Trust scores

A trust score is a number from 0 to 100 that summarises the security, reliability, and transparency posture of a Sentinel agent. Every score is produced by the same versioned, deterministic rubric, so scores are comparable across agents. The current rubric is v1.1.

How the score is computed

Sentinel uses a penalty model, not a weighted sum. Every agent starts at a perfect 1.0 and each verification signal can only deduct from that score:

score = 1.0
score -= (penalties from each stage)
score = clamp(score, 0.0, 1.0)         # never below 0 or above 1
display_score = round(score * 100)     # the 0–100 number you see

This makes the score transparent and monotonic: a clean agent scores 100, and every deduction maps to a concrete, disclosed finding. The internal score is a float in [0.0, 1.0]; the displayed score is round(score × 100).

Penalty weights (rubric v1.1)

Penalties are per-stage: a finding only ever triggers the penalty for the stage that produced it (a supply-chain CVE never triggers the static-analysis penalty). The overall finding counts shown in a report are the union of all stages, counted once.

Deferred stages are never penalised

Some stages depend on infrastructure that may be unavailable — for example, dynamic detonation and live red-team require the execution runtime, and AI review requires a model key. When a stage cannot run, it is marked deferred, not failed: it deducts nothing, and the report discloses it in public_summary.stages_deferred so a badge never claims coverage it did not actually achieve. Only a genuine failure of a stage that did run deducts points.

Certification tiers

The final score maps to a certification tier:

An uncertified agent is not eligible for the marketplace listing gate.

What the score means for buyers

• 90–100 (certified_managed): No critical/high findings and every stage that ran passed. Safe for production; review any low/info findings.
• 75–89 (certified): No critical findings; at most a few capped deductions. Suitable for production after a quick read of the open findings.
• 50–74 (provisional): One high finding or several capped medium deductions. Read the full report before deploying in a security-sensitive context.
• Below 50 (uncertified): Not recommended for production — significant unresolved findings, or a critical finding.

Because deferred stages are disclosed, always check stages_deferred on the report: a high score with dynamic/redteam deferred means those lanes have not yet run (they run once the execution runtime is available).

How to read a report

The score is served with two views:

• Public summary (GET /reports/{agent_id}/badge via the gateway): score, cert tier, critical/high counts, and stages_deferred.
• Full report (GET /reports/{agent_id}): per-stage results and finding counts.

Every report is stamped with its rubric_version and, when signing is enabled, an Ed25519 signature so consumers can verify the score was produced by Sentinel and not tampered with.

How sellers improve a score

Rubric versioning

The rubric version is recorded on every trust report (currently v1.1). A change to any weight or threshold ships as a new rubric version and is shadow-run against existing agents before it takes effect, so scores never move without a corresponding, published rubric change.

:::note Roadmap Continuous re-verification (score decay + periodic re-check), a public score-history API, trust_score.updated webhooks, and an externally-fetchable rubric endpoint are planned and not yet generally available. This page documents the rubric as it is implemented today; see the roadmap for what is coming. :::