Civic June 2026 · Diagnostic results

Two completed runs · initial review history

Initial assistant dispositions and actual observed costs. Database verification establishes storage and source provenance, not label accuracy.
Run	Articles	Initially accepted and DB verified	Initial review flags	Length failures	Actual API cost (USD)
First run	33	21	10	2	$0.0231873875
Next run	40	23	16	1	$0.0292755625
Total	73	44	26	3	$0.05246295

All 73 outcomes were inspected. Initially, 44 were stored, 26 were flagged and 3 were unscorable. Reassessment withdrew 3 flags; those unchanged outputs were saved after source verification, bringing current DB totals to 23 from the first run and 24 from the next run (47 overall). The remaining 23 held outcomes and 3 unscorable outcomes are excluded from accepted DB results. No source documents were modified and no paid retries were made.

What the review flags mean

The initial 26 flags were reviewed against the exact inputs and prompt: 15 retain clear prompt-rule conflicts, 8 require adjudication, and 3 were withdrawn. The 8 adjudication cases are not asserted to be proven errors. These are assistant judgments without a human gold reference; the categories below explain the kinds of questions reviewed.

Entity and target attribution

Whether the selected actor and target are the entities carrying out or receiving the action, rather than entities merely mentioned.

Government scope

Whether local or provincial officials are being confused with the national incumbent government.

Routine contact versus political support

Whether diplomatic interaction alone supports a political alignment label.

Action spans and conditions

Whether evidence captures the full action and distinguishes conditional requirements from events that happened.

Whose viewpoint is represented?

Whether a stance belongs to the source, the narrator, or an attributed speaker.

Final-response completeness

Three requests hit the completion-token limit without a usable final response. They remain unscored and their costs are included.

Inputs and source checks

Requests used title_translated and maintext_translated exactly.

Source and context checks were limited to reviewed samples; they do not certify the full archive.

All 40 inputs in the second run were English originals with identical stored translated fields. That run adds no new multilingual translation-fidelity evidence.

Inputs, outputs and previous attempts were preserved under their original versions.

How to interpret these results

This is assistant review, without a human gold reference or scientific acceptance.

Acceptance and reassessment counts are not an accuracy estimate. Mechanical validity alone does not establish substantive correctness.

These small selected samples do not establish whole-month performance or a precise cost forecast.

The baseline used gpt-6-luna, medium reasoning and a 2,048-token completion cap including reasoning.

Next step

Use the completed reassessment to evaluate a transparently versioned prompt or profile on a bounded diagnostic before expanding archive processing. The next prepared chunks have not been submitted.

All 47 stored attempts retain their original model outputs. The three acceptances after reassessment have separate review provenance; the initial reviews and their original 44 stored attempts remain preserved.