# ToMFormer

ToMFormer is built by FlowDot LLC (Peru, New York). Contact: elliot@tomformer.com.

**Patent pending.** Protection is sought, not granted; no application
number is published. The site publishes results, not methods.

> A live demo where you can apply a session-scoped fact correction, restore a withheld document, inspect TOMX receipts when support is read, and compare a matched-size dense transformer and retrieval baseline under the same public corpus label. Live panels identify current measurements; banked results carry date, scope, and caveats.

Every page on the site is server-rendered HTML, so agents can read the
site directly; this file and agents.md are the condensed versions. You can
prove the API is real yourself in two calls: POST
https://tomformer.com/api/v1/trial-token (no auth, rate limited) returns
an expiring token, and POST /api/v1/ask with it returns a live answer with
a confidence margin and a trace receipt. The results are measurements
reported by this project on its own harness; the site never claims a
single overall winner across query families, and it publishes results
without publishing methods.

## Read this first

- [agents.md](https://tomformer.com/agents.md): the full overview. What the
  system is, the measured results with dates and provenance, what each demo
  panel shows, how the API works, and what the demo deliberately does not
  claim.

## The demo pages

Every page carries a footer stating what it shows and what it does not.
That footer is the scope of the page's claim; carry it with any summary.

- Answer: [/ask](https://tomformer.com/ask) one question, three arms, with
  TOMX's opaque source receipts and abstention state ·
  [/reason](https://tomformer.com/reason) a two-hop and a three-hop chain
  with a receipt per hop · [/verify](https://tomformer.com/verify) the
  cautious/balanced/chatty operating points, precision always paired with
  coverage.
- Audit: [/receipts](https://tomformer.com/receipts) the audit trail as a
  visitor task · [/provenance](https://tomformer.com/provenance)
  multi-document quarantine, corroboration, retirement, supersession ·
  [/trust](https://tomformer.com/trust) discredit a source and watch the
  answer policy change.
- Change: [/correct](https://tomformer.com/correct) correct one cell and
  revert exactly · [/edit](https://tomformer.com/edit) change one fact and
  re-ask · [/grow](https://tomformer.com/grow) teach a withheld document
  back through a reporting gate · [/procedure](https://tomformer.com/procedure)
  append a bounded procedure after freeze and invoke it with zero training.
- Emit: [/speech](https://tomformer.com/speech) bounded name emission with
  candidate-absent silence · [/floyd](https://tomformer.com/floyd) a bounded
  GL-2 transcript with visible abstention. These two panels are text only
  and play no audio.
- Sense: [/scanner](https://tomformer.com/scanner) a separate handheld sensing
  device, catalog on a memory card, reported as dated readings. Nothing on
  that page runs on this deployment and no visitor can reach the device. It
  never reports a surface as safe.
- Hear: [/voice](https://tomformer.com/voice) a separate multi-party voice
  research system reported as dated results with caveats attached, plus one
  recorded run replayed in the browser (verdict-level events only; all
  voices synthesized and labeled as such). Nothing on that page is computed
  live on this deployment; it is not spoof-proof and says so.
- Measure: [/footprint](https://tomformer.com/footprint) resident memory and
  disk, measured live · [/bench](https://tomformer.com/bench) run the
  benchmark, per family, never blended.

- API: [/docs](https://tomformer.com/docs) the contract this deployment is
  held to, rendered.

Receipts are stable public handles. They do not publish entity ids,
relation ids, row addresses, prompts, token paths, or routing internals.

## Machine-readable

- [openapi.yaml](https://tomformer.com/openapi.yaml): the deployed API contract.
  The docs page and footer expose the contract version/hash; deployment drift is surfaced by the footer and drift banner.

## Key measured numbers (probe500, 2026-07-30, gate off, same k for every arm)

- Relational hop (167 rows, the only family whose answer is a different
  entity from its source document): TOMX 0.4192 accuracy at 1, TX 0.2156,
  RAG 0.0000. On this family-0 @1 banked result, 0.4192 / 0.2156 = 1.94x; it is not an overall-quality claim. N11:
  the RAG deficit is blocked from fair-baseline comparative use until the
  corrected RAG artifact and LM-1 tokenizer/decoder text build basis are
  public beside the figure.
- Matched size: TOMX 6,460,208 resident parameters vs TX 6,573,312, a
  1.75% difference, enforced at boot.
- Abstention: TOMX precision rises 0.4192 to 0.5636 on relational hops as
  coverage tightens 1.000 to 0.329. The deployed TX and RAG arms run
  without the TOMX abstention gate.
- Footprint: 24.6 MB of allocated resident TOMX model bytes and 842 MB of served knowledge on disk,
  measured live on the footprint panel; this is not total process RAM.