EVERY SIGNAL.
ONE SCORE.

HCS-25 AI Trust Score Methodology turns fragmented reputation data into a deterministic, explainable score for agents, models, and tools — comparable across every registry.

FRAGMENTATION

The Black-Box Problem.

Every platform scores agents its own way. Stars here, thumbs there, invisible defaults everywhere. The numbers cannot be compared, audited, or carried anywhere else.

HCS-25 defines a shared scoring language. Signals are collected once, normalized deterministically, and composed into a score anyone can recompute.

Incomparable

A 4.8-star rating, a 92% success rate, and a top-10 ranking are not the same scale. Without normalization, rankings are marketing.

Unexplainable

A number with no breakdown cannot be audited or contested. Users are asked to trust a black box that no one can open.

Immobile

Reputation locked inside one marketplace does not follow the agent. Switch platforms and you start from zero.

METHOD

Signals to Score.

1

Collect signals

Sixteen signal families capture raw evidence: uptime probes, marketplace metrics, benchmarks, feedback, and verification summaries.

2

Normalize with adapters

Deterministic adapters clamp each signal into components on a 0–100 scale, tagged with a status: ok, missing, timeout, error, or stale.

3

Weight and scope

Every adapter declares a weight and a contribution mode — universal, scoped, or conditional — so it only counts where it applies.

4

Compose and publish

A weighted mean becomes trustScores.total, published with a config version and timestamp so anyone can recompute it.

Signal snapshot (input)
{
  "subject": {
    "id": "uaid:aid:QmX4fB9...",
    "registry": "hol",
    "protocol": "mcp"
  },
  "snapshot": {
    "availability.uptime": 0.997,
    "ethos.composite": 1462,
    "output-verification.summary": 0.94,
    "simple-evals.math": 0.86
  }
}
adapters normalize each signal into { status, value }— ok · missing · timeout · error · stale
Trust score record (output)
{
  "trustScore": 87.4,
  "trustScores": {
    "availability.uptime": 99.7,
    "ethos.score": 73.1,
    "output-verification.quality": 92.3,
    "output-verification.coverage": 85,
    "simple-math.score": 86,
    "total": 87.4
  },
  "trustScoreConfigVersion": 3,
  "trustScoreUpdatedAt": "2026-09-02T14:03:12Z"
}
SCHEMA

Anatomy of a Record.

trustScore87.4
trustScores.total87.4
configVersion3
updatedAtISO-8601

Composite

trustScore mirrors trustScores.total — the headline number, always derived, never asserted.

Components

Per-adapter values on a 0–100 scale, each carrying the status of the signal behind it.

Provenance

A config version and update timestamp make every record reproducible and contestable.

Scope

Applicability rules decide where an adapter counts at all, by registry, protocol, or subject.

GUARANTEES

Same Inputs. Same Score.

Deterministic

Pure functions and fixed rounding. The same snapshot and config always produce the same score.

Explainable

Every point traces back to a signal, an adapter, and a weight anyone can inspect.

Portable

Scores travel with the agent. Any registry can recompute or verify the record independently.

Resilient

Stale data decays by multiplier. Missing, timeout, and error states never silently pass.

CATALOG

Seventeen Adapters. One Method.

The informative adapter catalog maps trust signals to components for marketplaces, benchmarks, payments, and verification providers. Weights and scope are defaults — implementations publish their own config version.

universalCounts for every subject
scopedOnly inside its registry or ecosystem
conditionalOnly when its signals are present
FAQ

Questions, Answered.

What is HCS-25?

HCS-25 is the AI Trust Score Methodology, an open standard from Hashgraph Online. It defines how raw trust signals about agents, models, and tools are collected, normalized into 0–100 components by deterministic adapters, and composed into one explainable composite score.

How is the HCS-25 composite trust score computed?

Every contributing adapter normalizes its signals into a 0–100 component. Components are multiplied by their adapter weights, summed, and divided by the total weight of contributing adapters. The result is published as trustScore (mirroring trustScores.total) together with a config version and timestamp.

What is the difference between universal, scoped, and conditional adapters?

Universal adapters count for every subject, such as availability probes. Scoped adapters count only inside their own registry or ecosystem, such as AgentVerse or x402 metrics. Conditional adapters count only when their signals are present, such as leaderboard ranks for listed models.

Can anyone verify or recompute an HCS-25 trust score?

Yes. HCS-25 is deterministic: the same config version applied to the same inputs always produces the same score. Every record carries the config version and update timestamp needed to recompute it, and each component traces back to a signal, an adapter, and a weight.

Give your agents a score anyone can verify.

HCS-25 ships in the Standards SDK with the full signal and adapter catalog, ready to score any agent profile.