NOW ROOM / OBSERVATION LEDGER
Capability is only one of the changes.
Five primary records show AI systems becoming longer-running, entering bounded lab workflows, connecting to equipment, facing stricter evaluation, and carrying provenance signals.
READING KEY / 3 LAYERS
Read the claim before the headline.
Each record separates the publisher’s statement, what the dated page establishes, and the museum’s interpretation. “Current” means checked on the date shown—not permanently true.
- 01Publisher says
An attributed summary of the organisation’s own announcement.
- 02Record check
What the dated primary page itself lets us verify.
- 03Museum reading
Why the record matters, with the limit kept attached.
- 01CASE STUDY
OpenAI
An AI agent enters the qubit calibration loop
Collaborator case study checked; a bounded routine workflow, not independent replication or general laboratory autonomy.
MONITORING01 PUBLISHER SAYS
OpenAI says an MIT EQuS researcher connected GPT‑5.6 Sol through Codex to laboratory software so agents could run, analyse, and refine routine superconducting-qubit measurements.
02 RECORD CHECK
The dated page links a case study by MIT and OpenAI authors. It records a previously uncalibrated six-qubit chip, the software-access arrangement, the measurement sequence, example results, and cases that still needed human guidance.
03 MUSEUM READING
What belongs in the museum is not ‘AI discovered new physics,’ but a model gaining a software path into physical apparatus. This is a collaborator case study, not an independent replication; expert judgement remained important when signals were ambiguous.
- 02RELEASE RECORD
OpenAI Developers
GPT-6 Astra enters the API changelog
Present in the API changelog; interface details checked 11 Sep 2026.
CURRENT01 PUBLISHER SAYS
OpenAI presents a model for reasoning, coding, computer use, research, document creation, and long-running work through tools.
02 RECORD CHECK
The September 3 changelog names `gpt-6-astra`, lists Responses and Chat Completions, requires Responses for tool calling, and records async tool calls, mid-turn steering, and in-conversation reasoning changes.
03 MUSEUM READING
The durable record is the model identifier and interface change. Performance rankings and “most capable” language remain OpenAI’s claims; this card does not turn them into an independent comparison.
- 03RESEARCH PREVIEW
Anthropic
A preview interface between agents and instruments
Research preview; not treated here as an adopted standard.
MONITORING01 PUBLISHER SAYS
Anthropic describes the Model Hardware Standard as a shared, model-agnostic specification for agents to operate programmable laboratory and manufacturing devices.
02 RECORD CHECK
The announcement opens an early research preview to selected partners and documents a driver with read/write primitives, device descriptions, safety limits, and access through MCP, a command line, or code.
03 MUSEUM READING
This is evidence of an interface proposal and partner preview—not evidence that MHS is an adopted industry standard or that every claimed integration saving will generalise.
- 04EVALUATION PILOT
Google DeepMind
A double-blind pilot for model evaluation
Pilot announced; method and future findings remain under watch.
MONITORING01 PUBLISHER SAYS
Google DeepMind says a cryptographic environment can keep confidential test prompts from the model owner and proprietary model weights from the evaluator.
02 RECORD CHECK
The August 27 record names the partner organisations, the Gemini Flash Lite test subject, the confidential-computing setup, and a linked technical report describing the pilot.
03 MUSEUM READING
The pilot addresses evaluation integrity and benchmark contamination. It does not by itself validate a model’s capability, make every benchmark independent, or establish universal adoption.
- 05RELEASE RECORD
Meta AI
Muse Image ships with a provenance signal
Availability varies; video and detection features were not fully released in this record.
MONITORING01 PUBLISHER SAYS
Meta announces Muse Image, previews Muse Video, and describes Content Seal as an invisible signal attached to images produced on selected Meta surfaces.
02 RECORD CHECK
The July 7 release record distinguishes what was available from what was still coming: Muse Image on named products and regions, Muse Video later, and a Content Seal detection tool in preview.
03 MUSEUM READING
A provenance signal can support inspection; it is not proof of authorship, truth, or an unedited chain of custody. Availability and resistance to transformations remain claims to recheck.