# OpenRouter reconciliation and completion plan



Saved response receipts: **US$67.35803440** across 1001 dispatched call records.

API key usage increase since recorded preflight: **US$67.52636440**.

Unassigned difference: **US$0.16833000**. This is not individually matched to failed requests and must not be called zero.

The former **US$26.70488375** reserves were request maximums, not confirmed charges. Retain the original ledger; aggregate reconciliation is a separate evidence layer.

Account-wide balance omitted from public copy.



## What remains

43 affected model/task/repetition pairs; 72 unfinished stages. Details: [missing-runs.csv](missing-runs.csv).

The one legitimate initial turn-limit result remains reported as such. Do not replace it with a selected better result. Infrastructure retries are separately named and linked. Existing completed stages remain intact.



Replaying missing work at observed first-pass cost is estimated at **US$40.44278170**, not a fixed quote. This puts cumulative usage near **US$107.96914610**.

Recommend cumulative inference cap **US$125** and **US$50 credit top-up**. The actual checkout preview showed **US$2.75 fee + US$6.33 tax = US$59.08 payment**. No purchase submitted. Credit purchase and inference consumption must not be double-counted; fee/tax remain separate cash costs.



## Execution changes before resuming

- Preserve initial outputs, completed revisions, all unsuccessful attempts and every receipt.

- Use a durable shared cap across both task groups, not two artificial allocations.

- Persist provider generation/request IDs and error bodies immediately, before parsing the full response.

- Retry only infrastructure failures, with unique attempt IDs and bounded retries. Never retry a content failure to improve its score.

- Stop on 402. Check account balance and maximum output reserve before any next request.

- Keep exact model/provider and reasoning configuration; document dates and runner fixes.

- Independently grade every completed stage, rerun relevant website checks, populate the workbook, and rebuild the portable guide/evidence bundle.

- Verify every guide link. No placeholder URLs or unreported missing cells.



## Evidence limits

Aggregate API-key reconciliation, not exact per-call attribution. The shared key may include other work. Unassigned usage is conservatively retained; unknown calls have no saved generation IDs. Do not set their actual cost to zero.

The browser activity export did not provide a usable requested-period file. One observed generation was independently matched to its local receipt with the documented generation metadata API. It matched exactly. Full per-call reconciliation of the24uncertain requests remains unavailable without provider-side IDs/log export.

No new paid model calls or credit purchases were made during this reconciliation.
