diff --git a/FINDINGS.md b/FINDINGS.md
index c7ceecb..7521f07 100644
--- a/FINDINGS.md
+++ b/FINDINGS.md
@@ -1,13 +1,52 @@
 # FINDINGS — pxpipe (text→PNG token compression)
 
 **Status:** ⚠️ **VERDICT REVERSED — see correction below.** Originally ruled "dead"; live measurement shows pxpipe is a working *lossy gist-compressor* saving ~68% on real (dense) Claude Code traffic, with a known verbatim-recall gap.
-**Date:** 2026-05-28 (original) · 2026-05-29 (correction) · 2026-06-09 (Fable 5 update) · 2026-06-10 (gist-recall A/B, SWE-bench pilot)
+**Date:** 2026-05-28 (original) · 2026-05-29 (correction) · 2026-06-09 (Fable 5 update) · 2026-06-10 (gist-recall A/B, SWE-bench pilot) · 2026-06-12 (field observation, n=1)
 **Models tested:** `claude-opus-4-5` (original run), `claude-opus-4-8` (re-test after a model bump), `claude-fable-5` (2026-06-09)
 **Model scope (current):** Fable 5 only, enforced in library + proxy (Opus disabled 2026-06-09 — see update below).
 **Harness:** `eval/needle-haystack/` (receipts preserved from `/tmp/needle_eval`)
 
 ---
 
+## Update (2026-06-12) — field observation: live verbatim misreads in a real session (n=1, anecdotal)
+
+**Not a controlled eval.** Logged because it is the first *unprompted,
+in-the-wild* capture of the exact boundary the harnesses bracket, with
+ground truth available for post-hoc grading.
+
+Setup: a working session was resumed from imaged history (older context
+rendered to PNG pages, recent turns text — the production architecture).
+Mid-session, the assistant was challenged to transcribe, from the imaged
+history alone, ten 12-char hex IDs that had appeared in an earlier tool
+output. The transcription was committed *before* re-reading the source
+files on disk, which then served as ground truth.
+
+| tier | result |
+|---|---|
+| gist (resume a paused run: plan, counts, file states, next steps) | fully functional — zero observed errors, consistent with gist-recall 98/98 |
+| verbatim (10 × 12-char hex through imaged history) | **7/10 exact** · 4 char-level errors / 120 chars (~3.3%) |
+
+Misread detail: `0→8`, `e→4`, `e→8` substitutions plus one dropped `d` —
+the substitutions are the closed-counterform confusion classes the planned
+glyph-matrix experiment targets (see `eval/glyph-matrix/PLAN.md`). **2 of
+the 3 misses were silent** — stated confidently, no hedge — matching the
+needle-haystack failure mode (*silent confabulation*: plausible wrong hex,
+not an error). The one read the assistant self-flagged as low-confidence
+was indeed wrong (a weak calibration signal, n=1).
+
+**What this adds:** a live demonstration that flawless gist-tier recall and
+silent verbatim-tier corruption co-occur *in the same session, on the same
+imaged pages* — the boundary holds in the field exactly where the harness
+drew it (gist = functional, exact-bytes = unsafe, failure = silent).
+
+**Caveats:** n=1 session; the reader was the interactive session model, not
+the fable-5 production gate; render parameters of the imaged history were
+not recorded; session-internal images make this non-reproducible. Treat as
+illustration/motivation only — the numbers that matter come from
+`eval/glyph-matrix/` when it completes.
+
+---
+
 ## Update (2026-06-10) — gist-recall A/B: zero information loss measured at the gist tier
 
 The README's weakest claim ("useful when you know where to blur") had no
diff --git a/eval/glyph-matrix/PLAN.md b/eval/glyph-matrix/PLAN.md
new file mode 100644
index 0000000..c5ae2c4
--- /dev/null
+++ b/eval/glyph-matrix/PLAN.md
@@ -0,0 +1,46 @@
+# Glyph confusion matrix + render-style A/B (Task #7) — PLANNED, paused for usage budget
+
+**Status:** paused 2026-06-12 at 05:45am · resume after weekly limit reset (Jun 18 ~5:59pm Toronto)
+**Working state:** `~/glyphmx-paused/` (canonical backup, 330 files, 21MB) · scratch `/tmp/glyphmx` · resume via `run.sh` (limit-aware, skips banked trials)
+**Detailed handoff:** `HANDOFF.md` (repo root)
+
+## What this measures
+
+Per-character confusion matrix for 12-char hex IDs read back through the
+pixelpipe render path, across render-style arms — the precision-tier
+failure mode that gist-recall (98/98) structurally cannot surface and
+needle-haystack only bracketed (~87% exact-match on worst-case hex at
+prod style).
+
+- **40 trials/arm × 5 arms** (prod 500px baseline + style A/B first pass: onebit, color, grid, + winner expansion to 30 later)
+- Each trial: 1 page, ~222 image tokens, 40 dense JSON lines, 5 labeled 12-char hex IDs (LCG-seeded, seed 20260612) + 8-hex distractors
+- Reader: `claude-fable-5` only (production gate model), direct API route
+- Grading: exact-hit rate per trial (12/12 chars), per-char confusion matrix, pooled misses; hard-fails on empty/poisoned output graded as misses
+
+## Banked so far
+
+- 2/40 prod-arm trials valid (`out_prod_0.txt`, `out_prod_1.txt` in `~/glyphmx-paused/`), not poisoned; run halted on session limit (`HALTED-ON-LIMIT` in `run.log`)
+- All 120 rendered pages for the first pass are pre-rendered and backed up — resume only spends read calls, not render work
+
+## Why it's not in README yet
+
+README rule for this repo: measured numbers + n only, no marketing tone.
+This experiment ships to README **after** the matrix completes — as
+"at render style X, content carries at ~N× fewer input tokens with
+measured per-character read error E% on worst-case zero-redundancy
+content." Until then it lives here so the plan is public without an
+unearned headline.
+
+## Decision criteria (when trials finish)
+
+1. Per-arm exact-hit rate (5 labels vs golds), per-char confusion on misses, which hex glyph pairs confuse (0↔8, e↔4, e↔8, f/t suspects).
+2. If any zero-cost style beats prod meaningfully (and prod ≈ known ~87%), recommend flipping default in `src/core/render.ts` (e.g. `DENSE_RENDER_STYLE` gains `colorCycle:true`) — token-cost-neutral, but re-run harness to confirm before changing defaults.
+3. Update README verbatim row + FINDINGS.md dated entry; copy receipts (`golds.json`, outs, `grade.py`) into `eval/glyph-matrix/` and commit.
+
+## Related field evidence
+
+See FINDINGS.md "Update (2026-06-12) — field observation": an n=1
+in-the-wild session captured 7/10 exact / ~3.3% per-char misreads on the
+same confusion classes (0→8, e→4/8), 2 of 3 misses silent. Anecdotal,
+non-reproducible — motivation only; this harness is what produces the
+publishable numbers.