Projects

MORPHEUS checks every step an AI agent plans or takes against the requirements you gave it, and tells you, for each requirement, whether it is met, broken, or cannot be checked — with the evidence.

• Check. Your requirements are kept across the whole conversation, and a verdict rests only on what code can confirm — the plan's own “total: within budget” never counts. When it cannot tell, it says so instead of guessing. On a blind test it let none of 8 planted budget overruns through (2 flagged as broken, 6 as cannot-check); the LLM judges let 2 to 8 of them through.
• Stop. Inside Claude Code, Codex and Cursor it blocks the tool call that would break a rule you set — a new dependency, a force push, an edit to the file that computes your metric — and checks the agent's “done” before you believe it. Replayed on 4,750 real tool calls: 10 blocks, all correct.
• Fix. When a step breaks a requirement, it proposes alternatives at that step and re-checks them before you see them.
• Record. For ML experiment loops, it turns a run into a record anyone can re-verify: a frozen evaluator, the noise across seeds, and an audit for metric gaming.

Where it came from: a planning engine for science (Morphism-Oriented Reasoning with Provable, Honest, Epistemic Understanding for Science) that builds a knowledge graph from papers — 18,600+ actions from 665 documents, every one traced to its page — plans over it, and accepts a plan only after a verifier passes it. Research: EMNLP Findings 2025 and four 2026 preprints. Write-ups: what MORPHEUS is and does · the judge that agrees most is not the safest · related work: judges, verifiers and guardrails.

One OpenAI-compatible model name. Behind it, Rung tries the cheapest capable model, checks the answer, and escalates only on a verified failure, under a hard budget per task. Every response says who answered, what each attempt cost, and whether anything checked it.

• No prediction. Guessing the right model from the prompt recovered none of the available savings in our tests; climbing on evidence did. Ahead of every single model in 37 of 39 matched-spend comparisons on HumanEval+.
• A check that needs nothing from you. If you have tests, Rung uses them. If not, it asks the same model three times and requires it to agree with itself. 0.78–0.98 correlation with correctness, at about 3 % of one top-model call; an LLM grader came out worse than chance.
• Routing by kind of task. Your own traffic builds a table of cost per correct answer for each kind of task, so the expensive model is used where it actually wins. Its advantage was 58 points on one category and 2 on another.
• Drop-in. Point your existing OpenAI client at it; nothing else changes.

Status: working and measured, not yet battle-tested. Write-up: Don't predict, verify · related work: routers and cascades.

Drop in your agent's logs. A minute later you know whether training a smaller model would beat simply letting the agent try again at the same cost, whether you have enough of the right data, and what to record next.

• Train, or try again? Two cards side by side: what retrying already gets you, measured on your own log, and what training could get you, with the published results it is based on.
• Do you have enough? Which tasks have no example of success, which states are never reached, which tools are never used — as a collection plan with sizes.
• What to record. The ten things a log needs before it can answer, and the ones yours is missing.
• Honest by construction. Every sentence points to a reading of your log or a cited paper, and it says “can't tell” when the log cannot decide. It never promises that training will help, and no training platform pays us, so the advice does not depend on which way you go.

Status: runs on your own computer, or on a hosted site for invited teams where your log is deleted as soon as the report is made. Comparing a trained agent with the untrained one at the same cost is next. Write-up: Before you train on your agent's logs · related work: from agent logs to training.