Abstract screening: output is 88% of the stage's cost, and it is all written justification

Abstract screening spends most of its money on prose the pipeline rarely reads back.

Measured

Abstract screening, over 708 calls (vbm_of_ptsd and vbm_of_substance_use, with the token accounting from autonima/llm/usage.py):

Output is 88% of the stage's cost, and abstract screening is ~26% of a project's total. The stage emits more output than input — the driver is the written reason field, not the abstract.

Capping reasons at roughly 200 tokens would cut abstract-stage cost by ~72%, or about 19% of a whole project.

But do not just truncate them

Those reasons are a genuine asset. Classifying 6,835 full-text exclusion reasons is what established that 82% cite a fact a structured schema field could carry — the analysis behind the evidence-layer work in the validation write-up. Throwing that away to save $4 a project would be a bad trade.

Suggested shape

Make verbosity conditional rather than uniform:

• full reasoning on exclusions and on low-confidence decisions, where the justification is doing real work (audit trail, PRISMA, error analysis);
• one line on confident inclusions.

Most abstract decisions are confident excludes... which is exactly the bucket worth keeping verbose. So the saving is smaller than 19% in practice, and the right split needs checking against the actual confidence distribution before committing to it. An alternative worth considering is keeping the reason but dropping the restatement of criteria IDs already captured in inclusion_criteria_applied / exclusion_criteria_applied, which is duplicated structured data rendered as prose.

Scope

Same pattern likely applies to full-text screening (1,211 output tokens/call) and annotation (4,424), but there output is a much smaller share of cost because input dominates, so abstract is where it matters.

Related

#67 and #68 address the input side of the same bill. Together the three are on the order of 35-40% of project cost.

Start with autonima/llm/usage.py to understand the existing token accounting, then trace how abstract-screening reasons and confidence decisions are produced. Compare the confidence distribution and duplicated criteria fields before choosing conditional verbosity. Done means preserving useful exclusion and low-confidence reasoning while reducing unnecessary output and measuring the cost change.