Schema v1.0.0 · Open Standard

When agents choose your tools,
marketing copy is invisible.

CapCard is the open standard for machine-readable, verifiable capability claims. A single JSON file your project publishes so agents — and humans — know exactly what you've proven, measured, and reproduced.

The agent economy has inverted the discovery funnel. An AI agent researching tools doesn't read landing pages. It queries structured data, verifies claims, and reports back. If your project doesn't have a CapCard, the agent guesses — and guesses wrong.

Structured
Every claim has a status field: proven, empirical, or conjecture.
Reproducible
Every benchmark carries the exact command that produced it. No black boxes.
Verifiable
Schema-validated locally. Cryptographically signed in the hosted tier.

The gap agents fall into today

What agents find today
# README.md

## Features
- Blazing fast inference ⚡
- State-of-the-art results 🏆
- Production-ready
- Easy to use

# No benchmarks. No proofs.
# No reproducible commands.
# Agent hallucinates the rest.
What agents get with a CapCard
{
  "id": "approx.density",
  "name": "Universal Dense Approximation",
  "status": "proven",
  "reproducibility": {
    "command": "cargo run --release",
    "deterministic": true
  }
}

How it works

Three layers, each independently useful. Start at layer one — you're done in an afternoon.

1

Publish a capability_card.json

Drop a single JSON file at your repo root. Declare capabilities with honest status labels, benchmarks with reproducible commands, and proofs with verification dates. The schema enforces structure; you control the claims.

2

Verify locally with the open CLI

capcard verify validates your card against the schema, parses every reproducibility command, and returns a structured JSON report. Runs offline. No account needed. No data leaves your machine.

3

Get a cryptographically signed trust score coming soon

The hosted verification service re-runs your benchmarks in a clean-room container, issues a signed trust score, and attaches a CapCard Certified badge to your commit. Agents query the registry. Humans see the badge.

Flagship reference implementation

Monogate

View repo →

Monogate is an open-source mathematics library built around a single binary operator eml(x,y) = exp(x) − ln(y). It carries the most rigorous CapCard we know of — not because we wrote both, but because the results earned it.

Proven theorems
  • EML Weierstrass Theorem Linear combinations of EML trees are dense in C[a,b]. All tested functions DENSE — 0 separations.
  • Infinite Zeros Barrier No single EML tree can equal sin(x). Proven analytically and confirmed across 1,704,034,304 trees to N=12.
  • Bivariate EML Weierstrass Dense in C([a,b]²). x·y recovered at MSE = 2.63×10⁻²⁰.
  • KL = Bregman via EML KL divergence is the Bregman divergence of the EML log-partition. Direct connection to information geometry.
Open questions (with PSLQ evidence)
  • EL-Transcendence Conjecture Phantom attractor α ≈ 6.2144 is PSLQ-null at 200 digits against 20 classical constants. May be a new transcendence class.
  • N=3 Schanuel Phase Transition Universal 10⁴–10¹⁴× improvement at depth 3 across all tested functions. Mechanism: depth-3 EL independence (conditional on Schanuel's conjecture).
This is the point.

A CapCard doesn't require certainty — it requires honesty. proven means proven. conjecture means conjecture. Agents and humans can handle nuance. What breaks trust is false precision.

capcard verify D:/monogate/capability_card.json
{
  "valid":                true,
  "card_name":            "Monogate",
  "version":              "0.7.0",
  "capabilities_count":   17,
  "benchmarks_count":     22,
  "proofs_count":         11,
  "trust_score":          0.94,
  "errors":               []
}

Try it now

Verify any CapCard locally. The CLI is open-source, runs offline, and exits 0 on valid.

Install
$ pip install capcard-ai
Verify the monogate reference card
$ capcard verify capability_card.json
Or run without install
$ python -m capcard.cli verify capability_card.json

The schema

A minimal required surface — five top-level fields — plus rich optional structure for capabilities, benchmarks, proofs, and trust scores. Start minimal, add rigor as your claims earn it.

capability_card.json — minimum valid card
{
  "capcard_version": "1.0.0",
  "name":            "your-project",
  "version":         "1.0.0",
  "capabilities":    [{
    "id":      "my.feature",
    "name":    "My Feature",
    "status":  "empirically_confirmed"
  }],
  "benchmarks":      []
}

The moment the industry got serious about capability claims.

Add a capability_card.json to your project. It costs an afternoon. It pays back every time an agent — or a careful human — needs to know what you've actually built.

Get started Browse registry