RATIN

RATIN converts mixed-format procurement documents into an evidence-linked decision brief with exact page, text-span, and spreadsheet-cell navigation.

Comment

Inspiration

The core idea is simple: every AI-generated finding should be directly verifiable at the exact source sentence, bounding box, or spreadsheet cell. Procurement decisions often depend on information scattered across PDFs, HWP documents, spreadsheets, public notices, and external reference data. Even when an AI produces a useful summary, users still need to verify where every important statement came from.

RATIN was built to close that gap: not just summarize documents, but connect every decision-relevant finding back to exact source evidence.

What it does

RATIN transforms a mixed-format procurement package into an evidence-linked decision brief.

The interface uses a Dual A4 layout:

  • The left pane presents bid metadata, participation requirements, cost references, conflicts, risks, questions for the issuer, and required actions.
  • The right pane preserves the complete Evidence PDF stream.
  • Clicking an Evidence link opens the exact PDF page, text bounding box, or spreadsheet-cell location.
  • Multi-target evidence can highlight several physical source locations from one logical finding.

The final validated public build contains:

  • 21/21 logical Evidence jumps
  • 26 exact physical text or cell targets
  • 0 fallback jumps
  • 18/18 automated tests passed

How we built it

RATIN separates deterministic document processing from model-based interpretation.

The deterministic layer owns:

  • mixed-format document normalization
  • source identity and provenance
  • PDF page and bounding-box binding
  • XLSX cell binding
  • quantities, prices, and arithmetic
  • exact Evidence navigation

GPT-5.6 is connected through a strict structured-output contract. Its role is to interpret already verified evidence, identify cross-document conflicts and risks, and generate structured questions and required actions. It does not invent source coordinates, prices, quantities, or document facts.

The public judging path uses a validated frozen cache for reproducibility and works without an API key. An optional live GPT-5.6 path is also included and fails closed when schema or semantic validation does not pass.

How Codex was used

Codex was used extensively during Build Week to navigate a large existing repository and turn the underlying normalization results into a focused downstream product.

Codex helped:

  • build the standalone Build Week consumer
  • connect PDF bounding boxes and XLSX cells to the UI
  • implement single-target and multi-target exact Evidence jumps
  • restore and extend the Dual A4 interface
  • create the continuous Evidence Stream
  • integrate the GPT-5.6 structured-output contract
  • implement live, cached, and fixture execution paths
  • run browser QA and automated validation
  • prepare the sanitized public repository and reproducible judging package

The primary Codex session preserved the existing proprietary pipeline while rapidly implementing and validating the new Build Week workflow.

What existed before Build Week

Before Build Week, the project already had a proprietary document-normalization and evidence-binding foundation, including research around mixed-format documents and source-coordinate mapping.

That proprietary canonical pipeline is not included in the public repository.

What was built or materially extended during Build Week

During Build Week, we built or materially extended:

  • the standalone downstream RATIN consumer
  • the Dual A4 Decision Brief and continuous Evidence Stream
  • exact PDF page and bounding-box navigation
  • exact XLSX cell navigation
  • multi-target Evidence handling
  • the GPT-5.6 strict structured-output workflow
  • fail-closed live, cached, and fixture modes
  • browser-level Evidence-jump validation
  • the sanitized public demo bundle
  • automated tests, security audits, and judging documentation

Challenges

The hardest problem was preserving trustworthy provenance across different document formats.

A PDF sentence, an HWP-derived Evidence page, and an XLSX cell all require different coordinate and identity models. RATIN had to normalize those differences without hiding them from the user.

Final provenance review also removed one derived public-price-to-XLSX link because the API item did not exactly match the spreadsheet item. The video shows the earlier 22/22 build; the final validated public build contains 21/21 logical links resolving to 26 exact physical targets. Both builds use zero fallback jumps.

This reinforced an important principle: removing an uncertain Evidence link is better than presenting a convenient but unsupported connection.

Accomplishments

  • Mixed PDF, HWP, and XLSX package support
  • Exact page, text-span, bounding-box, and spreadsheet-cell navigation
  • 21/21 validated logical Evidence jumps
  • 26 exact physical targets
  • 0 fallback or unverified jumps
  • 18/18 automated tests passed
  • API-key-free cached judging mode
  • Fresh-clone reproducibility verified
  • Secret, personal-information, internal-path, and private-source audits passed

What we learned

Reliable document intelligence requires a strict boundary between evidence and interpretation.

The model should explain verified evidence, not become the source of coordinates, prices, or document truth. Exact provenance, fail-closed validation, and transparent limitations make the resulting AI system substantially more useful for real decisions.

What's next

Next steps include:

  • testing broader real-world procurement packages
  • expanding production-grade live GPT-5.6 validation
  • supporting additional document and table structures
  • connecting bidder profiles to eligibility analysis
  • extending evidence-linked decision intelligence beyond procurement.

Prior controlled validation

In a separate frozen GPT-5.4 controlled benchmark, anchor-routed inputs reduced input tokens by 84.15%, estimated API cost by 77.91%, and API response latency by 52.9%–72.7%, while preserving issue-level F1 at 1.0.

The benchmark’s strict overall gate was not passed because one semantic-equivalence case correctly returned insufficient evidence when subject context was missing. These figures are prior controlled-validation results, not measurements from the current GPT-5.6 Build Week demo.

Built With

Share this project:

Updates