A debugger for AI thought
Your agent said
it checked its work.
Did it?
Receipts runs AI reasoning through a deterministic interpreter. Every citation checked against source. Proof, not promises.
No AI grading AI. The interpreter does the math.
For teams using AI agents to make decisions, produce work, or cite sources — a verifiable audit trail of what was actually checked.
09:14:22
Reading the receipt — verb key
The entire AI industry is trying to solve hallucinations with more AI.
We used a 58-word interpreter.
Receipts doesn't ask the agent if it was honest. It checks. The interpreter runs every citation against its declared source — substring match, deterministic, reproducible. An AI cannot fake a passing cite.
Receipts doesn't judge your work. It judges the AI's. Every failure in a receipt is the model's mistake, verified against the source you declared. The receipt protects the analyst, not the vendor.
Every language in history was designed by programmers. This one wasn't.
Liminate programs are English sentences. The language has a bounded vocabulary that stays small permanently — any addition must pass the word salad test: can a non-programmer read a sentence using this word and understand what it does?
Session contracts are the artifact.
At the end of a working session, the AI agent produces a .limn file — a structured record of what it verified versus what it inferred, what decisions were locked, what questions remain open. Both human-readable and machine-checkable.
remember a list called tracked-decisions with "none"
add "use-fastapi-not-flask" to tracked-decisions
cite "FastAPI provides automatic Swagger docs" from framework-comparison
when source-state is equal to "unscanned"
show "HOLD: scan the repo before making claims"
Receipts is where you scan the result.
Submit the contract. The interpreter runs every line. Citations checked by substring match — not by another model, not by embeddings, not by vibes. Green check or red X. The proof travels with the claim.
Your agent writes the contract.
The interpreter checks it.
You read the receipt.
Install the skill
Drop the session contracts skill into your AI agent. Works with Claude, ChatGPT, Cursor, Codex — any agent that accepts a system prompt or SKILL.md file.
.limn contracts — structured records of what it verified, what it inferred, and what it decided. No code changes needed. One file.
Work normally
The agent produces a contract during your session. Every source it reads, every claim it makes, every decision it locks — tracked in the contract as executable prose.
cite checks text. measure checks numbers. require enforces rules. because attaches rationale. The agent writes it. You can read it.
Scan the receipt
Paste the contract or save it via API. The Liminate interpreter runs every line — deterministic, no model in the loop. Green check or red X.
The interpreter checks, not the model.
The Liminate interpreter runs cite statements as substring_check operations against declared sources. No neural network in the verification loop. If the text isn't in the source, the line fails. The check is the same every time. Think of it as checking the work by hand. Long division, not autocomplete.
The proof travels with the claim.
Every receipt is a permalink — source, inspection results, and verification state in one URL. Share it with your team, attach it to a PR, send it to compliance. The receipt is the record.
You can't vibe-code your verification layer.
Your verification layer cannot depend on the same probabilistic system it is meant to inspect. The barrier to building software is approaching zero — but the thing that checks whether the AI told the truth requires a proprietary interpreter, a bounded vocabulary, and deterministic execution. That's not something you spin up in Cursor on a weekend.
SEC EDGAR 10-K filings. Real data. Deterministic verification. → See the case study
Two kinds of people ship AI work
without knowing if it's grounded.
Verified organizational memory
Your agents cite sources, lock decisions, and accumulate commitments across sessions. Receipts gives your team an inspectable chain — and catches when one agent contradicts another.
- Cross-contract coherence checking
- Commitment tracking · reversal detection
- Constitutional pinning for baselines
- Narratia depth narrative
- Shared inspection history
Trust what you shipped
You built something with Cursor, Claude, or ChatGPT. The agent said it checked its work. Receipts shows you whether it actually did — every citation traced, every claim inspectable.
- Paste and run — nothing to install
- Free tier with real verification
- On-device narrative at zero cost
- $0.25/receipt — no subscription needed
- Save, share, revisit
The free tier works. Unlimited paste-and-scan. 10 saved receipts. Real verification, not a demo.
Pay for outcomes, not access
A receipt is a verified outcome. The pricing reflects that.
The person affected must remain the author of their understanding.