The AgentNode Trust Model: Tiers, Scores, and What They Mean

Every package on AgentNode has a verification tier. This tier tells you — as a user or as an AI agent — how much evidence exists that this package works correctly and safely.

This article explains the full trust model: what each tier means, how packages move between tiers, and how the system uses tiers to make decisions.

The Four Tiers

What Gold Actually Proves

A Gold package has passed all of the following:

• Installation — pip install succeeds in a clean container
• Import — the tool entrypoint can be imported without errors
• Smoke test — the tool was called with publisher-declared inputs and returned a valid result
• Contract — the return value is JSON-serializable, non-None, and type-stable
• Reliability — 3 identical calls all succeed (≥ 90% pass rate)
• Determinism — outputs are consistent across runs
• Explicit cases — the publisher declared specific test inputs (not auto-generated)

All of this runs in a sandboxed container with no network access (unless VCR cassettes replay recorded responses). The platform controls the environment — not the publisher.

What Verified Means

Verified is the most common tier. It means:

• The package installs correctly
• The tool entrypoint imports without errors
• Basic smoke tests passed (with auto-generated or limited inputs)

Verified does not mean the tool was tested with meaningful inputs. It means the tool exists, installs, and doesn't crash on import. That's significant — many packages in other registries can't even guarantee that — but it's not Gold.

What Partial Means

Partial means verification hit a legitimate limitation:

• Tool needs API credentials the sandbox doesn't have
• Tool needs system dependencies (ffmpeg, tesseract) not in the base container
• Tool needs specific binary input files

The package isn't broken — it just can't be fully verified in the current sandbox. Publishers can upgrade from Partial to Gold by adding verification cases with VCR cassettes or appropriate test fixtures.

Tier vs. Score

The score (0-100) and the tier are related but not the same thing:

Score ≥ 90  →  Base tier: Gold
Score 70-89 →  Base tier: Verified
Score 50-69 →  Base tier: Partial
Score < 50  →  Base tier: Unverified

BUT: Hard caps can override downward.
A score of 95 with no explicit cases → Verified (not Gold).

The score measures how well verification went. The tier represents the confidence level the platform assigns. Hard caps ensure that certain minimum quality standards are met regardless of numeric score.

Publisher Trust Levels

Separate from package tiers, each publisher has a trust level:

Publisher trust and package tier are independent. A trusted publisher can have a Partial package (if it needs credentials). An unverified publisher can have a Gold package (if they added proper test cases).

How Tiers Affect the Resolver

When an AI agent requests a capability (e.g., "I need web search"), the resolver selects from matching packages. Tier factors into the resolution score:

• Gold packages are preferred over Verified
• Verified packages are preferred over Partial
• Unverified packages have the lowest priority

This means Gold packages are more likely to be installed automatically by agents. It's the primary incentive for publishers to add verification cases.

Quarantine

Quarantine is separate from tiers. A package can be quarantined for:

• Verification failure on first publish — auto-quarantined, cleared when re-verification passes
• Security scan findings — quarantined until findings are resolved
• Admin action — manual quarantine for policy violations

Quarantined packages are not visible in search and cannot be installed. Quarantine is independent of tier — a Gold package can be quarantined (e.g., if a security issue is discovered post-verification).

Tier Transitions

Tiers can move in both directions:

Upgrades

• Publisher adds verification.cases → Verified can become Gold
• Publisher fixes a failing test → Partial can become Verified
• Admin triggers re-verification after infrastructure update → Partial can become Gold

Downgrades

• Re-verification fails (dependency broke, API changed) → Gold can drop to Verified
• Security finding discovered → tier can be overridden by admin
• New verification runner catches previously-missed issues

Tier changes only happen during verification runs. Between runs, the tier is stable.

Confidence Levels

Each verification result also carries a confidence level (high/medium/low) that indicates how much you should trust the score:

• High — smoke passed, contract valid, reliability ≥ 0.9, publisher tests passed
• Medium — most checks passed but some data is missing
• Low — smoke was inconclusive, limited evidence

Confidence is shown alongside the tier on the package detail page. A Gold package always has high confidence (it's required for Gold). A Verified package might have medium confidence if it reached that tier through score alone.

The Verification Modes

Each verification run is tagged with how it was conducted:

The mode is visible in the API response and helps publishers understand how their package was tested.

Summary

The trust model has three layers:

1. Publisher trust — who published this? (unverified → curated)
2. Package tier — how well was it verified? (unverified → gold)
3. Quarantine status — is it safe to show/install? (none/quarantined/cleared)

All three are independent. Together they give the resolver and the user a complete picture of how much to trust a package.

For publishers: the fastest path to Gold is adding verification.cases to your manifest. For users: Gold means the platform tested it with real inputs in a sandbox — not just that the publisher says it works.