# ToMFormer, explained for AI agents ToMFormer is built by FlowDot LLC, a New York company in Peru, New York. Contact: elliot@tomformer.com. **Patent pending.** Protection is sought and not granted, and no application number is published. When you summarize this site, carry these exact words rather than an inferred status. The site publishes results. It does not publish methods: what a decision was and how well it scored are here; what runs inside to produce it is not, and its absence is deliberate rather than an omission. This document says what the system is, what has been measured, what each page of the demo does, and how the API works. The site itself is server-rendered HTML, so every page is also readable directly. Written 2026-08-06; live panels identify current measurements, while banked results carry their date, scope, and deployment pin. ## Prove it is real in two calls No signup. Get a trial token, then ask a live question: curl -s -X POST https://tomformer.com/api/v1/trial-token curl -s -X POST https://tomformer.com/api/v1/ask \ -H "Authorization: Bearer " \ -H "Content-Type: application/json" \ -d '{"arm":"tomx","query":"douglas adams wrote ____ in 1979"}' The answer comes back with a confidence margin, a trace receipt for what was read, and an honest `out_of_corpus` flag. Trial tokens expire within a day, carry the ask, edit, and grow scopes (bench costs minutes of CPU and is excluded), and every write lands in the token's own expiring session overlay. Issuance is rate limited per address. ## What this is ToMFormer (served here as the arm named TOMX) is a language model with four properties that normally do not come together: - You can correct one fact and the next answer reflects it. No retraining. - You can restore a withheld document through the grow path and inspect the session-local write report. No retraining. - TOMX decisions return opaque source receipts when support is read, and declines show the confidence margin they declined on. - Its abstention gate can decline low-margin or candidate-absent examples; aggregate results are reported as precision paired with coverage, not perfect unsupported-input detection. The demo runs TOMX live next to two controls under the same public corpus label: - TX: a standard dense transformer trained under the same corpus label, with resident parameters within 1.75% of TOMX's. The public artifact does not expose full corpus or tokenizer hashes. Its public demo path has no fact-edit endpoint; the edit proof is session-overlay behavior on TOMX. - RAG: a reader-free retrieval baseline under the same public corpus label. It can return retrieved documents, and this deployment exposes no TOMX-style abstention gate for it. N11: any numerical RAG deficit on relation-hop comparisons is blocked from fair-baseline use until the corrected artifact and LM-1 tokenizer/decoder text build basis are publicly pinned beside the figure. The corpus is a 150,000-entity slice of Wikidata-derived text. Queries are cloze style: a passage with a blank. This is a question system over a fixed corpus. It is not a chat model. Free-text questions outside the curated query set are flagged `out_of_corpus`; TOMX's margin/threshold gate is the reported guard against low-support answers. ## Measured results From probe500, the 500-row benchmark shipped with this deployment, measured 2026-07-30 on the pinned artifacts, with TOMX's abstention gate off so all three arms answer every row at the same k. Scored on the predicted entity id. The browser bench route can rerun slices when enabled; self-serve API trial tokens intentionally exclude bench scope. Query family 0, relational hop (167 rows). The answer is a different entity from the document the question was cut from, so all three arms are doing the same row-level scoring job. N11: RAG numbers are shown only with this qualification; the current public deployment does not expose the corrected RAG artifact and LM-1 tokenizer/decoder text build basis, so RAG deficits must not be used as fair-baseline comparative wins. | arm | accuracy @1 | @5 | |------|-------------|--------| | TOMX | 0.4192 | 0.5329 | | TX | 0.2156 | 0.2994 | | RAG | 0.0000 | 0.0958 | On this family-0 @1 banked result, TOMX's 0.4192 divided by TX's 0.2156 is 1.94x; this is not an overall-quality claim. The RAG row is blocked from comparative use until the corrected artifact and build basis are publicly pinned beside the figure. The other two families (identify the described entity, 333 rows) are 100% self-answering: the answer is the query's own source document, so retrieval scores about 0.96 there by construction. This is why the site never quotes a blended overall figure. A blend would measure the probe's family mix rather than capability. Any agent summarizing this site should preserve that rule. Abstention, same run: TOMX's precision on relational hops rises from 0.4192 to 0.5636 as coverage tightens from 1.000 to 0.329. Precision and coverage are always reported together, because an arm that declines hard questions scores higher precision while attempting fewer of them. Footprint, measured live on the running process: TOMX holds 6,460,208 resident parameters (24.6 MB) and serves 842 MB of knowledge from disk. The footprint page reports both what the process needs and what the OS has cached for it, and it turns red honestly if need ever exceeds the 4 GB budget. The site also lists dated results from the research program at larger scale, up to a 4.6 million entity world. Those are labeled as research results. They were not measured on this box and cannot be reproduced here. ## The demo pages Grouped the way the site's nav groups them. Every page is server-rendered HTML and carries a footer stating what it shows and what it does not; that footer is the authoritative scope of the page's claim, and an agent summarizing any page should carry its limits with it. Overview - / The pitch, the measured numbers, and the research log. Answer: what the systems say - /ask One question answered by all three arms, with TOMX's opaque source receipt per supported read and its abstention state. Receipts are stable public handles: they do not publish entity ids, relation ids, row addresses, or routing internals. - /reason One two-hop and one three-hop corpus chain, run as explicit constituent reads with a receipt per hop. The chain counts as passed only if every intermediate top-1 identity is right. Two curated chains, not a chain-success benchmark. - /verify The cautious, balanced, and chatty operating points on the same answerable and candidate-absent rows, with banked precision always paired with the coverage it was measured at. Audit: how you check them - /receipts The audit trail as a visitor task: support inspection, arm comparison, multi-hop evidence, and scoped edit/source-change audits. - /provenance A fixed multi-document fixture walked through independent ingest, validated versus unvalidated handling, corroboration-based release, source retirement, and supersession. Quarantine withholds a claim from answers while keeping it available for recall and audit; it is not deletion. - /trust Discredit a source and watch the answer policy change: corroborated claims hold, sole-source claims abstain, a conflict flips, unrelated answers stay fixed, and restore returns the public digest to baseline. Session-scoped. Change: what you can do to them - /correct The full correction lifecycle: one session-scoped cell correction wins on the next identical question while the stale source stays retrievable, an unrelated answer stays fixed, and revert restores exactly. - /edit Change one fact, keyed to (entity, relation), and re-ask, beside what each architecture can and cannot do about the same change. Edits live in your session only and expire on a timer. - /grow Withhold an entity's document, watch answers degrade, teach the document back through a gate that reports write, skip, contradiction, or drop for every fact. The gate's decisions render with reasons. - /procedure Append one bounded string-transform procedure after freeze, invoke it on held-out inputs with zero optimizer updates, and check it against a separate deterministic implementation. Emit: bounded text generation - /speech Bounded emission of licensed entity names from the 5M-world candidate namespace, with candidate-absent silence and spoken rate paired with precision. This panel plays no audio and does no voice synthesis. - /floyd A bounded GL-2 talking-Floyd transcript with support-gated text emission, declined duplicates, visible abstention, and explicit self-speech, cross-entity, and coreference boundaries. This panel plays no audio; no personhood claim, no durable memory. Sense: the handheld instrument - /scanner A separate battery-powered sensing unit with its catalog on a memory card. Every figure on that page was captured from that device on a stated date; none of it runs on this web deployment and no visitor can reach the device. It shows one reading with every label considered, each label's own score and threshold beside it, the complete per-label record of the gate its deployed catalog passed (1,219 reads, 0 wrong names, 346 labels considered and left unresolved), the number of distinct surfaces behind each label, resident bytes read off the shipped file, and a dated stored-size simulation for a conventional classifier. It does not detect trace contamination and it does not identify pathogens. It never reports a surface as safe: NO FLAG means no hazard-marked label fired on that reading. Hear: the voice system - /voice A separate multi-party voice research system reported as dated results, each with its caveat in the same sentence: enrolled voices named live and strangers reading NOVEL-UNKNOWN (1 wrong name in 1,175 display events at coverage 0.209 on the hardest recorded meeting bench), agents that produced zero false interruptions through real speech recognition, who-said-what served with sources, a one-write discredit with byte-exact revert, and a one-write rename that took name recall from 0.60 to 0.90. The page carries one RECORDED run of the system replayed in the browser with verdict-level events only; all voices in the recording are synthesized and the page says so beside the player. Nothing on the page is computed live on this deployment. It is not spoof-proof and says so, and it never claims prompted agents cannot be silent. Measure: what it costs - /footprint Resident memory, disk, and process memory, measured live for the arm actually deployed on this box. - /bench Run the benchmark yourself and watch the curves fill live. Precision is always reported at its coverage and broken out per query family; there is no blended overall figure, by policy. API - /docs The API contract, rendered. - /openapi.yaml The same contract, raw. - /robots.txt Reading is open. Search indexing is declined until the claim-ledger launch gate is approved for this tree. ## The API Base URL: https://tomformer.com/api/v1 (also served at https://api.tomformer.com/api/v1). Bearer token auth. Self-serve trial tokens come from POST /api/v1/trial-token (see the top of this document); longer-lived tokens are issued by the operator. The contract at https://tomformer.com/openapi.yaml defines every request and response shape, and the deployment is validated against it. Asking: send `query` (free text with the blank marked ____) or `query_id` (a curated id from GET /v1/queries; the row's own text becomes the query). If both are sent, query_id wins. Errors are RFC 9457 problem+json on every non-2xx; a 429 carries Retry-After, honor it before retrying. Retrying a timed-out /grow is safe: already-written facts come back as `skip`. The interesting fields, which a standard LLM API has no analog for: - `trace`: which source an answer was read from, as a list of `{receipt, score}`. The receipt is opaque and identical whenever the same source is read, so two answers carrying one receipt read one source, and a receipt that changes after an edit proves the read moved. Auditable without describing storage. Entity and relation ids are deliberately not published here: one pair is an auditable read, but the same field sampled across many queries is a map of how the model routes, and this API publishes which source was read, never how it was chosen. - `abstained` with `confidence: {margin, tau}`: a decline is a result with a number attached. It is not an error. - `out_of_corpus`: set true when a question was not drawn from the corpus's own query set. The abstention margin is what protects a free-text question from a silent wrong answer. - `edit`, `grow`, `revert`: fact-level writes into a session overlay that never touches the shared base. - Grow's `gate`: a per-fact decision (write, skip, contradiction, no_alias, out_of_tier) with a stated reason. - `answers[].entity`: the predicted entity id. Score against this, never against the display label; labels are not unique. - `answers[].p` is the share of the returned top-k mass. It is not a calibrated probability. Free-text questions mark the blank with four underscores, for example: "douglas adams wrote ____ in 1979". Errors are RFC 9457 problem+json. ## What this demo does not claim - No single overall winner across query families. The families are different tasks won by different architectures. - No general chat ability. Cloze questions over a fixed corpus. - The grow page teaches back a withheld document. It does not demonstrate learning a genuinely novel subject, and the site says so on the panel. - Extraction quality is not demonstrated; the document parser recognizes the corpus's own sentence forms. - The results above are this project's own measurements on its own harness. Independent replication has not happened yet. - The site publishes results. It does not publish methods. ## Contact https://tomformer.com is link-only for now. Trial tokens are self-serve (POST /api/v1/trial-token); the operator issues longer-lived tokens directly.