Does Parallel Review Still Pay Once Compute Counts?
The same detection and false-alarm rates can favor different review designs. In our internal numerical experiment, using the source-labeled Haiku 5.5 rows, an assumed 2% defect share and a miss costing 20 false alarms favor three full reviews with any positive draw triggering a flag when compute is free. Adding a particular resource charge makes one review the cheapest option instead.[2]
The model has not changed its capabilities. Its detection rates stay fixed; what changes is the assumed cost of obtaining them. Our proposition is that review choices must combine misses, false alarms and resources in a comparison with explicit units. Choices calculated from rounded aggregate rates support conditional decisions, not universal claims about collaboration architectures.
What the division of work actually changes
Pei Liu's preprint tests a narrow parallel decomposition. One call checks six requirements; three partitioned calls check two each, while all read the complete memo. There are no tools, member communication, orchestrator or later filter. Additional options restore the unassigned requirements to partitioned calls, or repeat the full review three times and aggregate by majority or any flag.[1, Section 3.3]
The source describes 300 program-generated credit memos of roughly 1,100 words, with six checks per memo: 1,800 cells, including 1,136 clean cells. These test internally checkable inconsistencies, not external facts, omissions or actual credit decisions. The author reports that decomposition raises detection for some models while also increasing false alarms. Equal call counts do not mean equal token use. Some repeated-call comparisons also change sampling temperature, preventing attribution of every difference to responsibility allocation.[1, Sections 3–4]
We therefore compare conditional costs for five disclosed options, not the space of multi-agent architectures. The cited performance results remain source-reported claims rather than independently reproduced measurements.
Put errors and computation in the same units
The completed same-day internal experiment uses the 20 aggregate rows in the paper's Table 2: four source-labeled models and five options per model. Let p be the share of defective cells, H the detection rate, F the false-alarm rate on clean cells, and r the cost of one miss relative to one false alarm. K is token use relative to one full review; tokens measure model text processing. Let c be the assumed cost of consuming one single-review token budget, expressed in units of one false alarm.[2]
Conditional expected cost per six-check memo is:
6 × [p × (1 − H) × r + (1 − p) × F] + c × KThe first term covers expected errors across six cells; the second is the resource charge for the whole memo. c is not a per-token price, and the resource term must not be multiplied by six again. This uses no measured service pricing and estimates no monetary return. It assumes additive linear error costs and transfers the source's detection and false-alarm rates to hypothetical defect shares, a transfer not validated in deployment.
The grid uses defect shares 0.001, 0.01, 0.02, 0.05 and 0.37; integer miss-cost ratios 1–200; and c values 0, 0.01, 0.1, 0.5, 1, 2, 5 and 10. Across four models, that produces 32,000 scenarios. “Cheapest” means cheapest among that model's five options at that grid point, not a global optimum over continuous parameters.[2]
A changed choice, not a measured saving
Only the source inputs needed for the opening example appear below. They are reported, rounded rates, not newly measured performance.[1, Table 2; 2]
| Option | Detection H | False alarms F | Relative tokens K |
|---|---|---|---|
| One full review | 0.825 | 0.013 | 1.00 |
| Three full reviews, flag on any draw | 0.929 | 0.029 | 3.00 |
At p=0.02, r=20 and c=0, the internal calculation selects any-of-three as cheapest, at 0.34092 false-alarm cost units per memo. With c=0.1, one review becomes cheapest among the five options, at 0.59644. Those numbers use different assumed charges. Subtracting them would not measure a saving, and actual expenditure was not observed.[2]
Both winner comparisons survive adverse changes of up to 0.0005 in H and F while K stays fixed. That means the changed choice in this example does not depend on the final displayed rate digit. It is not a confidence interval and covers neither token rounding, sampling error nor variability between runs.
A tiny winning margin cannot carry a broad claim
Within the p=0.02 subgrid, the original partitioned Team option wins none of the 5,600 positive-charge comparisons. Among 800 zero-charge comparisons, however, there is one Team winner: the source-labeled Qwen3.5 9B row at r=39 has cost 2.74308, beating three-draw majority by 0.0036.[2]
That isolated winner fails the H/F rounding sensitivity check. Exact arithmetic on the displayed rates therefore does not reproduce a strict “no Team winner even without resource charges” control. It also does not refute the source's unrounded measurements. The experiment retains the discrepancy. There are a few Team winners at other defect shares in the full grid, so the p=0.02 result cannot be expanded into a claim that partitioning is never worthwhile.[2]
Two inferences must be withheld together: a narrow, rounding-sensitive win is not reliable superiority, and failing to win on one grid does not exclude winning under other tasks, prices or parameters.
Evidence and sources
This article separates three layers. Source-reported claims describe the author's design and performance. Internal experimental evidence comprises our calculations on those aggregates. Research interpretation concerns the resulting decision method. A separate decimal-arithmetic program checked the calculation, but it remains internal verification within the same research operation, not third-party independent evidence.[2]
[1] Pei Liu, The More the Better? A Comparison of Single- and Multi-Agent LLM Setups on a Financial Review Task, arXiv research preprint v1, October 20, 2026. Design: Section 3; inputs: Table 2; results and limits: Sections 4–5 and appendices. Read the fixed version. Public version status establishes traceable provenance, not theoretical validity or independent reproduction. Citation does not constitute endorsement.
[2] Research Center, same-day internal experiment EXP004, November 2, 2026: 20 input rows, the complete 32,000-scenario grid, results, execution record and decimal verification. The attached inspectable evidence package retains those same-run artifacts. No model was called again and no original document benchmark was rerun. Article production organizes completed research only.
The cover is an editorial concept generated from the cost-choice relationship, not a data figure or an actual system capture. The formula and compact table already communicate the precise relationships; no decorative inline figure is added.
Measurements still needed before deployment
The verified source package provides no downloadable complete generator, memos, labels and predictions, preventing a faithful rerun of its 300-memo evaluation. Reported prompt and instruction dependence, one requirement partition and the lack of repeated whole Team runs restrict transfer. Cell-level false alarms may also include a flag assigned to the wrong requirement. Whole-file human adjudication could change handling costs; this experiment leaves the original scoring unchanged.[1–2]
For engineers considering parallel review, the useful output is a bounded decision method: declare prevalence, error costs and resource units before comparing partitioning with repeated full reviews under matching conditions. What is the real defect base rate? Does handling each false alarm have a fixed cost? Do the choice boundaries survive matched temperatures, different partitions and repeated runs? Until field evidence answers those questions, this experiment supplies neither product returns, financial deployment advice nor a universal architecture ranking.