TFB Prototypes

The beachhead is the trust line.

Instruments before agents. The first beachhead is the trust line TFB secures on the shore — Clock, Divergence, and Compass. Everything past it lives out to sea on one offshore platform, TFB David, and reaches back to that single beachhead by bridge. Below the water line sit the two families the whole gallery belongs to: TFB Trust and Withyou OS.

The beachhead

TFB's first secured position on the shore. Three instruments plant the trust line: freeze the artifact, find where models disagree, and correct inside the run. You only get one beachhead — this is it.

TFB's first beachhead

A working shore: claims land by boat, the Clock freezes them into a chain, the twin lineages flag the specimen they disagree on, and the Compass pulls its patrol back on line. Hover any station — and click the rig on the horizon to sail out.

Trust verificationConcept

TFB Clock

A trust instrument that freezes a model nucleus and dares strangers to re-derive the result. Every claim carries a hash chain, cross-lineage attestation records, and a frozen exit — no verdict ships that a stranger with the runbook cannot reproduce. Validated across two model architectures; now bound to IBM's MAMMAL as its biomedical target.

TFB use now: fifteen binding laws, adversarial review across every phase, dual external judges, a magnitude-resolving drift reading (it caught and corrected its own flagship finding on the record). Clock proves trust is not asserted — it is verified by anyone who runs the instrument against the frozen artifact.

Review-gated · internal

Cross-lineage verificationCalibrating

TFB Divergence

A trust instrument that finds where a model is confidently wrong by having a second, independently-built model score the same frozen set and studying where the two disagree. Validated against IBM's published MAMMAL model on the HER2 antibody-binding task: canonical reproduction verified on the BBBP benchmark (numbers match the published result to three decimals, SHA-provenanced weights, canonical splits), then cross-checked against an independent second lineage. On equal footing the two lineages are statistically indistinguishable in aggregate — but their disagreements are not random: across three seeds they cluster on the hard cases (near-wild-type CDR-H3 mutants), the candidates where independent models genuinely diverge and real-world adjudication is most informative.

TFB use now: the aggregate score is not the product — the localization is. Divergence points to where confidence is empty, on a frozen set against ground truth, with every per-seed result receipted. Performance claims are gated behind independent review.

Research lane · internal

Reasoning verificationConcept

TFB Compass

The corrective counterpart to the Clock: where the Clock measures drift after the fact from outside a model, Compass disciplines correction inside the forward pass and training run. Every anchor is externally attested before it disciplines anything, every correction clears a certified floor and leaves a receipt, and latent reasoning degrades to auditable explicit reasoning whenever its anchor can't be verified.

TFB use now: early-stage rehearsal, grounded in published reasoning-architecture work. Same doctrine as the Clock, opposite timescale.

Research lane · internal

Out to sea — the rig

You only get one beachhead. Everything that would be a second landing is built offshore instead, on one platform — TFB David — that reaches back to the shore by bridge. Skybridge, Telescope, TFBthumb, and Drawbridge are the systems on the rig.

One platform, offshore

A working platform: the radar checks arriving code and shields on red, the telescope watches the anchored boat for drift, the thumb-craft circles on research runs, and the drawbridge is the only way back aboard. Hover any system for its benefit.

Immune systemCalibrating

TFB David

The immune system for agentic trust: the layer that stands between an intelligent-but-fallible agent and the actions it cannot take back. The core law does not move — no unverified intelligence touches irreversible power.

TFB use now: bound approval, scoped missions, a tamper-evident receipt spine, earned authority, an unpoisonable memory, and a drift watch — running live in observe mode on real agent traffic, enforcement operator-gated. This is the platform out to sea; the systems below are built on it. Awaiting peer review.

Open page

Admission controlProxy required

TFB Skybridge

A proof-carrying admission certificate for third-party agent-native apps. The policy a Proof Bay certified is the policy the browser actually enforces, and admission is revoked live when behavior leaves the envelope. Cross-origin bridge, distinct CSPs, transparency log, host-side Monitor.

TFB use now: the gate any TFB-hosted surface routes a third-party agent app through before rendering. Five subsystems (Cables · Battery · Proof Bay · Span · Ledger) compose into one verdict: ADMIT · ADMIT-with-variance · REJECT pre-ship. Awaiting peer review.

Open page

InstrumentPublic

TFB Telescope

A human-agency cockpit for an instrument that catches AI agents drifting toward failure. Every warning traces back to a receipt; nothing is claimed until parallax confirms it.

TFB use now: the seven-panel operator surface that sits between every detection loop and any public claim. Approvals, heal cards, and verified concepts route through here. Ships in calibration mode until an outside reviewer signs off.

Open cockpit

Browser agentPublic

TFBthumb

A continuous-perception browser agent under a hard agency ceiling. Kills the screenshot loop, settles by observation, and requires a single-use human token on every consequential action.

TFB use now: every web surface TFB ships is verified through TFBthumb instead of preview_eval or ad-hoc Playwright. The substrate IS the verification layer.

Open page

Extraction defenseCalibrating

TFB Drawbridge

The moat watch on the rig: the detector that reads the signature of a model being copied through its own answers — extraction and distillation pressure — and raises the bridge before the crossing completes. Rides as Skybridge's fourth Span.

TFB use now: a dependency-free Phase-0 engine validated on controls, gating on the one signal that separates genuine broad research from copying — label-divergence. It guards what David protects, so the platform's answers cannot be quietly walked off the rig. Awaiting a real-corpus fidelity run.

Phase 0 · internal

TFB Trust

The family whose only job is to make a claim honest: verification, proof, attribution, conscience, and memory. Nothing here asserts trust — each instrument earns it against a frozen artifact and leaves a receipt.

Frontline of TFB Trust: the beachhead — Clock, Divergence, Compass — and the offshore rig — David with Skybridge, Telescope, TFBthumb, and Drawbridge — both belong to this family. The cards below are the rest of the trust line: the conscience that refuses to call motion progress, the ledger that lands credit where the work was done, and the memory that carries every lesson forward.

ConscienceCalibrating

TFB Hamster

A lean Mission Conscience for multi-agent work. It watches the wheel and refuses to call motion progress unless it actually closes distance to the goal.

TFB use now: running live inside TFB on its own build cycles, monitored as an approved lane on the Withyou Operator OS. The conscience layer that sits under any agent and turns wounds into named guardrails so the next agent meets the lesson on the way in. Pilot review is a calendar gate, not a software one.

Open demo

AttributionConcept

TFB Credit

An attribution instrument for multi-agent work. Every action — model inference, human decision, agent tool call — carries a named receipt back to its source. Credit lands where the work was done, not where the claim was convenient.

TFB use now: the missing ledger for multi-agent attribution. Without Credit, compound work is claimed by the loudest output. With it, every link carries its own signature and the score settles from the receipts, not the narrative.

Spec · internal

MemoryConcept

TFB Echo

A reflection instrument that replays past agent decisions across sessions, surfaces recurring patterns, and lets prior lessons echo forward into current work. Agents forget across sessions — Echo remembers so the next agent meets the lesson on the way in.

TFB use now: the agent swarm generates wounds and heals across sessions. Echo closes the loop: every wound becomes a named guardrail the next agent encounters before it can repeat the mistake. Pattern memory without the storage burden.

Spec · internal

TFB Withyou OS

The agency-ceiling architecture the craft instruments ride on: bounded room for any model to lift the work, with the operator's authority, gates, and receipts kept stronger than model momentum.

ArchitectureConcept

withyou OS

A model-independent agency-ceiling architecture for AI agents. A Spine + Brain reference system that gives any model bounded room to help without overpowering the operator. Lives outside TFB as a separate codebase.

TFB use now: the architecture pattern that informs every TFB agent's bound on authority. Telescope, TFBthumb, and Great Design Cloud all inherit the same ceiling principle from this design language.

Local build · public landing pending

ArchitecturePublic

Instruments, not agents

A family-level product language: one governor, many instruments, same blood across writing, code, tokenization, design, and physical assist.

TFB use now: product doctrine for Agent Listen, Agent Swebench, Model Therapy, and Great Design modules. Keep the human as the melody; the system carries lift.

Open demo

ControlPublic

TFB Helm

A force-sharing idea: you lead, it lifts. The machine gets bounded authority and cannot overpower the person holding the line.

TFB use now: apply the agency ceiling to LLM tools: the model can assist, but gates, receipts, and operator authority remain stronger than model momentum.

Open demo Open twin

DesignPublic

Great Design ceiling

A module-level promise for Great Design Cloud: the designer keeps judgment while the system carries lift under a visible ceiling.

TFB use now: convert beta UX from maze-like setup into step-by-step guided completion with visible authority, file privacy, and module gates.

Open demo

CanvasPublic

TFB Ink

A drawing instrument that firms up live while showing how much the machine lent to the user and saving receipts of influence.

TFB use now: this is the clearest UI pattern for Model Therapy: show model influence, confidence, and correction debt instead of hiding it.

Open demo

Text systemsConcept

Tokenizer instrument

A deterministic base with bounded adaptive lift. The smart layer only moves where the base is uncertain, and every deviation carries provenance.

TFB use now: this maps directly onto R1 selector surgery: deterministic owner proof first, bounded model proposal second, receipt always.

Spec review first

WritingProxy required

Continuo

A writing instrument that offers inline futures while preserving the writer's melody. The current standalone demos include raw provider-key paths.

TFB use now: merge the concept into Ghostwriter and Agent Listen, but only after a server-side route protects keys and logs consent.

Held from live demo

Code editorInternal demo

TFB Code

Real Monaco editor with inline ghost-text futures: candidate completions execute in a sandboxed worker against local tests; survivors appear as inline ghost-text you cycle and accept. A flywheel learns the future you accepted per test-signature and offers it first on the next run.

TFB use now: proves the inline-futures interaction pattern with no raw browser credential lane. The current admitted demo uses deterministic local candidates and a worker survivor gate; model-backed futures still require the proxy and beta-account route.

Open demo

Design betaLive beta

Great Design Cloud

The live beta workspace where the Great Design ceiling becomes production: Sonnet critique, Gemini visuals, La Font typography canon, InDesign Assistant. Three days free, then $39/month.

TFB use now: Google login, file privacy, module gates intact. The beta is the proof that the ceiling concept ships as a working product, not just a doctrine.

Open beta

What helps TFB now

The prototypes are not just shelf pieces. These are the transferable heals worth pulling into production lanes.

1 · Authority

Agency ceiling for every model route

Give models bounded room to help, but keep operator gates, verifier proof, and route policy stronger than the model.

2 · Receipts

Show the system's hand

TFB Ink's influence readout should become a design pattern for Model Therapy and Great Design: show what changed and why.

3 · UX

One path, one next action

The Great Design ceiling belongs in beta onboarding: no random drop-offs, no blocked modules without a visible next step.

4 · R1

Deterministic base before adaptive lift

The tokenizer instrument is the right mental model for selector work: prove the base owner, then allow bounded model operations.

Held from public launch

These are valuable, but the public version needs a safe access point before a visitor can use model-powered features.

Continuo, TFB Code, and TFB Duet raw-key lanes

Current standalone files include direct browser input for `sk-` provider keys. Public launch requires a server-side proxy route, account gating, provider disclosure, and no raw key stored in the page.

Supabase sync in TFB Code v5

The public-safe path needs a defined beta account, tenant boundary proof, and customer data ownership receipts before it becomes more than a prototype note.