---
name: verify-counts-two-ways-mechanical-sweeps-are-finicky-in-this-shell
description: grep tallies / frontmatter scans misfire in this zsh (glob quirks; all-tree
  greps catch HTML class= + code-block noise). Verify every count two ways before
  it enters a record; and don't dismiss correct data on a prior expectation. Three
  near-false-findings caught this way 2026-05-29.
metadata:
  node_type: memory
  type: feedback
  originSessionId: a8eb1d46-314e-4545-a133-f747c69ad5b4
permalink: claude-memory/feedback-verify-counts-two-ways
---

Mechanical corpus sweeps (class-tallies, frontmatter scans, "does X exist" greps) are **unreliable in this zsh shell** and must be cross-checked before any count enters an audit or a record.

**Why:** observed repeatedly on 2026-05-29 during the code-represents-spec audit —
- `--include=*.md` unquoted → zsh "no matches found" (glob expansion); the tally silently didn't run.
- `grep -rhoE` across all of `content/` caught non-frontmatter `class:` (HTML `class=` attributes, code blocks) → spurious "types" (offering, lex, not-found, sequence-index) that looked like clause-1 findings.
- Two runs of "the same" tally disagreed wildly (124 essay vs 193 glimpse) — one was noise, one was right.

**How to apply:**
- Prefer **per-pattern `grep -lE '^class:[[:space:]]*X' files | wc -l`** (count files matching one exact pattern) over `-o`/`-m1`/`sed` pipelines.
- Quote globs: `--include='*.md'`.
- Scope explicitly (e.g. `content/posts` + `content/pages`, not all of `content/` with its chamber HTML).
- **Confirm any surprising count a second way** before reporting it — *both* directions: don't assert a finding from a noisy tally, and don't *dismiss* a correct tally because it violates a prior expectation (the "essay must be dominant" anchor nearly buried the true glimpse-dominant count).

The cost is real (tool-churn), but it is the "once, correctly" tax. Related: [[feedback-trust-prior-pass-frame]] (verify-before-compose); the logged drift `asserting-fs-state-from-a-misread-listing`.