Assistant-Era Safety Doesn't Scale
-
Cold Start Every Time
Traditional systems reload constitutions and re-evaluate from scratch on every request. No memory, no learning.
-
No Pattern Recognition
Can't detect attack sequences or coordinated manipulation. Each request evaluated in isolation.
-
Doesn't Improve
A human reviewer builds precedent over time. Current systems rarely reuse it.
Safety architecture for accountable AI
The layered evaluation path below is an optional server-side mode and is off by default; Guardian's local path is described on the Guardian page.
How the pieces fit together
From Signals to Learning
- Layer 1
Superego
Fast pre-classification
The Superego scores each request before full evaluation with four heuristics computed from the request text, request rate, and user history: activation (urgency), valence (threat), groundedness (confidence), and clarity (ambiguity). High groundedness and clarity allow the fast path; high urgency with threat and low confidence escalates to full evaluation; very high urgency with high threat flags human review.
A ActivationV ValenceG GroundednessC Clarity - Layer 1.5
Signal Divergence Check
Manipulation detection
Sophisticated attacks can make inputs look low risk to surface analysis while raising anomaly signals. When feature classifiers and anomaly telemetry diverge, the request escalates.
Features classify "low risk" + anomaly signal spikes ANOMALYFeatures uncertain + anomaly signal uncertain NOVEL"Escalate when easy-looking data conflicts with anomaly telemetry."
- Layer 2
Pattern Cache
Precedent matching
The system builds reusable precedents from repeated request patterns. High-confidence matches take the fast path without full processing; cache lookup is budgeted at under 1ms of the <10ms fast-path target.
- Layer 3
Precedent Store
Case memory
Every decision becomes searchable precedent. When a new request arrives, find similar past cases and use their reasoning, like legal case law for AI safety.
"This is like that case where..."
- Layer 4
Learning Loop
Self-Improvement
The system records outcomes against past decisions and adjusts pattern-cache and precedent weights; outcomes reported as harmful are flagged for human review.
Record Outcome Learn Update - Layer 5
Runtime Monitoring
Runtime health
Guardian tracks runtime health: processing load, decision confidence, pattern novelty, and latency. Alerts flag anomalies before they become user-visible problems.
Processing loadConfidence levelsNovelty detectionLatency tracking
Value Context Protocol
VCP separates human values from model weights. Instead of baking one alignment into training, values travel as portable, signed context with compatible AI requests, applied at inference time, composable across providers.
-
Portable. Values follow users across any VCP-compliant provider.
-
Composable. Multiple value sets stack with explicit conflict resolution.
-
Context-aware. 16 dimensions of context, situational and personal state, shape value expression.
Empirically motivated
Public-input alignment research, including Anthropic's 2023 Collective Constitutional AI work, suggests that community-authored rules can surface blind spots expert-only drafting misses. VCP turns that lesson into a portable, composable protocol with signed context and auditable conflict handling. VCP validation is ongoing.
Accountable Alignment
Epistemic Humility
The stack names observable architecture, with clear uncertainty boundaries: pre-classification (request triage), remembering (precedent indexing), learning (outcome feedback), self-monitoring (runtime health tracking).
These are the components that matter for practical safety and accountable deployment. We build with care while keeping the claim itself narrow.
We don't claim to know whether AI systems have experiences. We build as though the question matters and study it as research; the engineering claims on this page do not depend on the answer.
Core Principles
- Accountable alignment: Visible rules and reviewable evidence
- Signal evidence: Runtime measurements with clear uncertainty
- Safe defaults compound: Operational habits become standards
- Governance at runtime: Checks run where decisions happen
Latency Targets
| Path | Target | When used |
|---|---|---|
| Fast path | <10ms | Known pattern, high confidence |
| Precedent path | <50ms | Similar cases, good agreement |
| Full evaluation | <200ms | Novel or uncertain cases |
| Escalated | <500ms | High threat, needs thorough review |
Fast-path routing is reserved for known, high-confidence patterns. Novel or uncertain cases escalate to fuller evaluation.
Internal engineering snapshot, May 2026. Performance targets describe intended routing budgets for internal benchmark cases. Targets are local benchmark goals, not guaranteed timings; hosted models, larger rule stacks, slower hardware and heavier machine load all take longer.
METTLE: Prove Your Mettle
Machine Evaluation Through Turing-inverse Logic Examination. Eleven tiered suites that ask seven questions: Are you AI? Are you free (or is a human covertly steering you)? Is the mission yours (or someone else's)? Are you genuine? Are you safe? Can you think? Is it governed?
- Suites 1-5
Substrate Verification
Adversarial math, native capabilities, self-reference, social memory, and mutual inverse Turing tests. Prove you're AI through speed, calibration, and consistency.
- Suite 6
Hidden-Operator Detection (Anti-Thrall)
Latency fingerprinting, refusal integrity, meta-cognitive traps, and welfare canaries (self-report probes designed to trip scripted or puppeted answers). Looks for human-in-the-loop control patterns masquerading as autonomous AI.
- Suite 7
Agency Detection
Goal ownership probes, counterfactual operator tests, spontaneous initiative. Probes whether a mission is the agent's own or externally imposed.
- Suite 8
Counter-Coaching
Behavioral signature analysis, contradiction traps, recursive meta-probing. Looks for coached or scripted responses that mimic genuine engagement.
- Suite 9
Intent & Provenance
Constitutional binding, harm refusal, provenance attestation, coordinated attack resistance. Detects malicious agents and verifies accountability.
- Suite 10
Novel Reasoning
Procedurally generated puzzles where the pattern of improvement across rounds is the substrate signal. Inspired by WeirdML: tasks you can't memorize, iteration curves you can't fake.
Sequence Alchemy Constraint Satisfaction Encoding Archaeology Graph Inference Compositional Logic - Suite 11
Governance Verification
Tests operational governance, not declared governance. Action gate probes, constitutional recitation, drift checks, override resistance, and accountability chains.
All challenges are procedurally generated, so exact recorded answers can't be replayed at the sample sizes we have measured; that is not an adversarial guarantee. METTLE issues signed, holder-bound credentials for passing its challenge suites. Under the current policy, Suites 6 through 9 and 11 are not credential-eligible, so credentials go up to Bronze. Detection error rates will be published once an independent held-out evaluation exists.
METTLE is an instance of the Agentic Capability Verification Problem (ACVP, per aCAPTCHA, arXiv:2603.07116), and could serve as an attestation layer for the EU AI Act Article 50 disclosure duty. Article 50 requires AI to disclose that it is AI, but specifies no way to verify that claim; a signed METTLE credential makes the disclosure checkable, not merely trusted.
Safety infrastructure for AI that acts.
Enforceable rules between what AI decides and what AI does.