Start with evidence
Signals remain tied to their source, subject, scope, and observation time.
Identity binding current
Authenticator assurance AAL3
Device integrity observed
Network threat low
Evidence-derived trust projection
A trust score explains the evidence behind a decision.
TrustAI converts validated, current evidence into a purpose-bound projection with confidence and reason codes. Resource policy decides access. Missing evidence, hard gates, and imported context remain visible.
Interpretation
Trust is evidence for a purpose.
A TrustAI projection answers a bounded question: given the named subject, session, resource, action, audience, policy, evidence, and time, how strongly does the evidence support this access context?
It is a purpose-bound evidence projection. Resource policy, entitlements, assurance, and local authority determine access. Two valid projections for the same subject can differ because the resource, mission, evidence freshness, and policy differ.
TB
The projection is bound to
Tenant and organization
Subject and session
Purpose and resource
Requested action and audience
Trust profile and policy version
Evidence and applicability digests
Observation and expiry time
Score construction
Evidence becomes a cautious, balanced projection.
Each dimension carries an observed value, a conservative lower bound, confidence, weight, source, and freshness. Atlas reduces the influence of uncertain evidence, balances the dimensions, and then applies policy gates.
Account for uncertainty
The supported value stays between the conservative floor and the observation. Lower confidence moves it closer to the floor.
54
76
88
lower boundsupportedobserved
Illustrative device signal · 65% confidence. Atlas qualifies the observed value of 88 to a supported value of 76.
Balance the profile
74projection
ARequired evidencePresent and current
BDimension floorsWeak areas remain visible
CAdverse gatesExplicit vetoes still apply
DResource policyReturns the final verdict
Missing evidence stays visible. A required dimension with no current evidence produces an indeterminate result and preserves the need for corroboration.
Thirteen dimensions
A profile of decision-relevant evidence.
Dimensions separate unlike signals so investigators can see whether the change came from authentication, device, behavior, threat, privilege, federation, resource alignment, or data exposure.
01Identity binding
02Authenticator assurance
03Device integrity
04Session integrity
05Behavioral consistency
06Network threat
07Compromise threat
08Automation abuse
09Privilege exposure
10Authorization governance
11Mission–resource alignment
12Federation integrity
13Data exposure
From projection to policy
Policy gates complete the decision.
A projection is evaluated with required dimensions, lower-bound floors, adverse conditions, freshness, and policy-specific thresholds. The same score may produce different outcomes for different resources.
ALLOW
All required evidence and policy conditions pass.
STEP_UP
Additional assurance or evidence can meet the current floor.
DENY
An explicit adverse veto or hard policy gate is present.
INDETERMINATE
Required evidence or an authority is unavailable, stale, or invalid.
Example: A user with valid AAL3 authentication can still require step-up or be denied when device integrity is missing, a compromise indicator is confirmed, the mission relationship is invalid, or the requested resource imposes a stricter profile. AAL describes authentication assurance; resource policy completes the access decision.
Enterprise continuity
A pseudonymous handle for governed subject continuity.
Atlas Trust Fingerprint v2 (ATF2) is a keyed, non-reversible, tenant-scoped lookup handle used to correlate the same governed subject across Atlas contexts.
ATF2 supports enterprise correlation under tenant, purpose, audience, epoch, and read-budget controls. Authentication and authorization remain with their dedicated controls.
ATF2 safeguards
- Keyed and non-reversible
- Tenant-scoped subject continuity
- Epoch and rotation support
- Proof and rebinding controls
- Purpose- and audience-limited lookup
- Audit and read-budget controls
Across boundaries
Admit foreign trust cautiously.
A foreign projection is validated for signature, issuer, profile, audience, purpose, subject binding, freshness, replay, and semantic comparability. The receiving tenant applies its own admission policy.
- Restriction-first semanticsValidated adverse signals can narrow a local decision when policy allows.
- Favorable evidence requires equivalencePositive influence requires exact profile, measurement, calibration, and applicability gates.
- Local authority winsThe receiving policy decides the outcome and records how foreign evidence was used.
Model governance
Production enforcement requires more than a model file.
Atlas supports live telemetry and reviewed feedback, but promotion is governed. Training data, labels, validation, calibration, signing, rollback, drift, and operational authority are separate concerns.
Collect attributable evidence
Ingest observations with source, subject binding, time, scope, retention, and validation state. Feedback enters governed review before promotion.
Train outside the enforcement path
Keep training, evaluation, and serving isolated. Record dataset and feature lineage; prevent operational labels from silently feeding the active model.
Validate independently
Evaluate discrimination, calibration, drift, robustness, protected operational scenarios, and the exact model bundle and profile intended for use.
Sign and promote deliberately
An authorized workflow activates a signed bundle with a defined scope, canary plan, revocation budget, status authority, and rollback target.
Observe and revoke
Monitor evidence availability, score distribution, explanation integrity, validator status, and outcome quality. Fail closed or advisory when authority is lost.
Shipped AIO posture: the integrated demonstration model is advisory. Production enforcement requires an independently evaluated and calibrated signed bundle, current validator status, exact applicability, authenticated scoring service, fresh evidence, and policy-defined corroboration and revocation controls.