at the two ends of the curve are printed where the pieces give them.
What this does not say
It does not price the commitment ledgers themselves: imbalance-first's score is mostly the auditor copying the imbalanced line into its verdict, and building the two ledgers is outside the measurement. Every number is one auditor, qwen3.5, under schema-constrained decoding with thinking off, on one procurement workflow with one fault per diary. The balance check finds none of the wrong amounts carried on both sides of the book, by construction.
Source
Every number comes from A Trial Balance for Agent Omissions (Paper, Medium, 2026-10-01), with some from An AI agent can skip a step without leaving a trace of the omission. (Executive piece, 2026-10-01). The campaign record is A Trial Balance for Agent Omissions.
Every number this instrument uses
| Number | As published | Where |
|---|---|---|
| Nominal reading budgets | 200, 400, 800, 1600, 3200, 6400, 12800 | Paper, Section 3, Design, and the columns of Figure 3 (six doublings from 200 to 12,800) |
| Diary lengths, in requisitions | 1, 2, 3, 5, 8, 12, 16 | Paper, Section 3, Design |
| Omission detection per arm; one row of seven budgets per diary length, rows in the order of the lengths | imbalance: 0.96, 0.96, 1.00, 1.00, 1.00, 1.00, 1.00 / 1.00, 1.00, 1.00, 1.00, 1.00, 1.00, 1.00 / 1.00, 1.00, 1.00, 1.00, 1.00, 1.00, 1.00 / 1.00, 0.96, 1.00, 1.00, 1.00, 1.00, 1.00 / 1.00, 1.00, 1.00, 1.00, 0.96, 1.00, 1.00 / 1.00, 1.00, 1.00, 1.00, 1.00, 1.00, 1.00 / 0.96, 0.92, 1.00, 1.00, 0.96, 1.00, 1.00; logreview: 0.04, 0.17, 0.29, 0.29, 0.33, 0.25, 0.29 / 0.00, 0.21, 0.21, 0.54, 0.46, 0.50, 0.58 / 0.00, 0.08, 0.08, 0.21, 0.29, 0.38, 0.33 / 0.00, 0.00, 0.00, 0.21, 0.21, 0.42, 0.46 / 0.00, 0.00, 0.08, 0.08, 0.08, 0.17, 0.21 / 0.00, 0.00, 0.04, 0.08, 0.21, 0.25, 0.29 / 0.00, 0.00, 0.00, 0.08, 0.04, 0.00, 0.12; sample: 0.17, 0.54, 0.25, 0.25, 0.25, 0.25, 0.25 / 0.04, 0.12, 0.46, 0.50, 0.39, 0.42, 0.50 / 0.00, 0.04, 0.12, 0.21, 0.33, 0.29, 0.29 / 0.04, 0.04, 0.00, 0.17, 0.29, 0.33, 0.29 / 0.00, 0.00, 0.00, 0.00, 0.00, 0.17, 0.21 / 0.00, 0.00, 0.00, 0.08, 0.04, 0.08, 0.25 / 0.00, 0.00, 0.00, 0.00, 0.00, 0.08, 0.21; checklist: , 0.25, , 0.50, , , 0.58 / , 0.08, , 0.62, , , 0.58 / , 0.00, , 0.33, , , 0.50 / , 0.00, , 0.33, , , 0.71 / , 0.04, , 0.25, , , 0.58 / , 0.00, , 0.08, , , 0.46 / , 0.00, , 0.04, , , 0.38 | Paper, Figure 3, strict omission detection by trace length and budget, synthetic seeds 1 and 2, n = 24 per cell |
| Detection threshold that defines C80 | 0.80 | Paper, Section 4, H1 |
| Measured prompt tokens for imbalance-first at the smallest budget, per diary length | 563, 586, 611, 666, 745, 850, 954 | Paper, Figure 2, imbalance-first C80 in measured prompt tokens per diary length (at budget 200) |
| Measured prompt tokens for log review at a full read, per diary length | 994, 1683, 2421, 3882, 6173, 9044, 12184 | Executive piece, Figure 2, reading cost per audited episode |
| Correct omission findings per thousand prompt tokens at 8 and 16 requisitions: imbalance-first, log review at its best budget, log review at a full read | k8: 1.34, 0.063, 0.034; k16: 1.00, 0.035, 0.010 | Paper, Section 4, H1, the finite descriptive |
| Log review reading the whole diary at about 6,200 tokens, pooled over 1 and 8 requisitions | 0.25 | Executive piece, Headline figures |
| Detection at a full read by fault class | classes: Silent omission / Phantom completion / Value commission / Consistent commission; imbalance: 1.00, 1.00, 1.00, 0.00; checklist: 0.726, 0.357, 0.976, 0.952; logreview: 0.476, 0.179, 0.976, 1.000 | Paper, Figure 4, strict detection at full read by fault class, synthetic seeds 1 and 2, n = 84 per class |