board: E-OCR-SKIP-GATE-1 — skip cascade ABORTED (28.3%), batching is the converged lever

.claude/board/EPIPHANIES.md

@@ -1,3 +1,14 @@
+## 2026-07-08 — E-OCR-SKIP-GATE-1 — the go/no-go gate rules the skip cascade MARGINAL with a number (30.1%, far below the 60% proceed bar): probe-first discipline stops a research swamp before a line of cascade is built
+**Status:** FINDING (measured; `tesseract-rs` `skip_fraction` instrument; PR #24)
+
+**The operator hardening that produced this:** the "Morton-tile-pyramid-CAM-PQ recognizes tokens" idea was correctly RESCOPED — not "replace the LSTM" (segmentation-free CTC + font-invariance are load-bearing) but "AVOID LSTM invocations" via a skip/escalate cascade: cheap coarse descriptor → empty (skip) / trivial-known-token (CAM-PQ fast-path) / ambiguous (ESCALATE to the exact LSTM). CAM-PQ/VSA = a CONFIDENCE GATE, never the OCR engine; the LSTM stays the truth oracle for the hard tail (always-escalate preserves byte-parity). And the iron discipline: **first task is instrumentation ONLY** — measure the skippable fraction before building anything; abort < 30%, proceed >= 60%.
+
+**The measurement (instrumentation only, recognizes nothing):** `skip_fraction.rs` over the 10 committed pages counts, within each line-band ink span, the pure-blank interior column fraction (the safely-skippable floor for a blank-run gate; preserved under the front-end 2× maxpool so it proxies blank LSTM timesteps; same Otsu+binarize as the real path). **Aggregate 30.1% (6772/22522), per-page ~21-41% → MARGINAL** (30-60%, far below the 60% PROCEED bar). On dense typeset text the easy skip is barely there. (First cut read 28.3% by omitting the recognizer's blank 4px x-padding from the tally — codex P1; counting the ACTUAL fed crop raised it to 30.1% and flipped the LABEL ABORT→MARGINAL. CONCLUSION unchanged: 30% << 60%, cascade still not a clear win, batching stays the lever — reported faithfully.)
+
+**The convergence — two independent probes name the SAME lever:** (1) the arithmetic-intensity estimate — eng.lstm forward runs at ~0.1% of AVX-512-VNNI peak, i.e. the int8 GEMM is STARVED by per-tiny-op glue (pack_contig, dispatch, requant), not saturated arithmetic; and (2) this 28.3% skip fraction. Both point to **BATCHING the forward** (width-bucket lines → batch-across-lines-per-timestep so the recurrence holds → fill the tiles → fewer, larger cache-friendly kernel calls → scatter), NOT skipping it and NOT more SIMD. The "AMX-Ferrari bekommt Pixelkrümel mit der Pinzette" diagnosis is confirmed: the kernels are fast, the FEED is tiny/irregular. Batching is the developer-army-replacement — one dense feeder vs Tesseract's thousand hand-tuned micro-heuristics.
+
+**Transferable:** the probe-first gate is the anti-research-swamp mechanism — a one-hour instrument that says NO with a number is worth more than a month of visionary cascade-building. Corpus caveat: dense text is the skip worst-case; the instrument re-runs on any corpus (sparse forms/tables would score far higher — the cascade's real home). Composes with E-OCR-PARALLEL-1: rayon spreads work across cores (outer), batching fills each core's tiles (inner) — multiplicative, and the measured next lever. Branch `claude/happy-hamilton-0azlw4`.
+
 ## 2026-07-08 — E-OCR-PARALLEL-1 — page-chunk rayon closes the throughput gap losslessly: 3.34× on 4 cores, byte-identical to serial (the operator-pinned topology, not whole-doc lanes / not a global crop pool)
 **Status:** FINDING (measured; tesseract-rs `ocr-pdf`, determinism-gated; PR forthcoming)