Skip to content

Repository files navigation

VeriEnvelope

“It passed.” What exactly passed?

VeriEnvelope is an open evidence system for verifying specific Tool claims under declared conditions.

Verify specific claims, not products. A result without its envelope is incomplete.

Claim → Method → Evidence → Envelope

VeriEnvelope keeps a narrow, testable statement attached to the frozen rule that judged it, the observation that supports it, and the conditions where the conclusion stops. It does not produce product scores, rankings, certifications, or fitness-for-use verdicts.

Explore the public Viewer at verienvelope.pages.dev or inspect the source-of-truth records in this repository.

A real case: Playwright MCP

The current record for the claim that browser_navigate reaches a fixed local page is not_demonstrated under VE-METHOD-MCP-002 0.4.0.

This does not mean Playwright is broken. It means this specific claim was not demonstrated under the frozen Method and runtime envelope. The cause remains unclassified.

Why this exists

Software results are often easier to publish than the conditions that made those results true. PASS, FAIL, verified, and safe travel quickly; the component identity, decision rule, runtime, permissions, untested areas, and revalidation triggers often do not.

VeriEnvelope exists to bind those parts back together. It asks:

  1. What exact behavior was claimed?
  2. What was fixed before execution?
  3. What was actually observed?
  4. Which frozen rule authorized the result?
  5. Where does that conclusion stop?
  6. If the conclusion changes, does the history remain visible?

Sometimes the tool isn’t wrong. The test is.

The Filesystem MCP case preserves a real correction chain.

Its first read_text_file observation matched the preregistered fixture bytes. But an independent review found that Method 0.1.0 did not specify the formal capability status while the Runner supplied demonstrated. The raw observation remained valid; the formal status was not authorized by the frozen Method.

The old evidence package stayed unchanged. Method 0.2.0 made the decision rule explicit, the Runner was repaired, a new run was authorized, and the claim was measured again. Only then did the current demonstrated conclusion return.

Evidence stayed. The conclusion changed.

Methodology

VeriEnvelope separates the normative Method from its implementation and from any downstream registry decision.

Layer Location Responsibility
Method Specification methods/ Public rules that can be reproduced without this repository's Runner.
Reference Implementation runner/ One implementation of a Method; never the sole normative authority.
Verification Result One run's result.json The conclusion from one measurement. Third-party results do not become official automatically.

The evidence chain is:

Measurement Object
  → Method
  → Raw Observation
  → Interpretation / Decision Rule
  → Claim Result
  → Claim Envelope
  → Admission Rule
  → Registry Decision

Observation is not itself a verdict. Runner behavior is not Method authority. Admission is not certification.

Start with the core methodology, the responsibility model, and the glossary.

Architecture

  • methods/ — frozen, public Method specifications and version history
  • runner/ — one reference implementation of those Methods
  • evidence/ — sealed run packages and raw observations
  • schemas/ — machine-readable record contracts
  • registry/ — identities, authorizations, history, and registry state
  • site/ — read-only public Viewer generated from five current sealed results
  • viewer/ — local single-record renderer

GitHub is the source of truth. The website is a Viewer and entry point; it does not rerun tools or recalculate conclusions.

Current boundary

  • v0.1 covers Tool claims only. Skill, Agent, Multi-Agent, and Workflow verification remain outside the active scope; GATE 2 is locked.
  • The data model can represent tool and mcp_server objects.
  • The reference run_case path rejects third-party component verification. Unknown code is not installed on the host.
  • A result with admission=admitted is not publication, certification, security assurance, or production readiness.
  • The local and public Viewers render existing records without judging them again.

The governing constraints are in CONSTITUTION.md. Method lineage is recorded in METHODOLOGY_LINEAGE.md. Current status is in CURRENT_STATE.md.

Run locally

Requires Python 3.11+.

python3 -m venv .venv
.venv/bin/pip install -e ".[dev]"
.venv/bin/pytest

Run the measurement-pipeline self-audit and render a local record:

.venv/bin/python -m verienvelope run \
  --method methods/VE-METHOD-001 \
  --case methods/VE-METHOD-001/cases/echo_match.json \
  --out evidence/_scratch

.venv/bin/python -m verienvelope view \
  --result evidence/_scratch/fixture.method001/0.0.0/*/result.json \
  --out viewer/preview/index.html

evidence/_scratch/ and viewer/preview/ are gitignored. Exit code 0 means only that this observation matched; it does not admit a component.

Build the public site

site/ generates a static, read-only presentation from five current sealed results.

Public URL: https://verienvelope.pages.dev/

python3 site/build.py
python3 site/verify_public.py

Cloudflare Pages configuration is documented in site/README.md. The publication evidence and hygiene checks are recorded in docs/release/PUBLICATION_AUDIT_V0_1.md and docs/release/PUBLICATION_HYGIENE_V0_1.md.

Explicitly out of scope

User accounts, comments, rankings, payments, ads, self-service verdict submission, bulk scoring, Agent or Multi-Agent standards, enterprise certification, a universal global standard, and model-generated adjudication. VeriEnvelope does not claim to be a “world first” or an authority that certifies products.

Working-name notice

VeriEnvelope remains a working project name. The repository does not claim that a complete trademark or domain clearance has been performed. Formal release planning should still include appropriate USPTO, WIPO, EUIPO, CNIPA, GitHub, domain, company, and product-name checks.

About

Evidence-bound verification for AI tools — specific claims, reproducible evidence, explicit boundaries.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages