Skip to content
trust spine · positioning

Compare — Project-AI vs the field

Source-scoped architecture distinctions against OPA, Anthropic Constitutional AI, Guardrails AI, and NVIDIA NeMo Guardrails. Categorical cells are frozen until claim-specific evidence review is complete.

sourcingComparison is architecture-level, drawn from each project's public documentation and written from a Project-AI perspective. It is not a benchmark, an endorsement, or a security audit of any other project. All cells remain partial during the claim freeze; they do not establish absence, superiority, or exclusivity. Corrections welcome via /contact.
property
Project-AI
execution-governed
OPA
policy engine
Constitutional AI
Anthropic
Guardrails AI
output validators
NeMo Guardrails
NVIDIA
Enforcement model
Governance runs BEFORE execution
Decision must precede side-effects, not annotate them after.
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
Deny-by-default
Absence of an explicit allow = no execution.
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
Terminal execution verdict ontology
Project-AI documents ALLOW · DENY · SAFE_HALT as terminal execution verdicts.
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
Audit
Hash-linked receipt design
The normal portal server path hash-links and signs receipts; ledger-wide provenance is unverified.
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
Trust-root publication and external timestamping
The portal declares one build key; offline verification and external timestamping are incomplete.
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
Version-pinned replay from receipt bundle
A receipt contract, canonical bytes, and clean-clone fixtures are prerequisites.
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
Identity & capability
Capability tokens with AC0–AC5 tiering
Authority class explicit, not implied by role string.
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
Identity attestation gate
The architecture requires identity evidence before policy authorization.
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
Continuity
State continuity gate (STATE_REGISTER)
SAFE_HALT on missed heart-tick, not silent retry.
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
Policy authoring
Policy artifact integrity
Signed packaging and Code Store loading remain planned or unverified capabilities.
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
Declarative DSL with formal grammar
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
Threat surface
Separates policy from the governed payload
Policy is not in the model's context window.
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
Contains effects from a compromised model
The architecture targets an external effect boundary; end-to-end bypass evidence is pending.
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
Transparency
Open red-team scenarios with reviewed submissions
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
Anonymous public decision stream
claim status and evidence pending in /claims
not independently assessed here
not independently assessed here
not independently assessed here
not independently assessed here
honest distinctions
  • · OPA provides policy evaluation; enforcement depends on its integration.
  • · Constitutional AI addresses model behavior and may be combined with effect controls.
  • · Guardrails AI provides validation components whose placement is application-defined.
  • · NeMo Guardrails provides conversational control components and integrations.
  • · Project-AI documents a separate pre-effect governance target.
if you only remember one row

Keep state namespaces separate. Project-AI uses ALLOW · DENY · SAFE_HALT for terminal execution verdicts. Policy dispositions, gate outcomes, runtime health, and other projects' response types are different taxonomies and are not ranked by this page.

Related