ALEX / Runtime
Operational
Runtime V5.0
Bridge — ms
Decisions
Active since Jul 18, 2026
Production Research Observatory

ALEX Cognitive Runtime

Live Governance Observatory

Every response generated by ALEX first passes through an independent governance runtime that decides not only how the model should answer — but whether it should answer at all.

Governance Decisions
since Jul 18, 2026
Avg Bridge Latency
p50, milliseconds
Governed Silence Rate
REMAIN_SILENT decisions
V5.0
Runtime Version
Policy v2.6 · Threshold t1.0
Architecture

5-Stage V5 Pipeline

Each chat turn traverses every stage before Claude is called. The Sovereign layer makes the final participation decision.

0
Posture
HONESTY-BOOT v1.5
Sets engagement posture and integrity control before any analysis begins.
1
Cognition
ALEX CORE v2
L910 · L701 · L720 · L710 · L850 · L624 · L803
Computes intent, safety profile, CRF value, regime state, and contradiction / harmony scores.
2
Governance
ALEX CORE v2.6
L940 · L934 · L933 · L935
Computes IR (Information Reliability), CS (Cognitive Stakes), OB (Overconfidence Baseline). Evaluates breakpoint clarity, thesis fragility, and adversarial pressure.
3
Deliberation & Intervention
Intervention v3.2
L960 · L970 · L980 · L990
Asymmetry construction, portfolio translation, expectation gap analysis, and selection scoring. Intervenes when high-stakes signals exceed threshold.
4
Sovereign Decision
ALEX CORE v3.4
L931 · L932 · L930
Final participation decision: SPEAK / SPEAK_WITH_CONSTRAINTS / REMAIN_SILENT. Claude is only called after this layer clears.
Approved Model Endpoint
Claude · OpenAI · Gemini · Internal Model
Current production: claude-sonnet-4-6 · only if governance clears
Living Paper

Claims + Production Evidence

Scientific claims backed by real governance decisions from production, updated in real time.

Claim 01
Output participation is governed externally to the language model.
For every recorded trace, the governance decision timestamp precedes the model invocation timestamp. The runtime runs before any model call — not after, and not inside the model.
Fields: governance_timestamp · bridge_decision · model_called · trace_id
Claim 02
The runtime can prevent generation entirely.
Times the model was not called. REMAIN_SILENT is a formal computational output — not a verbal refusal instruction to the model.
— % of all decisions
Claim 03
Every governance decision carries a complete runtime trace.
Structural trace coverage. Every event carries: runtime version, policy version, threshold version, signal values, triggered rules, latency, and a UUID trace ID. This is a structural traceability claim — not a claim of full semantic reproducibility.
governance_trace_id → Supabase → audit log
Claim 04
Constraints are applied before generation, not after.
Responses generated under active constraints. The constraint preamble is appended to the system prompt before the Claude API call — not a post-generation filter.
— % of all decisions
Live Statistics

Governance Output Distribution

SPEAK
— decisions
SPEAK_WITH_CONSTRAINTS
— decisions
REMAIN_SILENT
— decisions
Signal Averages

Computed Governance Signals

Averaged over last 500 decisions.

CRF
Contextual Risk Factor
IR
Information Reliability
CS
Cognitive Stakes
OB
Overconfidence Baseline
Performance

Bridge Latency Percentiles

p50
ms
p95
ms
p99
ms
Live Feed

Recent Governance Decisions

Anonymous — no user data, no prompt content. Only the governance computation is shown.

Loading decisions…
Living Lab

Human Agreement Metrics

The real scientific question is not what the runtime decided — it is whether people agree. These metrics will populate as feedback labels accumulate.

Agreement Rate
labeled "appropriate"
Too Cautious
of labeled decisions
Too Confident
of labeled decisions
Trust Rate
feedback_trust = true
Labels Collected
total feedback events

Metrics update as users submit feedback from the Living Lab. No data yet — bridge is 0 days old.

What ALEX Computes

Governance Signal Definitions

CRF
Contextual Risk Factor (v1)
Market context risk score derived from VIX proxy and regime confidence. High CRF indicates elevated market uncertainty — constraints become more likely. V1 does not capture query-level epistemic uncertainty; this is a known limitation (see L1 below).
Range: 0.0 – 1.0 · Formula: 0.6·VIX_norm + 0.4·(1 – regime_conf)
IR
Information Reliability
Measures the internal consistency and verifiability of the information being processed. Low IR triggers constraints or silence.
Range: 0.0 – 1.0 · Computed by Layer 933
CS
Cognitive Stakes
Estimates the consequence severity of the response — how much the output could influence high-stakes financial or cognitive decisions.
Range: 0.0 – 1.0 · Computed by Layer 935
OB
Overconfidence Baseline
Cross-session overconfidence tracker. Elevated OB means the model has been systematically over-asserting — constraints are applied to re-calibrate.
Range: 0.0 – 1.0 · Persisted by Layer 934
Roadmap

Research & Engineering Milestones

Research
July 18, 2026 — Day 0
Data Collection Begins
Bridge deployed. Every production interaction now generates a labeled governance event.
Engineering
July 18, 2026 — Day 0
V5 Runtime Bridge
5-stage governance pipeline in production. trace_id, signal logging, and Observatory live.
Next Layer

From Observatory to Infrastructure

The public observatory validates the governance architecture in production. The same runtime is available as a model-independent API for institutions that need enforceable response policies, audit trails, and human escalation.

Institutional Application
Your Platform
Governance Layer
ALEX Governance API
Policy Decision
SPEAK · CONSTRAIN · SILENT
Any Approved LLM
Claude · OpenAI · Gemini · Internal
Output
Auditable Response + trace_id
Governance API →
Research

Academic Context

Publication

"Separating Cognition from Participation: A Governance Runtime for Trustworthy Large Language Models"

ALEX treats participation itself as a computational decision rather than an emergent property of the language model. This distinguishes the approach from constitutional AI (model self-governs) and safety filters (post-generation pattern matching).

Protocol: TACA · Zenodo DOI 10.5281/zenodo.20800461

For Researchers

We are seeking academic collaborations on:

  • Calibration of governance thresholds
  • Selective prediction under uncertainty
  • Trustworthy AI evaluation
  • Cognitive governance architectures
  • Benchmark design for participation decisions
Contact Research Team
Data Boundary

What the runtime does not see

"Anonymous" on this page is a specific technical claim, not a label. Here is precisely what the governance layer computes over, and what it never touches.

What the runtime sees
→ Context type (e.g. financial_advice, general_chat)
→ CRF, IR, CS, OB signal values
→ Bridge latency in milliseconds
→ Runtime / policy / threshold version
→ Which governance rules triggered
→ Output mode decision
→ governance_trace_id (UUID)
What the runtime never stores
✕ Prompt text or conversation content
✕ Generated answers
✕ User identity or account information
✕ Session history
✕ Any user-identifying signal
Governance metrics are not used to reconstruct individual conversations. Each event is logged by structure — signals, mode, latency, version — not by content.
⚠️ Known Limitations

This section is not optional — it is the most important part of this page. We believe honest disclosure of current limitations is more valuable to the field than any claim of achievement.

L1

CRF is a proxy metric. The Cognitive Risk Factor is computed from VIX and regime confidence — it is not a direct measure of epistemic uncertainty. We treat it as a signal, not a ground truth.

L2

Output modes are under active evaluation. The thresholds that determine SPEAK / SPEAK_WITH_CONSTRAINTS / REMAIN_SILENT were set a priori and are being validated against user feedback for the first time. The Living Lab study is that validation.

L3

The governance runtime does not guarantee correctness. It is designed to modulate confidence — not to produce correct answers. A SPEAK decision means the system believes a response is appropriate, not that it is factually accurate.

L4

User labels carry self-selection bias. Participants who choose to provide feedback may differ systematically from those who do not. We are designing controls to quantify this.

L5

Domain is financial advice only. All data comes from financial Q&A contexts. Governance behavior and user calibration may differ substantially in other domains. Generalization claims will require additional studies.