Keep usable English.
Record defects without filtering countries.

What the numbers mean

2,531,940 passing inputs carry warnings. This is a subset of the passing count, not another exclusion category. One input-held record also has a retained source-scope issue; the exclusive table assigns it to source scope, so its input-exclusion row is one lower than the total input-screen holds. 21 earlier document-level holds were retained with warnings after checking for substantive English prose obscured by repeated or garbled tails.

Claim snapshot: 2026-10-17T04:58:47.174031+00:00. The candidate count is not a live queue size, a guarantee that all records fit one request, or a semantic acceptance total. Every actual submission rechecks paid/unknown claims, source identity, physical request limits and remaining budget.

Warnings versus exclusions

Original source language and country are not exclusion rules. Exact stored translated title and body are used; blank translated titles may remain blank when the body is usable. Text is not silently repaired, completed or replaced by the original-language fields.

Language-model scores are screening signals, not calibrated certainty or a count of proven bad translations. This screen does not establish translation fidelity, complete sentences or scientific label accuracy.

Input exclusion reasons

Why the old “remaining candidates” count was misleading

The earlier 1,624,549 was an intersection: records had to clear older language-window, sentence and quality-hold checks and previous-attempt exclusions. It was not the number of all usable English articles, and it was not a representative research sampling frame.

The earlier strict screen could flag a name, byline or short span even when the article body was readable English. That explains why an automatic flag was too broad a reason to remove an article. The old report remains available as historical evidence; current candidate decisions use the updated policy above.

Scope, processing order and traceability

The source scope is MLP civic-related, non-disaster, non-999 records from enabled sources. MLEED ENV source families are excluded. Shared MLP/MLEED domains remain when they are valid MLP source identities. The canonical snapshot covers January 2012–October 2026; September is the latest nonempty month.

Paid processing starts with June 2026 and moves backward within the inclusive shared $500 budget. Previous spending, pending reservations and unknown outcomes count toward that limit. Requests use the stable prompt, cached shared prefix and bounded output cap; no automatic paid retry is allowed.

Each result retains its source/version, request, prompt, model, policy, structural outcome and review flags. Operational results are labeled as structurally valid with pending or sampled review, rather than fully reviewed scientific gold. DB insertion and exact readback are reported only from actual write receipts.

First 40 production attempts: returned and stored

All 40 responses are accounted for. The DB contains all 40 immutable operational attempts, each checked by exact readback, with zero save exceptions and no source documents modified. The one structural failure is preserved as a failed attempt with its raw response; it is not silently repaired or retried.

The 39 structurally valid records include the four clear conflicts and nine adjudication cases. The original outputs and flags remain visible for later parsing or scientific review. Storage does not convert flagged or unreviewed labels into accepted gold data, and this 40-record shard is not a completed month.

Cache reuse: 39/40 requests. The batch used 262,782 cache-read tokens, 6,738 cache-write tokens and 17,628 ordinary input tokens. Actual cost for all 40 outcomes, including the failed extraction, was $0.014481185 (ledger-rounded $0.014482). The $0.117373 reservation was a conservative bound, not the charge.

The 47,459 output tokens already include reasoning usage; reasoning is not charged a second time. Observed cache reuse is not a guaranteed future hit rate or a full-month cost forecast.

These are the final amended review counts after the bounded independent reread. Four review dispositions were updated in separate, exactly read-back DB amendments. This includes correcting an initially overstrong conflict judgment and recording additional action/response and interpretation flags. Original model outputs, attempts and source text are unchanged.

A further 40 June requests have been submitted by the rolling worker and were awaiting provider completion at 05:17:25 UTC on 17 October. They are not included in the completed or DB totals above.

Cache reuse is already demonstrated

The completed 15-request diagnostic had 13 cache-read requests, 87,594 cache-read tokens and 13,476 cache-write tokens. Actual cost was $0.004370770. All 15 outputs were structurally valid; 11 reviewed attempts were saved with exact DB readback.

Open the completed cache test. This is evidence of actual reuse, not a guaranteed monthly hit rate or controlled accuracy comparison. The shared $500 budget is not the same as diagnostic-only spending.

All 40 responses from the first June production batch returned: 39 structurally valid and one evidence-length failure. All 40 original operational attempts are stored with exact DB readback, including failures and review flags. These are not 40 accepted scientific labels. This batch had 39/40 cache-read requests and cost $0.014481185 in total. A further 40 June requests have been submitted by the rolling worker and were awaiting provider completion at 05:17:25 UTC on 17 October. They are not included in the completed or DB totals above. Structural validity is not a full scientific accuracy certification; DB totals are reported separately.

Month-by-month input screen