Submission: patchg

submissions/patchg/MEMO.md

@@ -0,0 +1,51 @@
+# Technical Memo — MIB Doc Challenge
+
+**Author:** patchg  
+**Solution freeze:** v12.1 · train score **122.74 / 150** · **0 catastrophic false approvals** · **~4.0 s/PDF** under official Docker limits  
+**Runtime:** Offline Docker · Tesseract OCR · classical CV + rules (no network, no cloud APIs)
+
+## Approach
+
+I built a full offline pipeline that turns messy multi-page PDF packets into structured records and a three-way adjudication (`APPROVED` / `DENIED` / `NEEDS_REVIEW`).
+
+**1. Perception.** Each page is rendered (~200 DPI), optionally deskewed and tear-corrected, then OCR’d with Tesseract and merged with the native PDF text layer. Trust filtering strips injection / “answer key” style noise before fields are read.
+
+**2. Structure.** Text is segmented into typed components (form, B-13 biometric slip, fee receipt, registry, sponsor letter). Fields use a three-state evidence model: *read-and-trusted*, *read-ambiguous*, or *not read*. Risk flags are only trusted from B-13 (or labeled flag lines), never invented as `none` from a blank page.
+
+**3. Policy (deny-first).** Manual inspection stamps short-circuit when present. Otherwise FR6-style rules fire on trusted evidence: disqualifying flags, unpaid fees, TRANSIT-7, registry embargo, revoked sponsors (with corroboration), restricted homes. Approvals require a clean, complete evidence signature (including B-13-backed flags none and real fee evidence)—soft defaults used only for extraction emission, never to unlock approve.
+
+**4. Emission & calibration.** After policy, emit-only enrichments improve field strings (name cleanup, hatch → illegible biometrics, fee labels) without changing adjudication. Confidence is mapped by decision reason (deny high, incomplete low) for calibration points.
+
+**5. Ship path.** A single slim image (`Dockerfile` + `run.sh` + `config/default.yaml` + package) runs under `--network none --read-only --cpus 4 --memory 8g` with four workers and OpenMP thread limits so average time stays under the 6 s/PDF budget.
+
+## What worked
+
+- **Safety first:** keeping catastrophic false approvals at zero on full train by refusing soft “packet looks complete → APPROVED” rules.
+- **Asymmetric flag recovery:** parse labeled lines such as garbled `Disqualifying risk flag:` and truncated `O flags:` into deny tokens, without free-text answer-key harvest.
+- **Sponsor / multi-page corroboration and geometry (deskew/tear)** that improve extraction without special-casing case IDs.
+- **Measured plateaus:** PaddleOCR and heavier B-13 multi-ROI on hard flag-miss slices added no new tokens—so we stopped spending budget there.
+
+## Failure modes
+
+| Mode | Symptom | Why it remains |
+|------|---------|----------------|
+| Blank B-13 / image-only flags | Gold disq flag, we emit `none` or NR | True OCR ceiling on many train residuals |
+| Soft fee | Gold paid, we `unknown` | No receipt text; inventing paid would reopen FA risk |
+| Gold APPROVED → our NR | Clean-looking fields, no trusted B-13 none | Correct conservative behavior; complete-packet A is FA-toxic |
+| Known false denial | e.g. fee OCR as unpaid vs gold paid | Rare; locked with regression tests |
+| Name OCR noise | Near-miss spellings | Open vocabulary; cleanup helps tails, not full identity |
+
+## With another week
+
+1. Layout-aware B-13 / fee **panel detection** (classical or tiny offline detector) before OCR—only if a canary slice shows tokens appear under crops we don’t take today.  
+2. Stronger **restricted-home / revoked-sponsor** residual when home/SPN are RT but still NR.  
+3. Calibration split by incomplete subtype (flags NR vs fee NR) for a small Brier gain.  
+4. Broader validation monitoring (latency histogram, missing-case rate) under the 5k-PDF wall clock.
+
+I would **not** spend another week on soft approve paths or train-label hardcoding: the private test set and FA gates make those false wins.
+
+## Design principles
+
+- **Evidence over labels:** no case-ID special cases; rules must fire on text/geometry available at runtime.  
+- **Asymmetry:** deny tokens can promote; recovered `none` never opens approve.  
+- **Measure before ship:** full-train Docker freeze under challenge flags before submission.

submissions/patchg/SUBMISSION.md

@@ -0,0 +1,44 @@
+# Submission — patchg
+
+## Solution repository
+
+**Public repo:** https://github.com/patchg/mib-doc-solution  
+
+Includes:
+- `Dockerfile` + `run.sh` (two-arg offline contract)
+- **MIT `LICENSE` at repository root**
+- v12.1 freeze pipeline (Tesseract + rules, no network)
+
+## Docker
+
+```bash
+docker build -t mib-submission .
+docker run --rm \
+  --network none \
+  --cpus 4 \
+  --memory 8g \
+  --pids-limit 512 \
+  --read-only \
+  --tmpfs /tmp:rw,nosuid,nodev,size=2g \
+  --mount type=bind,src=/path/to/pdfs,dst=/input,readonly \
+  --mount type=bind,src=/path/to/output,dst=/output \
+  mib-submission /input /output/predictions.jsonl
+```
+
+## Freeze summary (public train labels)
+
+| Item | Value |
+|------|-------|
+| Version | v12.1 |
+| Train score | **122.74 / 150** |
+| Catastrophic FA | **0** |
+| Throughput | **~3.98 s/PDF** under challenge Docker flags |
+| Image size | ~190 MiB |
+
+## This folder
+
+| File | Contents |
+|------|----------|
+| `predictions.jsonl` | Validation-set predictions (5000 cases) |
+| `MEMO.md` | Technical memo |
+| `SUBMISSION.md` | This file |