What it does
CalibraLoop turns a natural-language quality goal into a bounded, auditable calibration loop over a fully synthetic optical field. GPT-5.6 decides what to inspect and adjust; deterministic Python code owns every measurement, safety limit, pass/fail decision, state transition, and report.
The default scenario begins with uneven illumination, soft focus, a power constraint, and a hidden spectral-efficiency fault. The agent must diagnose the field, choose useful probes, make the smallest bounded adjustments, re-measure, and revise its plan when a metric remains stalled.
Why it matters
Calibration work often breaks down between an operator's intent and several specialized measurements. CalibraLoop gives teams one reviewable loop:
goal → measure → diagnose → probe or adjust → verify → audit report
It demonstrates how an AI agent can provide flexible judgment while deterministic software remains the authority for safety and acceptance criteria.
How it was built
- GPT-5.6 interprets goals, compares failed metrics, chooses focus or spectrum probes, proposes bounded changes, and records its rationale.
- A Python standard-library simulator calculates the official results and rejects unsupported or unsafe changes.
- A repository-scoped Codex Skill exposes the workflow to the agent.
- A standalone HTML report presents before/after heatmaps, metrics, adjustments, and the full decision trail.
- Synthetic data and a no-key smoke path let judges reproduce the project without hardware, accounts, vendor SDKs, or proprietary files.
Codex and GPT-5.6
Codex was the primary environment for clean-room design, implementation, testing, browser review, and submission preparation. GPT-5.6 is also part of the live product experience: it diagnoses evidence and selects the next action, while deterministic code preserves the audit and safety boundary.
An independent GPT-5.6 Codex run used both diagnostic probes and reached all six acceptance targets in two adjustment cycles.
Try it
Python 3.10+ is the only requirement:
python tools/optical_lab.py demo --output-dir run/demo
start run/demo/report.html
The repository includes setup instructions, tests, synthetic sample data, and the guided Codex workflow. It is MIT licensed.
Log in or sign up for Devpost to join the conversation.