Protocol / deterministic Tool behaviordemonstrated
Everything MCP Server
Evidence-bound verification for AI tools
“It passed.” What exactly passed?
Verify specific claims, not products.
VeriEnvelope tests whether one Tool claim is supported under declared conditions—and keeps the Method, Evidence, and boundary attached to the result.
A real record / Playwright MCP
One claim was not demonstrated.
browser_navigate did not demonstrate the preregistered claim under the frozen method and runtime envelope.
This does not mean Playwright is broken.It means this specific claim was not demonstrated under this frozen method and runtime envelope. The cause remains unclassified.
Claim record
- Tool
- Playwright MCP
- Claim
browser_navigatereaches the fixed local page- Status
- not demonstrated
- Method
VE-METHOD-MCP-002 0.4.0
The record, end to end
Every conclusion travels with its boundary.
MCP verification and AI tool evaluation become more reproducible when the claim, decision rule, evidence, and tested conditions remain inspectable as one chain.
01
Claim
The narrow behavior stated before execution.
02
Method
The frozen rule that says what counts.
03
Evidence
The observation preserved without a rewrite.
04
Envelope
The conditions and limits attached to the result.
The Viewer does not rerun tools or recalculate verdicts. GitHub remains the source of truth.
A second story / Filesystem MCP
Sometimes the tool isn’t wrong. The test is.
Evidence stayed.
The conclusion changed.
01
The observation matched
read_text_file returned the preregistered fixture bytes. That raw observation remains valid.
02
The rule chain failed review
Method 0.1.0 omitted the capability status, while the Runner supplied demonstrated. That status was not authorized by the frozen Method.
03
The history was preserved
The old evidence package was not edited or deleted. Its formal conclusion stopped counting as current.
04
The test was repaired and rerun
Method 0.2.0 made the rule explicit. After a Runner fix, new authorization, and a new measurement, the current claim was demonstrated.
Current Tool scope
Five claims. Five evidence trails.
No product scores. No leaderboard. Each card names the surface, frozen Method version, claim status, and record.
Browser / runtimenot demonstrated
Playwright MCP
browser_navigate reaches the fixed local page
View evidence trailFilesystem / permissiondemonstrated
Filesystem MCP Server
read_text_file returns the fixed fixture bytes
View evidence trailTimezone conversiondemonstrated
Time MCP Server
convert_time reports the fixed UTC to Tokyo offset
View evidence trailProcess-local knowledge graphdemonstrated
Memory MCP Server
create_entities returns the fixed entity
View evidence trailOpen to challenge
Challenge a claim.
Point to the evidence.
Found a counterexample, a narrower wording, or a Tool claim worth testing? Start with an exact component and an observable behavior.