Understanding the Output#

After a run, ./results/<session_id>/ contains:

File	Description
`paper.pdf`	Compiled PDF (requires pdflatex + bibtex)
`paper.tex`	Full LaTeX source — edit and recompile if needed
`references.bib`	BibTeX bibliography
`theory_state.json`	Full proof state — lemmas, proofs, confidence scores
`research_brief.json`	Planning state — directions scored and selected
`experiment_result.json`	Numerical validation results (if experiment ran)

Paused sessions also write a checkpoint to ~/.eurekaclaw/sessions/<session_id>/checkpoint.json.

Reading `theory_state.json`#

Key fields:

{
  "status": "proved",
  "proof_plan": [
    {
      "lemma_id": "concentration_bound",
      "provenance": "known",
      "statement": "For sub-Gaussian ..."
    }
  ],
  "proven_lemmas": {
    "main_result": {
      "verified": true,
      "confidence_score": 0.91,
      "verification_method": "llm_check",
      "proof_text": "..."
    }
  },
  "failed_attempts": [ ... ],
  "counterexamples": [ ... ]
}

Proof Status#

Status	Meaning
`proved`	All lemmas verified, assembled proof complete
`refuted`	A counterexample was found; the conjecture is false or needs refinement
`abandoned`	Hit `THEORY_MAX_ITERATIONS` without completing; partial proof saved

Lemma Provenance#

Provenance	Meaning
`known`	Directly citable — no new proof needed
`adapted`	A known result modified to fit this context
`new`	Genuinely novel — fully proved from scratch

Low-Confidence Warnings#

If a lemma has verified=false, the PDF contains:

[Unverified step]   ← orange text

and a Limitations section explaining all unverified steps. Review theory_state.json → proven_lemmas to see which lemmas are flagged.

Using the Python API to Access Results#

import json

state = json.loads(result.theory_state_json)
print("Status:", state["status"])
print("Lemmas:", len(state["proof_plan"]))

brief = json.loads(result.research_brief_json)
direction = brief["selected_direction"]
print("Direction:", direction["title"])
print("Novelty score:", direction["novelty_score"])

See Python API for full details.