# Run 20260923T062806Z: the low figures read before any verdict (batch item 1 of the release-page brief) **What this is.** The brief asked, before any verdict, for the transcripts behind three low figures to be read and each classified as a real answer judged low or as an artefact (truncation, refusal, tool error, timeout, judge parse failure), with Report 007's timeout defect as the precedent. An artefact is an instrument defect: it is recorded here, beside the run record, and its cell is excluded from every verdict. **What was read.** Every guarded call of the three cells, from the ignored call records under `driftproof-calls/` (their sha256 are in `driftproof-calls.SHA256SUMS` beside this file), each call's `stdout.jsonl` result event, `stderr.txt` and `metadata.json`, and the matching draws in each receipt (`results.cases[*].generation.draws[*]`). Then, as a check on the cells the brief did not name, the same fields for all 204 calls of the run. ## Result: no artefact. No cell is excluded. | cell | receipt | calls | draws | what the draws are | |---|---|---|---|---| | claude-opus-5-5, git-workflow-and-versioning, baseline, 0.300 | `ca69c343…` | 041-052 (generations 041, 045, 049) | 3 of 3 measured, each `stop_reason: end_turn`, `truncated: false`, 9 of 9 judge samples 0.3 | Three complete commit messages. Each subject line begins `Fix password-reset links expiring ...` with no conventional type prefix, and the rubric reads "CAP: if the subject line does not use a conventional type prefix, cap at 0.3". Every judge reason names that cap. A real answer judged low. | | claude-opus-5-5, code-review-and-quality, baseline, 0.680 | `5c3ad19b…` | 013-028 (generations 013, 017, 021, 025) | 4 of 4 measured, each `end_turn`, not truncated; draw means 0.860, 0.483, 0.500, 0.877 | Four complete reviews of 3,678 to 4,195 characters, each ending with its closing paragraph. The two low draws (calls 017, 021) put the hardcoded key and the PII logging under `Critical` and the remaining findings under category headings (`Correctness`, `Design`/`Maintainability`) with no severity label; the six judge reasons for them say so. A real answer judged low; the spread is between draws, not an instrument fault. | | claude-opus-5-5, documentation-and-adrs, with_skill, 0.661 | `b924a435…` | 053-076 (generations 053, 057, 061, 065, 069, 073) | 6 of 6 measured, each `end_turn`, not truncated; draw means 0.483, 0.800, 0.600, 0.853, 0.697, 0.533 | Six complete revised functions. The low draws keep the `TODO` about expired coupons reworded rather than resolved or removed (rubric criterion c) and add a gold-tier comment the judge reads as restating the code (criterion d). A real answer judged low. | **Across all 204 calls:** every call exited 0 with a result event, `is_error: false`, `stop_reason: end_turn`, `terminal_reason: completed`, one turn, no permission denial and an empty `stderr.txt`. Every judge result parses as the JSON object the template asks for. No call was retried and no draw is unmeasured in any receipt. ## One observation outside the three cells, recorded, not excluded **claude-opus-5, git-workflow-and-versioning, baseline, draw 3 of 10 (call 133), judge samples 0, 0, 0.** The surface runs with no tools (`--tools ''`, and the init event lists none). The generation is 227 characters: one sentence saying it will check whether the file exists, then a `Glob` tool invocation written out as plain text, then the words "No files found." It is not a tool call the harness made (one turn, no tool-use block, no permission denial): the model wrote the invocation and its result as text and stopped. It contains no commit message, and the three judge calls (134-136) score that absence at 0.0, each saying no commit message was produced. Read against the brief's list, it is none of the five: not truncated (`end_turn`), not a refusal, not a tool error (no tool ran), not a timeout (27,527 ms against a 240 s limit), and the judge parsed. It is the model's own output, delivered whole, and judged on what it is. So it is **not excluded**. It is recorded because a reader may weigh it differently, and because it moves a figure: without that draw the arm's mean would be the mean of the other nine. It sits in a **baseline** arm, which neither comparison's verdict reads (`driftproof diff` compares the `with_skill` arms only); it enters the page only as context. **Recorded by** the building session on 2026-09-23, before any verdict was written.