ToMFormer / inspect the result
Five narrow proofs. Each starts from this page, states what you should observe, and says what the result does not establish. Live measurements identify the running deployment; banked results carry their own date and scope.
Correct → Teach → Reason → Verify → Scale
A mock response demonstrates the interface only. It does not complete a live proof.
Ask one question, apply one fact correction, ask the identical question again, then revert.
Withhold a curated document, measure its target questions, restore it, and rerun a fixed prior probe.
Inspect explicit two-hop and three-hop questions where the answer differs from the starting document entity.
Inspect precision and coverage together as TOMX serves fewer answers under recorded operating points.
Read allocated model bytes, served disk, shipped-but-unserved data, and process memory from this deployment.
Research-grade generative view
The speech panel emits only licensed entity names from the 5M-world candidate namespace, reports spoken rate with precision, and keeps candidate-absent rows silent instead of free-running an invented name.
Boundary: text-only name emission, not voice synthesis, celebrity impersonation, open-ended chat, or unlicensed generation.
GL-2 conversational readout
The Floyd panel shows a fixed GL-2 transcript with support-gated text emission, declined duplicates, and explicit ABSTAIN rows when a request crosses the displayed support boundary.
Boundaries: self-speech is only a card readout; cross-entity speech is blocked; ambiguous coreference abstains; coherence is local, not unrestricted dialogue planning.
Separate device / catalog deployed 2026-08-18
The same idea in hardware. A battery-powered unit reads a surface and answers with every label it has evidence for, each carrying its own score and its own threshold, while its catalog sits on a memory card. 1,219 gate reads produced no wrong name, the resident file has stayed the same size to the byte across every deployed catalog from 37 labels to 50, and a surface it has never seen can be taught in the field in seconds with nothing retrained.
Boundary: none of this runs on this web deployment and your browser cannot reach the device. It does not detect trace contamination and it does not identify pathogens. It never reports a surface as safe.
Separate system / recorded 2026-08-19
The same idea in conversation. Every person or agent in the room is one editable record: enrolled voices are named live, a stranger reads NOVEL-UNKNOWN until enrolled, and an agent speaks only when a read of its own record licenses it, with silence as the default. On the hardest recorded meeting bench it named wrongly once in 1,175 display events at coverage 0.209, and its agents produced zero false interruptions through real speech recognition. The page carries a recorded run you can play.
Boundary: none of this runs on this web deployment; the page replays a recording of the real system, labeled as a recording. It is not spoof-proof and never claims to be, and it never claims prompted agents cannot be silent.
LIVE
A current request against the running driver. Capture and deployment context belong with the result.
BANKED
A prior measurement with a stated date, scoring unit, sample count, and available deployment pin.
RESEARCH-ONLY
A different artifact or harness. It cannot complete one of these deployed proofs.
BLOCKED
A validity or provenance gap prevents comparative use. N11: the deployed RAG text path remains blocked from comparative claims until the corrected artifact and build basis are publicly pinned.
A trial token exposes scoped ask, edit, grow, and revert operations. It expires and excludes bench runs. Read the contract or inspect the machine-readable OpenAPI document.
curl -s -X POST https://tomformer.com/api/v1/trial-token
# scopes: ask, edit, grow; bench excludedAgent-readable versions: /agents.md and /llms.txt.
Each page states what it shows and what it does not in its own footer. That footer is the scope of its claim.