---
name: read-source-at-the-measured-sha-not-the-default-tree
description: "A defect measured at sha X must be read at sha X — the default tree answers a different product, silently and with plausible line numbers."
metadata: 
  node_type: memory
  type: feedback
  originSessionId: 2b6a02a4-d35d-443f-a4e1-98178d6cf69b
  modified: 2026-08-19T11:58:49.403Z
---

Handed an RCA whose evidence was an x20 at `bd942f6`, I ran Grep/Read against the project root.
`main@80ab77a` did **not** carry that work: `git diff bd942f6 80ab77a` = −349 lines across `rc.rs`,
`broker.rs` and the failing cell. The lane worktree I had source-read earlier (`keystone-w1@ba2d998`)
differed again (`broker.rs` +125, the cell +81). Three trees, three products, one question.

**Why:** it fails SILENTLY and CONVINCINGLY. The function exists, the grep hits, the line numbers land
near the ones in the brief — a couple did match exactly — so nothing announces that you are reading a
different product. I had already read one arm's helper from the wrong tree before checking, and its
`sid_after` line was a plain read where the measured tree polls a 10s rotation. That 10s was half the
arithmetic that answered the flake's bimodality; from the default tree the answer is not there to find.

**How to apply:** the FIRST act of any source read for a measured defect is `git worktree list` +
`git diff --stat <measured-sha> HEAD -- <the files>`. If it is non-empty, extract blobs
(`git show <sha>:<path> > scratchpad/...`) and read those — do not `cd` into someone else's rig
(see [[a-bare-cd-silently-repoints-every-later-read]]). State the sha in the design so a reviewer can
falsify the reading. A brief that names a sha is naming the tree, not just the evidence.
Sibling of [[authored-unlanded-text-lives-in-the-lane-index]]: on a fleet with 20+ live worktrees,
"the repo" is never a location.
- 2026-09-07 face (hertz, IR-82 draft): line refs drafted off the WORKING TREE were ~39 lines off in cli.rs while EVERY daemon.rs number agreed with the measured sha — a PARTIAL agreement reads as confirmation, which is what makes the mixed read dangerous. Re-derive every site at the sha; agreement on N-1 files proves nothing about the Nth.