# Method corrections

## Before final grading — November15,2026 Pacific
1. Task1 private answer key said all net/per-order figures depended on missing RF05.
   Corrected to Course and combined only; Workshop is unaffected by that refund.
   RF04 remains a separate unallocated orphan. Candidate input files were unchanged.
2. The diagnostic CSV checker initially treated an unrecognized inclusion-column
   name as failure. Opus uses `applied`; Fable outputs included-only records. Both
   are permissible. Added `applied`; unsupported layouts now require manual review.
   Scores are based on substantive records and amounts, not column-name conformity.
3. Added tasks7–10 at Curtis's request after initial task1 runs began. Tasks1–6 inputs
   were preserved; all five models receive identical added tasks.
4. Budget increased from US$50 to US$100. The active original batch retained its
   US$50 cap; a second batch of added tasks has an independent US$50 cap. Separate
   ledgers are aggregated in the evidence index, without double-counting reserves.

No model was shown the private answer key or these corrections. Keep this log with
the public method and append substantive later changes; do not rewrite history.

5. Website checker initially rejected every digit in an invalid-input message,
   incorrectly treating the valid instruction ‘1 to 8 seats’ as a numeric price.
   Changed to detect currency/decimal-price patterns, kept the original result, and
   reran the affected output. This corrects evaluation only, not the candidate.

6. Two connection resets interrupted T10calls in the added batch. The full uncertain
   call reserves remain charged against capacity; no zero cost is assumed. The batch
   paused dispatch. Untouched jobs resume, and SolT10gets one separate fresh-context
   infrastructure retry. Original partial attempts remain in the evidence archive.
   Astra/Qwen partial revision outputs remain explicitly incomplete.

7. The original batch also reset on QwenT04revision; its reserve is retained.
   Untouched work and two interrupted initial stages restart once with separate
   retry request IDs. Resumed tasks use an800,000-byte request cap because the
   shared runner was expanded for images. Earlier220,000-cap runs retain their
   original outcomes. This is a protocol limitation and must be disclosed.

8. Curtis approved a cumulative US$125 inference ceiling and a separate US$59.08
   payment for US$50 credits. Top-up was observed after checkout verification.
   Completion uses a single ledger with US$67.5263644 historical usage accounted
   conservatively, including the US$0.16833 unassigned reconciliation difference.
   Historical reserve estimates remain in the old ledgers; they are not added again
   to this reconciled baseline as if they were additional confirmed charges.
9. The completion pass uses streamed responses, records generation IDs/events as
   they arrive, and limits concurrency to two. Maximum request size is800,000bytes.
   Pinned provider/model/reasoning and the fixed revision prompt are unchanged.
   Timing is not a controlled speed benchmark across execution waves. Interrupted
   initial stages restart from pristine task inputs. Interrupted revisions restore
   the archived initial output and original initial conversation, then replay the
   same revision prompt. The copied workspace can retain incidental scratch files;
   this is a recovery limitation, disclosed rather than called a pristine repeat.
10. Genuine turn-limit and output-truncation outcomes are retained and graded as
    delivered; there is no quality retry. The completion runner's live version
    treats any non-length/no-tool response as complete; artifact review and final
    finish-reason audit are required to detect empty or filtered responses. A
    stricter finish-reason check was prepared after launch, but does not retroactively
    change the executing process.

11. Final integrity review separated selected-attempt stage costs from older retry
    charges. All receipts still count toward total spend. Missing receipts are shown
    as unknown with a known subtotal, never silently as zero. Grades must match the
    selected artifact hashes before display. The original grade audit filled three
    omitted reviews and corrected one inconsistent evidence label without changing
    its total. Public stage traces are bounded at the initial/revision boundary; a
    revision trace includes its prior conversation for context.

12. Qwen's second initial website desktop screenshot showed a stale viewport-paint
    strip. Repeated capture after a500ms resize wait displayed the intact single
    page; all19functional checks passed both times. Both captures are retained.
    The recapture changes operator observation only, not candidate files or score.

13. The same viewport capture artifact occurred on Sol's second initial website.
    Repeated observations after the same500ms wait passed all19checks and produced
    intact layouts. The operator verifier now waits after resize for later captures.
    This is an evidence-capture fix; no model output or scoring criterion changed.
