Heedle ClothText

A loom for steering a small model by hand. The cloth is what the model does at each point of a square. Press treadles to change it, beat a pick in, and read what was checked.

A loom for steering a small language model by hand. Each pick weaves a line. Press treadles to change what comes next, beat a pick in, and read what was checked on every context the model has.

Cloth7
Shafts

What this page is

Heedle is a workbench for steering a small model by hand. A model is a set of numbers that turns inputs into outputs. Here you change those numbers a little at a time, and the page checks each change before it keeps it.

There are two models to steer. Cloth turns a point on a square into a colour, so you watch a picture change. Text reads three characters and guesses the next one, so you watch lines of text change. Both are tiny on purpose: the cloth has 467 numbers, the text model 4,700. They are toys built for this instrument, not a real chatbot.

Try it in three steps

  1. Press a treadle. A treadle is a named change, such as tilting the cloth. The page shows what the change would do before anything moves.
  2. Beat the pick in. The page checks the move: the numbers must stay near where they started, and no input may change by more than the limit. If a check fails, the pick snaps and nothing moves.
  3. Click an input to pin it. Pinned inputs are held still while the rest changes. Click a pick to go back to it.

The words on the page are weaving words. A shaft is a direction the model changes most. A heddle is a pinned input. A pick is one change that passed the checks. A snap is one the checks refused. The full list of what is checked is in the draft specification.

Threading shafts: the directions in which this model changes most

Tie-up shaft × treadle

Drawdown

Left selvedge distance from baseline, of the gate's radius Right selvedge predicted worst change, of limit Click the cloth to raise a heddle: that point must not move Click a context in the pattern to raise a heddle: what follows it must not change

Treadling the pick you are about to make

The percentage under a treadle is how much of its named change gets through the shafts and the heddles.

Move length
0
Proved safe to
0
Predicted change
0
At the worst probe
0
Tension
0.00 rest
Picks left
24
Heddles
0

Picks click one to unpick back to it

Nothing woven yet.

Futures woven ahead of you

Lamp what was checked after the last beat

Nothing has been beaten in yet.

How to read it

  • Cloth: a network of 467 weights turns each point of a square into a colour. Every pixel is an input. The picture samples the square, so what happens between pixels is not on screen.
  • Text: a network of 4,700 weights reads three characters and scores the next one. It was trained on Heedle's own documents. Each pick weaves a line from the new weights, on the same random draws as before, so a change in the text is a change in the model.
  • Pattern: the model's guess after each of its twelve commonest contexts. A larger square is a likelier character. Green rose, red fell.
  • Shafts: the directions in weight space that change the outputs most per step, on a fixed set of probe inputs. To first order, one unit of travel changes the outputs there by one unit, whichever shaft you move.
  • Treadles: named changes, each tied to the shafts. The percentage is the share of the named change these shafts can carry, to first order, on the probes. A low share means these shafts cannot carry it. It does not mean the model cannot.
  • Heddles: inputs you pin. The loom is threaded around them, so pinned outputs do not shift to first order.

What a beat checks

  • Weft's own gate: the weights stay within a fixed distance of the baseline and the budget covers the pick.
  • A claim: no point of the cloth changes by more than the limit since the last fell. A search looks for one point that breaks it, and it alone decides. If it finds one, the pick snaps and the cloth stays. Finding none is a sampled reading.
  • A claim: no context changes by more than the limit since the last fell. Every one of the 21,952 contexts is read, and that alone decides. If one breaks the claim, the pick snaps and the model stays. Finding none is complete, up to rounding.
  • The claim is made pick by pick. Nothing here limits how far several picks together take the model from where it began, except the gate.
  • Fresh probes, drawn from the new weights' hash after the pick is fixed, measure what the pick did. They do not decide it.
  • The closer a pick comes to its limits, the longer you must hold to make it.

The record

Every pick, snap and unpick is chained by hash into a draft. A beat the gate refuses records nothing. Anyone can replay a draft without this page: python3 -m heedle verify draft.json