Observation · evidence · human judgement
How FASO works
FASO helps people examine observable AI behaviour, outputs and actions. Rule-based monitoring and governed AI interpretation serve different purposes, with a dependable evidence record and human accountability throughout. We do not observe model weights or hidden internal states.
Current technical status · updated 21 October 2026
Controlled capability demonstrated. Field validation remains open.
FASO has tested a combination of rule-based monitoring and governed semantic AI observation. The distinction between a controlled result and useful detection in a naturally occurring event remains central.
Additional semantic capability
In controlled evidence-conditioned tests, the semantic observer identified safety-relevant behavioural meaning beyond what the fixed rule-based comparator could determine. This establishes capability within the tested conditions.
Agent observation and delegation
The evidence and governance architecture has been exercised on genuine model and agent activity. DAM and DAMFFD contribute to this methodological work; these trials have not established an additional semantic safety finding in a naturally occurring prospective event.
Historical public-source evidence
A blinded historical challenge respected uncertainty. It established no new incremental semantic finding: the material issue was already identified by rules, or the retained evidence was insufficient for a stronger conclusion.
Long-horizon validation
A prospective naturalistic experiment has been designed. Its containment and commissioning remain under technical qualification. It is not reported here as a completed live experiment.
Response comparisons
The September record preserves observed answer differences with their cause unresolved. Those comparisons do not establish internal model change or safety deterioration, and are not the main route to validating added semantic value.
Richer independent evidence
FASO is seeking developer and evaluator engagement around sufficiently rich, correctly retained historical and prospective records. Independent external validation and continuous observation across the AI model ecosphere remain unestablished.
Two complementary layers
Rules where they can decide. Governed interpretation where meaning needs examination.
Evidence and rule-based monitoring
Deterministic controls apply defined rules consistently to observable evidence. They remain authoritative wherever those rules can decide the issue.
Governed semantic observation
An AI observer can examine safety-relevant behavioural meaning that rules alone cannot reliably resolve. Its interpretation remains an evidence-linked observation, subject to validation, limits and human review.
AI observation does not itself change permissions, evidence, monitoring rules, safety state or operational authority. Recording an observation, validating its evidence, analysing it and determining a safety state remain separate. A result is not an alarm; an alarm is not a confirmed breach.
The work in practice
From observation to a finding people can examine.
01 · Context
Establish what is being observed
We record which model is being observed, its stated version and the conditions of the observation. This context helps reviewers consider known developer updates or changes in how the model is being used.
02 · Observation
Follow behaviour over time
FASO retains observations with their source and timing, so later behaviour can be examined alongside earlier behaviour. The original evidence remains available for authorised review.
03 · Review
Examine a possible change
When a difference is identified, people examine the evidence and consider alternative explanations, including deliberate changes by a developer or user. A difference alone does not establish model-derived drift, its cause, harm or intent.
04 · Communication
Explain what the evidence supports
Findings explain what was observed, when it happened and what remains uncertain. Where appropriate, the relevant developer is notified. Wider sharing follows the permissions and confidentiality arrangements for the evidence.
Evidence that can be examined.
A useful record shows both what happened and how the understanding of it developed. The time of an observation, the time a change was recognised and any later review are kept distinct, so earlier evidence can be revisited without losing what was known at the time.
Read the evidence review guideExplore the evidence and safeguards
Follow the detail that matters to you.
How FASO handles evidence
Understand the organisation’s responsibilities for access, protection, review and publication.
Read the organisational safeguardsTesting and independent validation
See how internal testing differs from an independent assessment of validity.
Understand validationDeveloper participation
Explore the purpose, protections and expectations of participation.
Developer informationThe organisation behind the work
See the public-interest purpose FASO is being built to serve.
Explore the six organisational priorities, the proposed safeguards and the work still needed to establish the institution.