這一章送出的訊息
About
What you see on screen is what a person would type. Nobody sits at a terminal for three hours, so a second Claude Code session sends each message and waits for the artefact that stage was supposed to produce, refusing to advance until the check passes. It types; it does not decide.
Stages 02–04 are verification passes over a corpus retrieved in an earlier run — search, screening and full-text retrieval are 73% of the wall clock and were carried over rather than repeated. Everything they assert is re-derived on camera.
What the gates caught
- The search was filtering when it should have been widening. Naming the anti-HER2 agents had been ANDed onto the query instead of OR-ed into it, so every record had to name a drug in its title or abstract. Measured live in stage 02: the mandatory blocks alone return 1,382 PubMed records, the correct full query 1,524, and the broken form 643 — a 53% loss, taking every chemotherapy-alone comparator arm with it. The guard is now an invariant the query is measured against, not a yield range someone has to remember.
- Dual review had once not been dual. Reviewer 2's decisions had overwritten reviewer 1's, and the kappa still computed, because it was measuring a column against itself. Recomputed from the raw decisions in stage 03: kappa 0.673, with 274 of 2,756 records where the two reviewers disagree. It is the 274 that matters.
- 51 of 122 downloaded "PDFs" were HTML login pages — 35 of them from one publisher, returned by URLs an open-access index had listed as free full text. Caught by magic number, not by file extension.
- Retrieval failure was not allowed to masquerade as exclusion. 561 records were never obtained and 189 were excluded on eligibility; the audit in stage 04 confirms zero records coded as one when they are the other. That distinction is the difference between an honest PRISMA diagram and a flattering one.
What it would not do
- Claim a result it did not have. The regimen that ranked first — trastuzumab + pertuzumab + chemotherapy, SUCRA 0.79 — did not reach significance against trastuzumab + chemotherapy (RR 1.82, 95% CrI 0.95–3.63), and the abstract says so in its first conclusion rather than leading with the ranking.
- Hide how little full text it could reach. 71 genuine full texts from 872 records advanced — 8.1% — the rest extracted from trial registries and abstracts or recorded as unavailable. The number is in the paper, not in a footnote.
- Report more than it measured. Abstract-only sourcing left survival data too sparse to pool, so this is a single-outcome network and says so.