Submit midasavocado MIB document pipeline

submissions/midasavocado/MEMO.md

@@ -0,0 +1,127 @@
+# MIB Doc Challenge — Engineering Roadmap
+
+**Submission:** midasavocado
+**Solution:** <https://github.com/midasavocado/mib-doc-challenge-solution>
+
+This memo records how the system evolved, which shortcuts were rejected, what
+the frozen candidate actually proves, and where the next week of engineering
+should go. Runtime instructions and the component catalogue live in the public
+repository's README.
+
+```mermaid
+flowchart LR
+    A["OCR baseline"] --> B["Source-bound evidence"]
+    B --> C["Clean-room policy"]
+    C --> D["800 / 200 gate"]
+    D --> E["Bounded dual engine"]
+    E --> F["Frozen 1,000 + 5,000 runs"]
+    F --> G["Cross-fit and robustness work"]
+```
+
+## 1. From strings to evidence
+
+The initial extractor recovered many correct-looking values but flattened how
+they were obtained. A clean intake row, incidental policy prose, an OCR guess,
+and a foreign-case page could become the same string. Classification therefore
+sent many true approvals and denials to review: the emitted value looked
+complete, while its authority had been lost.
+
+The first durable change was to bind every observation to the active case,
+physical page type, labeled row, source program, and legibility state. Narrow
+rotation, deskew, faded-ink, and high-resolution readers improved extraction.
+More importantly, they distinguished observed, unreadable, absent, and
+contradictory evidence. Late field repair moved behind a frozen decision
+boundary so prettier output could not silently become policy authority.
+
+## 2. Clean-room policy rebuild
+
+An earlier participant-derived provenance package was removed. The active
+evidence audit, terminal policy, bridge, and writer were authored locally from
+the organizer's PDFs, schema, field manual, and evaluator. Engine A now follows
+a stable precedence: authenticated active-case finding; positive visible
+denial witness; material conflict or uncertainty; affirmative multisource
+approval; otherwise `NEEDS_REVIEW`.
+
+This phase also exposed seductive public residuals: name fragments, exact
+sponsors, tiny conjunctions, and hidden answer-like text could separate cases
+almost perfectly. Case-specific rules and failed trained models were deleted.
+The primary engine retained broad source topology, program authority, and
+symmetric safety vetoes instead.
+
+## 3. Generalization gate
+
+Development used a deterministic 800-packet partition. A separate 200-packet
+boundary returned aggregate section scores, validity counts, and catastrophic
+false-approval count only; it was not used for per-case diagnosis. The boundary
+is disclosed as repeatedly queried aggregate validation, not a pristine
+scientific holdout.
+
+A full-fit classifier looked excellent on its fitting rows but produced only
+72.20/80 classification and seven catastrophic false approvals in five
+internal 640/160 audits. Text, graph, neural, and residual-cell experiments
+were likewise rejected when they failed folds, encoded identity-like cells, or
+created approval without affirmative authority. The promoted Engine-A safety
+anchor measured 138.2286/150 on development and 135.1749/150 on aggregate
+validation, both with zero catastrophic false approvals.
+
+## 4. Bounded dual engine
+
+Engine B is a separately feature-flagged second opinion fit to all 1,000 public
+training labels. It contains two locally generated CatBoost heads and public
+residual policy hypotheses, but no case-ID answer map, validation labels, or
+manual output rows. It is explicitly benchmark-adaptive; private transfer is
+unproven.
+
+The first bridge allowed B to resolve every Engine-A review and replayed at
+146.5924/150 on the public artifact. That was a useful ceiling, not transfer
+evidence. The frozen arbiter is more conservative:
+
+- an Engine-A denial or authenticated approval always wins;
+- a decisive B result may resolve only an Engine-A review, after common fee,
+  risk, medical, conflict, and authority vetoes;
+- B abstention may demote an unsigned approval only in repeated identity-free
+  review families; and
+- after extraction freezes, a cache-backed refresh examines only materially
+  incomplete unsigned approvals. A refreshed B denial can produce review,
+  never denial or approval.
+
+Bridge confidence is not fixed at 0.90. A and B share inputs, so their signals
+are correlated. The arbiter discounts that correlation, combines A
+reliability, B strength, and evidence completeness, subtracts an approval-risk
+margin, and bounds bridge confidence between 0.62 and 0.93.
+
+![Conservative dual-engine architecture](https://raw.githubusercontent.com/midasavocado/mib-doc-challenge-solution/main/docs/assets/dual-engine-architecture.svg)
+
+## 5. Frozen release evidence
+
+| Boundary | Extraction | Classification | Calibration | Total | CFA |
+|---|---:|---:|---:|---:|---:|
+| Exact constrained public 1,000 | 46.9478 | 76.9800 | 18.3819 | **142.3097** | 0 |
+| Generalized Engine A, development 800 | 46.9028 | 73.4500 | 17.8758 | **138.2286** | 0 |
+| Aggregate-only 200 | 46.7389 | 71.5000 | 16.9360 | **135.1749** | 0 |
+
+The exact 1,000 run completed in 3,624.11 seconds (3.62411 seconds/PDF)
+under the organizer's offline 4-vCPU/8-GiB contract. The identical
+217,919,202-byte ARM64 image then processed all 5,000 unlabeled validation
+packets in 19,717.37 seconds (3.943474 seconds/PDF). The output contains 5,000
+unique complete rows, exactly matches the manifest, and passes both the
+organizer validator and an independent schema audit. Artifact SHA-256:
+`85ca045b1a5a652d6cc9d041966bee05cba17fc75675ef3be10ecccbb517b536`.
+No validation labels or private score were available.
+
+## Next week
+
+1. Replace public-fit Engine B with identity-free, source-state heads trained
+   and calibrated inside nested folds.
+2. Learn agreement, disagreement, and abstention confidence jointly rather
+   than calibrating a moving routing stack.
+3. Fuse the Tesseract and RapidOCR schedulers around one immutable raster cache
+   while retaining selective audit.
+4. Replace submission-wide vocabulary repair with a fixed development-derived
+   vocabulary so singleton and batch behavior are identical.
+5. Generate unseen layout and damage controls; require zero-CFA transfer before
+   granting any broader approval authority.
+
+The main lesson is gloriously unglamorous: durable gains came from preserving
+evidence provenance. Spectacular shortcuts usually melted when shown a held
+fold.

submissions/midasavocado/SUBMISSION.md

@@ -0,0 +1,52 @@
+# Submission — midasavocado
+
+**Public solution:** <https://github.com/midasavocado/mib-doc-challenge-solution>
+
+The root contains an MIT `LICENSE`, executable `run.sh`, and organizer-compatible
+`Dockerfile`. The image accepts exactly `<input_pdf_dir>
+<output_predictions_path>` and runs CPU-only with no network, API key, cloud
+OCR, LLM, VLM, or external service.
+
+## Default mode
+
+The default runs generalized Engine A with a bounded, public-training Engine B
+second opinion. Engine B cannot override an Engine-A denial or authenticated
+approval. It may resolve an Engine-A review only after deterministic safety
+vetoes. Its only reverse authority is fail-closed: B abstention or a late
+refreshed B denial may demote a qualifying unsigned approval to
+`NEEDS_REVIEW`, never to denial. Set `MIB_BENCHMARK_FIT_CLASSIFIER=0` for
+Engine-A-only operation.
+
+Engine B is disclosed as benchmark-adaptive. It uses public-label correlations
+and may not transfer; it contains no case-ID answer table, validation labels,
+or manually edited output rows. Selected native-PDF channels are also
+untrusted, disclosed, and separately ablatable. Visible or signed evidence
+retains deterministic veto authority.
+
+## Release verification
+
+The exact constrained public 1,000 run completed in 3,624.11 seconds
+(3.62411 seconds/PDF) with 1,000 valid complete rows. The evaluator measured:
+
+- extraction: 46.9478/50
+- classification: 76.9800/80
+- calibration: 18.3819/20
+- **total: 142.3097/150**
+- catastrophic false approvals: **0**
+
+The identical 217,919,202-byte ARM64 image completed all 5,000 validation PDFs
+under the organizer's offline 4-vCPU/8-GiB, read-only-root contract in
+19,717.37 seconds (**3.943474 seconds/PDF total**). The copied artifact has
+5,000 unique complete rows, zero missing/extra IDs, and passes
+`data/validation_manifest.csv` plus an independent exact-schema audit.
+
+- predictions size: 1,749,573 bytes
+- predictions SHA-256:
+  `85ca045b1a5a652d6cc9d041966bee05cba17fc75675ef3be10ecccbb517b536`
+- image ID:
+  `sha256:fc5c5eb8057d850b91e033ff3b49b28016afbd14cc4879b1264ba340c635bded`
+
+No validation labels or private score were available or inferred. The roadmap,
+failure modes, generalization boundary, and next-week plan are in `MEMO.md`.
+Third-party notices are retained in the solution repository under
+`third_party_licenses/`.