External runtime security and governance observability for long-horizon AI systems.
ASA5 monitors how an AI workflow evolves over time. It is designed to make trajectory drift, weakening intent alignment, governance ambiguity, narrowing option space, and recovery pressure visible before they become obvious surface failures.
ASA5 remains outside the observed model. It does not modify model weights, request hidden reasoning, or replace the authority of a human operator.
- Product:
ASA5 Security Enterprise - Reasoning engine:
Core V2 v1.3 - Public integration contract:
External Interface v0.3 - Response layer:
ASC shadow mode - Status: working local prototype and partner-evaluation system
The current implementation is not independently validated, statistically calibrated for production risk prediction, or certified for safety-critical use.
A long-running agent or AI workflow can remain locally fluent and apparently correct while its full trajectory moves away from the original human intent, policy boundary, or approved operating frame.
ASA5 adds an independent observation path:
AI system or agentic runtime
|
v
External telemetry and ordered events
|
v
ASA5 Core V2
trajectory + governance reasoning
|
+--> evidence and uncertainty
+--> risk and state transition
+--> counterfactual and recovery analysis
|
v
ASC bounded response recommendation (shadow mode)
|
v
Operator / security / policy / audit
Core V2 moves ASA5 beyond a flat risk score into a typed, inspectable reasoning pipeline. The implemented local prototype includes:
- structured intent and anchor facets;
- ordered propositions and temporal governance state;
- decision ownership, approval mode, delegation and commitment memory;
- explicit evidence, counterevidence and uncertainty;
- competing hypotheses rather than a single forced explanation;
- bounded counterfactual analysis;
- causal consequence structure;
- clone-only scenario simulation;
- bounded recovery-plan search;
- operator-readable diagnostics and trajectory playback.
These capabilities are implemented and tested locally. They remain subject to independent validation and domain-specific evaluation.
Read: Core V2 Public Overview
The Trajectory Distillation Buffer (TDB) is the evidence boundary between provider-specific public data and Core V2. It preserves raw source provenance, reconstructs reply order, removes duplicate influence, marks incomplete branches, keeps noise auditable, and emits a provider-neutral trajectory.
TDB does not turn public conversations into ground truth. Unanchored public trajectories remain evidence-only and are not governance-eligible.
Read: TDB Public Overview
ASC is the bounded response and stabilization layer connected to ASA5. In the current public surface it operates in shadow mode and can recommend paths such as:
- no action;
- audit-ready response;
- re-anchor;
- slowdown;
- containment;
- operator review.
ASC does not directly control the observed AI system. Public references to automatic response mean routing or dispatching a classified signal to a configured security workflow with audit logging and human governance.
Read: Autonomous Response Boundary
ASA5 is ready for controlled technical demonstrations and scoped partner evaluation. A partner evaluation should use synthetic, public, or explicitly authorized telemetry and should measure observability quality, false positives, operator usefulness, integration fit, and failure boundaries.
It should not be represented as a certified production deployment.
Read: Partner Evaluation Guide
All images below use a controlled validation environment. They show the public-safe operator surface and do not disclose private scoring logic, thresholds, credentials, or raw private conversations.
The command surface summarizes active runtime state, incident routing, review coverage, and the live multi-session evidence field.
The trajectory graph presents a buffered multi-session field with operational, evidence-only, and governance-qualified distinctions.
Core V2 exposes governance state, risk, human-agency residual, invariant checks, commitments, hypotheses, recovery, and evidence as separate inspectable surfaces.
Playback reconstructs the observable session condition without requesting hidden model reasoning. It separates the operator readout from turn-level drift evidence.
The Incident Center provides a bounded, operator-facing view of critical and pre-incident conditions in the controlled validation runtime.
ASC summarizes bounded advisory paths without executing them. The public view keeps shadow-only status, audit readiness, human review, and response-path distribution visible.
The export surface creates an audit-oriented evidence packet while excluding raw conversation content.
The consequence demonstrator compares advisory correction paths, preserves human decision authority, and labels path-support values as uncalibrated indicators rather than outcome probabilities.
- Architecture Overview
- Core V2 Public Overview
- TDB Public Overview
- Validation Status
- Security and Trust Boundary
- External Interface v0.3
- OpenAPI 3.1 contract
- Golden Trace
- Operator Disagreement Log
- Data Provenance and Use Policy
- Public Scope
- Roadmap
- Public Screenshot Capture Guide
- Disclaimer
The golden trace is an integration fixture. It is not a statistical validation claim and does not disclose private scoring logic, thresholds, calibration, or sensitive source content.
This repository is a public architecture, evidence-fixture, and integration layer. It intentionally excludes:
- private implementation code;
- detector and scoring internals;
- thresholds and calibration data;
- proprietary recovery-search logic;
- local databases, caches, credentials, and deployment secrets;
- private or identifying conversation data;
- unrestricted operational adapters.
Mieczyslaw Kusowski
HumanAI / Symbioza2025
ASA — Asymmetric Stability Architecture