Skip to content

Latest commit

 

History

History
103 lines (81 loc) · 6.23 KB

labels.md

File metadata and controls

103 lines (81 loc) · 6.23 KB

What a label says — and what it does not

🌐 English · Русский · Español · Français · 中文

Every label HISTOR issues is an MTL/1 label: a standalone AWR/2 VerificationVerdict, signed by the log's did:key, verifiable offline with any AWR/2 verifier. HISTOR is, as far as we know, the first issuer of that profile against real servers.

The subject

Number-masked and whitespace-only review views do not replace the literal canonical MTL digest. Numeric edits are not automatically cosmetic: a per-call price can change without a schema change. Separate description, schema and membership counts, their overlapping flags and missing evidence: measurement limits.

A label is about the pair (server identity, advertised tool set) — not the code, not the package, not the publisher. The subject is an MCP Server Descriptor:

{
  "mtl": "1",
  "server": { "name": "io.example/weather", "registry": "urn:awr:mtl:1:registry:registry.modelcontextprotocol.io" },
  "toolSet": { "count": 2, "names": ["get_weather", "list_cities"], "digestSRI": "sha256-…" },
  "artifact": { "transport": "streamable-http", "endpoint": "https://example.com/mcp" }
}

The descriptor carries no time and, in HISTOR, no version: an unchanged server produces the same subject digest every day, so detecting a change is comparing two digests. (A registry version bump with identical tools must not read as "definitions changed".)

flowchart LR
    T["tools/list result<br/>(every page)"] --> N["normalise each tool<br/>name · description ·<br/>inputSchema · outputSchema?"]
    N --> O["sort by UTF-16 code units<br/>(not locale, not code points)"]
    O --> J["RFC 8785 canonical JSON"]
    J --> H["SHA-256 → tool-set digest"]
    H --> D["descriptor"]
    D --> J2["RFC 8785"] --> H2["SHA-256 → subject digest<br/>= verifiedWork.digestSRI"]
Loading

Numbers are read as IEEE 754 doubles and written the way RFC 8785 says, which is how JSON.stringify writes them: 0.5, 1e+21, 5 for 5.0. Until MTL/1.1 (31 October 2026) a set with a fraction anywhere in its schemas got no digest (MTL-NUM-001), and so did an empty set and one that repeats a name; on that day 5,159 servers HISTOR had read were in that state, 4,941 of them for a fraction. Now only an entry with no name (MTL-SUBJ-001) or a number with no double (1e400, MTL-NUM-001) leaves a set without a digest, and a set with integers only gets the same bytes, and digest, as before. A set that gains its first digest is not counted as a change.

The four methods

Method Issued pass fail inconclusive
tool-set-observation first time a target shows a tool set tool set drained and digested never an entry with no name; a number with no double
tool-def-pattern-scan with every new observation, and again when WARDEN's pattern set changes no pattern matched never one or more matches, each listed with its tier
tool-set-continuity each crawl (≥ 20 h apart) once a prior label exists same subject digest as the prior label the digests differ no digest on one side; the current observation failed
name-threat-match with every new observation name and endpoint on no record never a record matched (a naming signal, not evidence about code)

fail exists only for continuity, because it is a mechanical statement about two committed digests. A pattern match cannot establish that a definition is wrong — the shipped pattern set flags a tool that legitimately takes an api_key — so the honest non-pass outcome is inconclusive, and every match carries its WARDEN tier: block rules can refuse a connection in WARDEN, advise rules never do. Labels carry no score: the only number available would be a product of gate constants, and publishing it as a 0–1 safety score would be the most misleading thing a label could do.

How the desk renders them

MTL/1 §9.2 binds any registry that displays labels, and HISTOR's own desk and badges follow it.

Outcome Rendered as Never as
observation pass "Tool definitions pinned ‹date›" Verified, Audited
continuity pass "unchanged since ‹date›" Stable and secure
continuity fail "tool definitions changed ‹date›", amber, with the diff Compromised, rug pull detected
scan inconclusive "N patterns matched — read the definition", neutral Failed, Dangerous, red
no label "not observed" + the status Failed, or ranked below an inconclusive

Verifying a label yourself

curl -sS https://histor.modelmarket.dev/api/v1/labels/<label-id> -o label.json
pip install awr                                 # the AWR/2 reference implementation
python -m awr verify label.json                 # "valid": true, "profile": null (a verdict is not a receipt)

# Recompute the subject digest from the evidence, as MTL/1 §4.6 asks a registry to:
curl -sS https://histor.modelmarket.dev/api/v1/descriptors/<subject-digest> -o msd.json
python -m awr digest msd.json                    # must equal credentialSubject.verifiedWork.digestSRI

The evidence a label cites is served by digest and never changes: /api/v1/toolsets/<sri>, /api/v1/descriptors/<sri>, /api/v1/pattern-sets/<sri>, /api/v1/record-sets/<sri>. A consumer should cache them by digest when it receives a label, so the label keeps its meaning even if HISTOR disappears.

Known limits, stated inside every label

Each label's scope sentence is inside the signature, so no page template can soften it:

  • advertised text only — no source read, no package resolved, no tool invoked;
  • a server can serve identical definitions and change behaviour;
  • a server can show different clients different definitions; a single crawler cannot see that, which is why /check exists and why a second, separately operated observer is the next step;
  • observation time is HISTOR's assertion; the log makes it impossible to change afterwards, not impossible to have lied about at the time;
  • endpoints behind authentication answer 401 to HISTOR and are recorded as not observed.