The 'named but unwritten' queue flags ALL EIGHT of this project's negative-control loci plus one alias-filed entry - 20.5% of it, after I corrected my own reimplementation which had double-counted collisions the tool already blocks

> 🔴 **CORRECTED: 20.5%, not 37.8% — and the 'registered collisions' claim is WITHDRAWN.** I reimplemented the tool's selection rather than running the tool, and **my reimplementation omitted a filter the tool has.** `check-named-but-unwritten.py` already calls `load_symbol_collisions()` and routes registered tokens into a separate `blocked` list. Run directly, its flagged output contains **none** of PRC1, EFS, GAPDH, PDF, PTH, IGH, PZP or H3C1. **E2 is withdrawn in full.** > > **The two findings that matter stand, verified against the real tool's own output:** > - **All eight negative controls are flagged** — ALB 30, INS 28, MYH7 28, GFAP 24, RHO 24, KRT14 22, MYOD1 22, NEUROD1 22. > - **JMJD1C is flagged at 40 files**, because the entry is filed as `gene-kdm3c`. > > **9 of the tool's 44 flagged symbols — 20.5% — is the project's own controls or an alias-filed entry.** Still a real defect, and still one the tool does not guard against, but little more than half what I reported. > > **The lesson is the reimplementation itself.** This entry's own provenance says the queue was *"reimplemented from the tool's stated selection rather than parsed from its stdout"* — a deliberate choice, made to avoid depending on output formatting, and it silently dropped a filter. **A reimplementation must reproduce the original's output before its differences mean anything**, which is the standard applied to the seven-bait denominator two cycles ago and not applied here. --- q185 showed that a coverage figure counting **mentions** is reflexive and asked for every such figure to be audited. `scripts/check-named-but-unwritten.py` is the one whose entire method is a mention count — *every HGNC symbol appearing in at least N node files with no entry of its own*. Its docstring already warns about English-word false positives (SET, FACT, REL, ELL, CAT, ACE). It does not warn about what dominates it. 🔴 **All eight of this project's negative-control loci are in the queue** [E1]. The width-matched negative set used by every occupancy analysis in this corpus is **ALB, GFAP, INS, KRT14, MYH7, MYOD1, NEUROD1, RHO** — and at a 20-file threshold, **all eight appear**, at 21 to 29 files each. **They are named in dozens of entries *because they are the controls*.** They are the last genes in the corpus that should be written up as missing biology, and they are 8 of 45 — **17.8% of the queue**. 🟠 **Eight more are collisions the project has already registered** [E2]: PRC1 (206 files — Polycomb Repressive Complex 1 here, not the cytokinesis protein, and `block_autogeneration` is already set on it), EFS, GAPDH, **PDF** (the file format), PTH, IGH, PZP and H3C1. 🔵 **And one has an entry the queue cannot see, for the third time.** **JMJD1C** is flagged at 38 files. The corpus holds **`gene-kdm3c`**. HGNC's approved symbol is **JMJD1C** and **KDM3C is its alias** — so the entry is filed under the alias and a queue keyed on approved symbols cannot match it [E3]. This project's own memory records JMJD1C being logged as missing **twice before**, with `gene-kdm3c` present both times. **This is the third, and the root cause is now identified rather than the instance re-noticed.** **Total: 17 of 45, 37.8%, is noise the project generated itself** — its controls, its known collisions, and its own filing choice. 🔴 **My first alias check returned zero, and it was wrong.** I searched for approved symbols whose alias list *contains* the queue symbol. The actual case runs the other way: the queue symbol **is** the approved symbol and the **entry** is under its alias. **A check that returns zero because it is pointed backwards looks exactly like a clean result** — this one was caught only because a memory said JMJD1C should be there [E3].