FBED.
    Models

    Five models, named for the founders of information theory.

    Each is a tool that agents call instead of reasoning from raw text. Three measure, one decides, and one models the buyer's world. Every model climbs the same ladder, and each step is gated by data, not by date.

    1. 01 Free
    2. 02 Logged
    3. 03 Drawn
    4. 04 Learned
    5. 05 Owned
    What they learn from

    Trained on behavioural tokens, not semantic text alone.

    Claude Shannon's theory solved the technical level of communication. Warren Weaver named two levels above it: whether a message carried its meaning, and whether it changed what the receiver did. Language models learn from text, which reaches the first two. Every FBED model is trained on behavioural tokens: what was said, what was done, and what the receiver did next. A sale is decided at the third level.

    Level A · Technical

    Was the signal sent accurately?

    Claude Shannon's theory, 1948.

    Level B · Semantic

    Did it carry the intended meaning?

    Where language models learn: semantic text.

    Level C · Effectiveness

    Did it change what the receiver did?

    Where FBED's models learn: behavioural tokens.

    The argument is set out in The Third Level.

    Benchmarks

    Each model names the benchmarks it must beat.

    The benchmarks below are being sought, not claimed. A result is published only once it passes on a sealed set. Before any result is used, every model is compared with four baselines:

    1. 01Taking no action
    2. 02A random allowed action
    3. 03The current way decisions are taken
    4. 04A frontier model with no further training
    Instrument

    Nyquist.

    Named after Harry Nyquist (1889–1976)

    Nyquist showed the minimum rate at which a signal must be sampled to be captured without loss. The model does the same for conversation: it captures what was said faithfully, at the smallest size that loses nothing.

    The conversation parser. A developer supplies a codebook; Nyquist parses any communication into it.

    parse(communication, codebook)
    Returns
    The communication as typed rows in the codebook's tables. Agents read codes, not transcripts.
    Gate before use
    The technical layer (speaker, turn, phrase, timestamp) must give the same answer every time on the same input. The judgement layer is measured against human readers and named comparators.
    Benchmarks sought
    • Agreement with expert readers, above the level the experts reach with each other.
    • Every named outside parser, scored on the same sealed set.
    • Codes that predict what the buyer does next better than the words alone.
    • Compression, speed and cost per thousand turns, reported beside accuracy.
    Instrument

    Huffman.

    Named after David Huffman (1925–1999)

    Huffman coding finds the shortest exact encoding of a message. The model encodes each action a seller takes, and finds the smallest change that moves a message to its target.

    The communication scorer. It scores communication at two levels, as signal and as behaviour, against a target, and returns the edits needed to reach it.

    score(message, target)
    Returns
    A signal score, a behavioural score and the minimal changes required. A verifier with edits, not an opinion.
    Gate before use
    Consistent scores on repeat, agreement with human graders, and a randomised test showing edited messages outperform unedited ones.
    Benchmarks sought
    • Suggested edits reach the target when the message is scored again.
    • Expert raters prefer the edited message at least as often as they agree with each other.
    • Edited messages outperform unedited ones in a randomised test.
    Instrument

    Viterbi.

    Named after Andrew Viterbi (b. 1935)

    The Viterbi algorithm recovers the most likely hidden path through a sequence. The model reads the timing of a commercial relationship, the hidden state it is in and the path it is most likely on.

    The timing and state model. It estimates when the next decision should be taken, including when the right action is to wait. Between calls, it estimates how changes in the buyer's real-world state alter which decision should be taken next.

    when(state)
    Returns
    The expected time to the next event, the estimated effect of real-world changes since the last contact, and whether acting now or waiting has the higher value.
    Gate before use
    Calibrated timing on held-out history before it informs any live decision.
    Benchmarks sought
    • Predicted timing against observed timing, day by day, on held-out history.
    • Ordering of who responds first, measured on held-out history.
    Policy

    Shannon Zero.

    Named after Claude Shannon (1916–2001)

    Zero is the first policy: the starting point from which every later version is measured. It learns only from decisions that were logged with their probabilities, and must beat the way decisions are taken today before it is used.

    The learned policy. It chooses the next action at each decision point, under a ceiling on harm and a budget per seat.

    act(state, goal)
    Returns
    The next action, with confidence that it improves on the current approach.
    Gate before use
    Its measured value must exceed the current approach with confidence, while harm stays below a fixed ceiling, in every period it runs.
    Benchmarks sought
    • Value above the current way decisions are taken, measured on logged decisions with confidence.
    • Harm below a fixed ceiling in every period.
    • Meetings held per 1,000 people entered.
    World model

    Shannon 1.

    Named after Claude Shannon (1916–2001)

    Shannon founded information theory and defined communication as a signal passing through noise. Shannon 1 models the buyer's world itself, so agents can rehearse inside it and plan through it.

    The commercial world model. It simulates the buyer's next turn and plans the next action toward a goal.

    step(state, action) · plan(state, goal)
    Returns
    The simulated next turn of the buyer's world, and the next action with its value.
    Gate before use
    Faithful to real outcomes across randomised arms, including buyers who do not buy, before planning through it is trusted.
    Benchmarks sought
    • Reproduces the differences between randomised arms seen in live work.
    • Reproduces buyers who do not buy, including silence after the first message.
    • A held-out test cannot tell its simulated turns from real ones.
    • Plans made inside it outperform both the reactive approach and the current approach on real records.
    • Outperforms published buyer simulators, which overstate how far non-buyers engage.

    Rules every model follows

    Predicted and observed are kept apart.

    A prediction never becomes a label and never counts as evidence.

    The label is the first observed outcome.

    Taken after the prediction was sealed, never before.

    Missing is never zero.

    Unread, censored and unascertained are recorded as such.

    One holdout allocation per lead, ever.

    The evaluation set never leaves FBED.

    Every gate must fail on a planted fault

    before its pass is believed.

    Readings stay with the customer.

    Models learn parameters, never a named person's readings.

    See the problems they solve, or read our research papers.