Add submission for dhyantsoni

[dhyantsoni]

Adds submissions/dhyantsoni/ — validation predictions, technical memo, and a link to the solution repository. Nothing outside that folder is touched.

Solution repository: https://github.com/dhyantsoni/mib-doc-challenge/tree/main (Dockerfile at root)

Approach

Every stage asks one question: is this evidence I'm allowed to use?

• Ink verification. Each text span is checked against the rendered raster; a span that leaves no ink is not evidence. One measurement subsumes white-on-white text, off-crop spans, invisible render modes, and occlusion, rather than a heuristic per trick.
• Template- and label-anchored extraction. A value is recorded only when a recognised form yields a recognised label followed by a legal value, so a free-floating SYSTEM: ... answer key line has no path into a record field even when printed in visible ink. Binding each field to the templates that print it also stops a barcode payload from fuzzy-matching its way into a waiver code.
• Evidence ledger with the field manual's precedence order, case-id scoping for multi-applicant packets, and conflicts counted rather than averaged.
• Dual-mode OCR for the ~47% of pages that arrive as degraded scans, read at three contrast cuts under two segmentation modes and merged.
• Decision layer: the manual's unambiguous clauses as hard constraints, a regularised model for the residue, and action selection that maximises expected score under the published payoff matrix — which differs from argmax-probability, since a wrong denial costs 0 while a false approval costs −4.

Measurement

Scores are cross-validated, not in-sample. Evaluating the fitted model on its own training packets read 15 points of accuracy too high and understated catastrophic false approvals by 18x; tools/honest_eval.py is the only thing quoted from.

Extraction	41.1 / 50
Classification	65.6 / 80
Calibration	15.8 / 20
Total	122.5 / 150
Held-out adjudication accuracy	81.4%
Runtime	1.36 s per PDF against a 6 s budget

Validation output: 5,000 predictions, all schema-valid, none missing. Runs offline on CPU with no LLM, VLM, or hosted API; the only model artifacts are Tesseract's bundled English data and a classifier under 5 KB trained on the public training split.

The memo covers failure modes and what a further week would go into.

[Copilot]

Pull request overview

Adds a new participant submission under submissions/dhyantsoni/, including the submission landing page and a detailed technical memo describing the approach and reported validation results.

Changes:

• Added SUBMISSION.md with build/run instructions and a contents inventory pointing to the solution repository.
• Added MEMO.md documenting the pipeline design, measurement methodology, and failure modes.

Reviewed changes

Copilot reviewed 2 out of 3 changed files in this pull request and generated 1 comment.

File	Description
submissions/dhyantsoni/SUBMISSION.md	Submission entrypoint with repository link, run commands, and artifact summary.
submissions/dhyantsoni/MEMO.md	Technical memo describing the solution approach, evaluation, and known limitations.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.