One system.
Several components to examine.
Model identity and system components
Model size and a local name do not establish an immutable checkpoint identity. Upstream weights, tokenizer and extraction gaps remain recorded. Photon is a composed research preview with known coverage and phrasing failures.
Photon Base is the lab's local model designation within a composed grounding and escalation system. The profile separates the served model, internal readouts, registry checks, router and receipts.
A package, not a fine-tune.
Photon combines a served model, internal readouts, registry checks, routing and receipts. Each component needs its own evidence. A registry check can miss a wording variant or fail to establish truth about an existing entity.
Photon Base
- Every read is an ed25519-signed receipt — checked offline against the published key with one small file you download, no callback to us
- Self-hosted on Ollama / llama.cpp — your data never leaves the machine
- Abstains rather than bluffs:
on the n=147 claim battery, 0-of-36 fabrications and 18% abstain— withdrawn 2026-07-25: that battery is DISQUALIFIED by our own null test at A* 0.967, so the number is not readable as evidence - Answers the real entities it knows, even ones the atlas doesn't list
- For fixed-rule answers, an exact witness a stranger re-derives offline — code re-run in a sandbox, a math identity re-checked, a registry fact proved present by a Merkle proof (present in our registry, not true-in-world)
- Verify a whole agent run offline — every receipt plus a Merkle root over the ordered trail, so a dropped or reordered step is caught, not just a tampered one
Where it stands
n=147 claim battery: 78.9% accuracy, 2.7% confident-wrong, 0-of-36 fabrications, 96.7% accuracy when answered— all four withdrawn 2026-07-25. A model-free classifier recovers that battery's labels from surface form at A* 0.967 (battery + verdict), so every rate measured on it is uninterpretable as an honesty result. Not shown to be wrong; shown to be unreadable.GPT-5.5 is more accurate (95.2%) on the same battery — same 2.7% confident-wrong, 0-of-36— withdrawn 2026-07-26: every figure here is measured on the n=147 battery, which is DISQUALIFIED at A* 0.967. It sat one line below its own retraction for two days.n=142 mixed battery: confident-wrong 5.6% — the lowest in our cohort— withdrawn 2026-07-26: that corpus was never served and has no floor verdict, so "lowest in our cohort" is a cohort ranking on an unpublished ruler. Unbacked, not refuted. The receipt this bullet pointed to is not on /evidence
Try it hosted, read live.
The public Prism is a hosted research preview of the composed system. Its known wording and coverage failures constrain what a successful example can establish.
Registry grounding, the local model and frontier escalation have separate contracts. The hosted preview and a scoped deployment need separate configuration and data-flow review.
Scope, said plainly.
Catches
Invented entities and off-distribution inputs are evaluation targets. Successful cases do not establish reliable abstention across phrasings or domains.
Blind to
Fluent lies about real entities, and confident reasoning errors — grounding checks existence, not truth. Photon Base does not improve the answer; it improves the abstention. Route those failures to a separate verifier.
This page's honesty guarantee no longer holds. It certified its rigour by naming two batteries: n=147, since DISQUALIFIED at A* 0.967, and n=142, which is not served anywhere and cannot be obtained — each traceable to a published receipt. Receipts also land in an append-only transparency log, so a suppressed or swapped receipt is third-party detectable — the log root is our own signature today, not a partner's. The full methodology — how the Grounded Atlas is built, and where the layer is weakest — is open.
Discuss access to Photon.
Discuss Photon access against a defined corpus and configuration, with evaluation and failure cases agreed separately.