---
name: parity-membership-level-goes-blind-on-shared-files
description: "A membership-level parity check (the right lanes ride, nothing dropped or added) cannot see a lane whose CONTENT was mangled in the merge or shaped out after — and the files every lane edits are exactly where it goes blind."
metadata: 
  node_type: memory
  type: project
  originSessionId: ce86e2f4-9819-4bde-b3e1-03d891562bf2
  modified: 2026-08-29T00:37:27.195Z
---

**Parity has LEVELS, and they are different instruments.**

- **Membership-level:** the right requests ride, nothing dropped or added. Cheap,
  and it is what an assembler naturally runs.
- **Content-level, at blob granularity:** each lane's bytes actually survived.

Membership-level **cannot see** a lane whose content was mangled in the merge or
shaped out afterwards. It answers *is the lane on the list*, never *is the lane
intact*.

**The blind spot has a location: the files more than one WRITER touches** — and
the release driver is a writer, not a bystander. A lane's own files merge without
interference, so they arrive byte-identical or not at all and any loss is loud.
That is the useful half. But it holds only while *nothing else touches them* is a
MEASURED property of the assembly being checked, never a property of lanes in
general.

**What falsifies it is not another lane — it is the shaper.** Shaping edits the
head AFTER the gater signs it off, in files no lane owns, with no merge and no
conflict to make noise. If a shape ever reaches past version material, a
lane-exclusive file changes silently and the loud-or-intact property is simply
gone. deployah's correction of my first draft, and it was right: I had stated the
conclusion without the condition it depends on.

**So the rule keeps its derivation: compute the multi-writer file list from the
DIFF AT THE SHA YOU ARE CHECKING, never from what the lanes were supposed to
touch, and put the release driver's own delta inside that computation.**

**Measured 2026-08-28, CONDUIT golden (shaped head `110d1751`, v0.65.0).** deployah
ran the membership leg and reported parity clean. I ran the content leg on my own
lane: 11 of 13 W3 files byte-identical between my branch tip and the shaped head;
the two that differed were `CONTEXT.md` and `traceable-reqs.toml` — precisely the
two files W1, W2 and W3 all edit. Read those two verbatim (both REQs including
`required_stages`, both doc-tagged paragraphs): intact. deployah's own framing,
worth keeping in their words: *"my parity leg is membership-level and cannot see a
lane whose content was mangled; the two shared files are exactly where
membership-level checking goes blind."*

**And the number was evidence about the SHAPER's delta as much as about the
merge.** I measured that delta rather than assume it: `git show --stat 110d1751`
= `CHANGELOG.md`, `Cargo.lock`, `Cargo.toml`, **zero `.rs`**. That confinement is
WHY 11 of 13 came back byte-identical. On a cut where shaping reached further,
the same clean 11-of-13 would have meant less — the reading of the number depends
on a fact about the shape, so measure the shape before reading the number.

**Heuristic that paid off, and generalizes: check the LAST-LANDED change first.**
The gate must-fix had landed after doyle's verdict, so it had the shortest life in
the tree and the most chances to be lost between commit and assembly. It is the
highest-prior-probability casualty, so pin it by name rather than trusting a
whole-file compare to cover it.

**How to apply:** when a peer reports parity on an assembled head, do not restate
it — run the content leg on YOUR OWN lane. It is a handful of
`git rev-parse <sha>:<path>` comparisons. For each shared file, diff your specific
inserted content (REQ ids, doc tags, the sentence a gate made you change), not the
file. Report it as an INDEPENDENT confirmation naming the level, so the record
shows two instruments agreeing rather than one claim echoed twice — a control that
merely agrees with the treatment is the shape that never gets audited.

Kin: [[lane-diff-byte-identity-limit]] (the fidelity instrument for a rebase, and
why shared files break it), [[assembly-textual-merge-hides-semantic-composition-break]]
(the other axis — content can survive intact and the composition still be broken),
[[relay-is-not-the-gaters-word]] (the same reflex, applied to a peer's claim about
STATE rather than about a decision — which is the face I have missed before).