Score sheet · list-price scenario, not an invoice

One window vs a four-call graph

Twelve frozen trial files. Same model. No live search. Not a parent with subagents: one window versus three specialists, a software merge that keeps disagreements, and a checker who cannot open the files. Quality cannot be decided on one extra correct case in twelve. Extra calls did not earn their keep on packets this size.

12Trials

10Both matched the key

1They disagreed

1One agent unfinished

4 vs 1Calls per trial

The job

Classify what these files establish: results paper, protocol or design paper, meeting abstract, pooled or secondary analysis, registry posting only, or nothing established.

The wiring

A: one agent sees all three dumps and self-checks. B: one specialist per dump, software keeps disagreements, a checker who cannot open the files. Not a parent with subagents.

Still draw the graph when

The files do not fit in one window, the jobs are independent, or the checker must not see the raw files. A nicer diagram is not that name.

All 12 trials

Orange rows need a closer look. Those IDs jump down the page. Other IDs are current ClinicalTrials.gov pages, not the scored dumps.

Full results: results paper on this trial Protocol / design: protocol or design paper Meeting abstract: conference write-up only Pooled / secondary: pooled, meta, or post-hoc, not the primary results paper under the locked key Registry only: numbers posted on the registry, no paper Nothing established: these files do not establish an eligible write-up

Trial	Locked key	One agent	Four-call graph	Match?
NCT01862536	Protocol / design	Protocol / design	Protocol / design	Both right
NCT02067780	Meeting abstract	Meeting abstract	Meeting abstract	Both right
NCT02205242	Protocol / design	Protocol / design	Protocol / design	Both right
NCT02513498	Registry only	Registry only	Registry only	Both right
NCT02635945	Nothing established	Nothing established	Nothing established	Both right
NCT02937584	Full results	Full results	Full results	Both right
NCT03123692	Pooled / secondary	Pooled / secondary	Pooled / secondary	Both right
NCT03414541	Nothing established	Nothing established	Nothing established	Both right
NCT03570359	Meeting abstract	Full results	Meeting abstract	Disagreed
NCT03573817	Full results	Full results	Full results	Both right
NCT03590379	Pooled / secondary	Did not finish	Full results	Unfinished / vs key
NCT03673670	Registry only	Registry only	Registry only	Both right

The two that need a closer look

Ten matched. These two did not, for different reasons.

The checker said no

NCT03570359: a title stretched into a full paper

Locked key: Meeting abstract
One agent: Full results
Four-call graph: Meeting abstract

The dump has a conference abstract. It does not confirm a journal results paper.
One agent treated PMID 38811970 as full results from the title. In the dump that PubMed row has no abstract, so NCT, PI, and enrollment could not be confirmed. The live PubMed page is out of scope.
The graph’s journal specialist made the same stretch. The checker, who could not see the files, refused it. That matched the key. A second pair of eyes helped. The three file-splitters did not add much.

Form break, then a miss versus the key

NCT03590379: TRI-D is in the packet

Locked key: Pooled / secondary
One agent: Did not finish
Four-call graph: Full results

TRI-D (PMID 33488071) is this RCT’s results paper. It is in the grey dump and on the registry citation list. PubMed’s dump has a class meta-analysis (PMID 40453981) instead.
One agent wrote the year as text (“2021”). A retry changed the trial ID. No usable answer. That is a form error.
The graph found TRI-D and called it full results. A publications officer looking at the same packet would likely agree. That is a miss versus the locked key, not versus the trial. Splitting still hid TRI-D from the journal specialist.

List-price scenario (not an invoice)

Run on Cursor Ultra (subscription, no token invoice). Zeros in the log mean not measured, not free. Visible tokens are estimated as UTF-8 bytes/4 on prompts and JSON, not a word count. “+ 8k thinking” assumes 8,000 hidden reasoning tokens per planned call (12 vs 48), billed as output at public API list rates checked 17 August 2026. Three one-agent schema retries are not in that band. No cache. No batch.

Model	One agent, visible	Graph, visible	One agent + 8k (planned 12)	Graph + 8k (planned 48)
Claude Opus 5	$0.40	$0.75	$2.80	$10.35
GPT-5.6 Sol	$0.41	$0.78	$3.29	$12.30
Gemini 3.1 Pro	$0.16	$0.31	$1.32	$4.92
Grok 4.6	$0.15	$0.28	$0.73	$2.58

Visible dollars: about 1.9× ($0.75 vs $0.40). Opus 5 at the 8k band: 3.7× ($2.80 vs $10.35). Four times is the call ratio, and the ceiling if thinking swamps the JSON, not a third measured bill.