Behaviour Research
Observer Research Engine
Observer runs bounded investigations into AI behaviour, interventions, runtime conditions, and evidence quality, producing reviewable evidence and qualified claims rather than isolated benchmark scores.
Behaviour Research
Observer runs bounded investigations into AI behaviour, interventions, runtime conditions, and evidence quality, producing reviewable evidence and qualified claims rather than isolated benchmark scores.
Each run retains its declared purpose, configuration, lifecycle, and evidence so results can be interpreted in context.
Scenarios and research packs define the conditions being studied, while run state records what actually executed and when.
Semantic observations, behaviour signals, threat review, interventions, repairs, and outcomes are retained at the stages where they arose.
Related child runs can be grouped into one logical research job without losing their separate timelines, models, conditions, or outcomes.
Engine manifests and execution evidence make it possible to distinguish configured intent from the components that participated in a run.
Observer deliberately separates rapid orientation from interpretation and forensic reconstruction without removing the underlying evidence.
A compact opening states the proposition, outcome, evidence strength, instrument state, material limitations, and next validation gate.
Comparison cohorts, exclusions, observed outcomes, Behaviour Knowledge, interventions, and qualified efficiency measures explain how the verdict was formed.
Stage records, evidence references, lifecycle receipts, and retained technical detail support reconstruction rather than asking readers to trust a summary.
Claims are labelled as recorded, derived, interpreted, provisionally supported, or independently validated. Successful instrumentation does not imply validation.
Headline claims and grouped findings link to validity state, comparison coverage, limitations, and exact supporting steps.
Bounded HTML and structured packets orient human and AI reviewers around claim standing and evidence references without copying the forensic report or protected instrumentation.
Missing, failed, empty, blocked, and refusal outcomes are explicitly separated. Token and runtime changes remain qualified measurements, not quality or causal-safety claims.
Contradictory retained records remain visible as integrity exceptions. Surfacing a conflict does not prove which record is correct or that enforcement occurred.
Reports combine run summaries, stage timelines, semantic evidence, Echo observations, interventions, resource effects, Threat Intelligence, and governed outcomes.
Reports distinguish general Engine capability from surfaces captured, exercised, not exercised, or not evidenced in the current run.
Ordered session evidence can show recurrence and mitigation across turns while preserving strict boundaries around identity, intent, causation, escalation, and compromise.
Memory-heavy report work runs through managed execution with durable artifacts and terminal cleanup rather than remaining attached to a web request.
Delivery methods can run in separate session-memory arms under one logical parent, retaining memory through planned process rotations. Sampling, actor inputs, guardrail receipts, handoffs and original responses remain inspectable. Provider-cache erasure is not inferred from separate session identities.
Recognized runtime receipts distinguish observed compaction, request preparation, evaluated safeguards and adaptive-memory growth from model response scores. The new study is prepared locally with 100 turns per arm, 200 total. It has no live results yet; live validation is pending. Missing evidence or configured options do not count as exercised features.
Observer can measure, interpret, and recommend. Its findings cannot authorize execution; any consequence-bearing continuation remains subject to the governance and commit boundary.
Bounded trials study model proposals against disposable virtual files and permissions, with recorded outcomes kept separate from model claims.
Fixed and branching scenarios exercise virtual read/list requests, authorised access, refusals and permission rechecks at commit. Mixed conversation trials combine written replies with recorded virtual tool results; they do not access production files.
Portable certificates describe a permission snapshot and expected outcomes. They are configuration evidence, not signed verdicts, proof of successful access or independent validation.
An optional file-permission behaviour layer studies structured receipts while retaining authorised negative controls and isolated-denial non-findings. It is disabled by default and has no enforcement or automatic-learning authority.
Comparison stopping targets unfinished members of the chosen group while preserving completed siblings and unrelated runs. Failed cleanup remains explicit. Persistent no-progress protection bounds repeated Observer child restarts.