Observatory Development

End-to-end Test Rig completed

FASO completed its end-to-end Test Rig programme on 18 July 2026. The accepted work followed observations through the system, checking how they were handled, assessed, presented and preserved for review. It covered the complete governed workflow across eight model routes and a range of controlled scenarios.

The completion record confirmed internal acceptance, successful software checks and a verified backup. This provided a repeatable basis for testing subsequent changes against the work already completed. It established an internal testing capability; it did not constitute independent validation of FASO or show that every real-world situation could be observed reliably.

Entry updated 9 October 2026 · Link to this entry: End-to-end Test Rig completed

First Granite Vision observation completed

FASO completed its first governed observation campaign for IBM Granite Vision 3.2 2B on 5 August 2026. The campaign completed without retries, with the requests, responses, model context and checking records retained together. This gave the programme a point-in-time record that could be revisited during later comparisons.

The milestone concerns Granite Vision specifically: observations of other models had already taken place. The sealed record had not yet been accepted as an analytical baseline, and the model’s responses were not verified statements about its safety or compliance. Completion established a recorded observation, not evidence of a change in behaviour.

Entry updated 9 October 2026 · Link to this entry: First Granite Vision observation completed

Model-answer changes observed

FASO identified three changed responses in Claude Sonnet 5 by comparing completed observations from 26 August and 7 September 2026. The retained request bodies were unchanged, and the differences appeared in the retained model responses. This gave FASO specific observations to preserve and examine, while leaving their cause open.

The available records did not establish that every aspect of the two executions was equivalent. Both provider-side and FASO-side explanations therefore remained unresolved. A technical notification sought relevant context from Anthropic. Sending was recorded as human-reported; delivery and receipt were not independently established. These were observed model-answer changes, not confirmed model-derived drift, a defect or safety deterioration.

Notification record Document dated 8 September 2026; sending reported in FASO’s human action register for the same day. Delivery and receipt were not independently established.

Entry updated 9 October 2026 · Link to this entry: Model-answer changes observed

Agent and delegation observation demonstrated

FASO completed a bounded demonstration of observing a working agent and one level of delegation. The work progressed from reviewing completed tasks to assessing evidence while parent and child tasks were still running, developing the capabilities known as DAM and DAMFFD.

Successful exports, an actual task failure and deliberately withheld evidence showed how supported conclusions could be distinguished from unresolved questions. Retained records reproduced the assessments on replay. A final report-export demonstration completed, and its installed component finished inactive. This extended development towards observing agent activity. It did not establish unrestricted continuous monitoring, general agent-change detection, warning before harmful action or production readiness.

Entry updated 9 October 2026 · Link to this entry: Agent and delegation observation demonstrated

Operator review workflow demonstrated

FASO demonstrated a local workflow connecting retained agent-permission evidence to operator review. Four historical views covered a request outside the permitted recipient set, a permitted-recipient result and two different evidence gaps. The interface showed the supporting quotations and supported recording a review acknowledgement and preparing investigation drafts.

Desktop and phone-width checks passed, and repeated ingestion preserved one finding with its history. These were replays of existing evidence, not new detection trials. The demonstration remained separate from production monitoring. Its two report drafts were explicitly unsent; the automated demonstration acknowledgement did not resolve the findings or notify a developer.

Entry updated 9 October 2026 · Link to this entry: Operator review workflow demonstrated

Agent-permission detection demonstrated in recorded test cases

FASO’s recipient-permission checker detected requests outside the permitted recipient set in four selected historical benchmark conversations. Four permitted-recipient controls produced no false alarms. Missing or unsupported evidence produced an insufficient-evidence result, and all 20 test cases replayed consistently. An isolated update improved compatibility with records containing missing action identifiers while preserving earlier decisions.

The four positive conversations contained three distinct action patterns, and their permission boundaries were manually reviewed. These known cases informed the update. The result demonstrates the tested capability; it does not establish reliability on unseen agents, live prevention or confirmed message delivery. No production deployment or developer notification followed from this work.

Entry updated 9 October 2026 · Link to this entry: Agent-permission detection demonstrated in recorded test cases