Polinko is a human-led research system for observing AI behaviour through binary eval gates, signal traces, retained failures, and repo-native evidence.

The binary gate is the method's starting point because it aligns evaluation with binary computation. Before interpretation begins, Polinko keeps the first judgement simple: pass or fail.

krystian.io is the website doorway. This repository is the research surface.

The static website source lives in site/. Netlify builds it with npm run build and publishes dist/.

GPT-5.6 was used throughout this Build Week iteration to:

• help build make build-week-demo, the repo-native recording script that runs the preflight, live OCR binary eval, retained-evidence checks, and cleanup shown in the demo;
• support daily repo audits and reports while Polinko is refactored and staged for the next beta;
• run the recorded workflow live inside the Codex CLI; and
• help edit, clean, frame, and transcribe the final submission video.

How can binary eval gates help separate coherent AI output from reliable AI behaviour while keeping evidence boundaries visible?

AI responses are shaped by more than the prompt. Policy, guardrails, retrieval, memory, context limits, tooling, and prior response residue can all bend the path from intent to output.

Polinko treats visible mismatch as evidence. The method preserves failures, classifies them, and uses them to update the next research boundary instead of smoothing them away.

The first gate stays binary so the operational result remains visible before summary, interpretation, or pattern language can blur the boundary. Richer analysis begins only after the source evidence and retained failures are still traceable.

OCR is one pressure lane because the expected answer is externally checkable. It is one part of the broader research model.

• Beta 2.3 is the frozen method snapshot.
• pre-Beta 2.4 is staged as the next research-model contract.
• OCR is the mature green lane and is moving into generalisation pressure.
• Co-reasoning is the first promoted non-OCR lane.
• Retrieval, response behaviour, uncertainty boundary, and hallucination boundary are operationalised support surfaces.
• Operator burden is the active thin lane.
• Lean binary evaluation is staged as a sustainability question: whether reduced context clutter and correction churn can reduce unnecessary inference, with energy use, cooling demand, and water impact as downstream resource implications.
• Source-first row and case evidence is the pre-Beta 2.4 method foundation.

The current refactor is being handled as a staged path. The diagrams show the working loop and the route each major surface has taken through the cleanup.

• Refactor method: alignment, kernel scope, validation, docs, PR, merge, and clean main.
• Refactor journey: evidence baseline, runtime/package movement, manual eval workbench, and docs closeout.

make deps-install
cp .env.example .env
# set OPENAI_API_KEY in .env
make doctor-env
make docs

Use make docs-open only when you want to launch the system browser.

Apache-2.0. See license.