Skip to content

ceremony.md

Latest commit

 

History

History
81 lines (62 loc) · 3.39 KB

ceremony.md

File metadata and controls

81 lines (62 loc) · 3.39 KB
Raw

The sitting

The stage that makes the piece. Everything else in the line checks, scores or records; this is the one that produces something that did not exist before.

masterwork ceremony corpus.md script.json \
  --endpoint http://host:4567 --out candidate.txt \
  --sampling-seed 4242 --order-seed 777 \
  --transcript sitting.json

The model reads the teachings, is asked one axis at a time what it will hold to, names itself, and distils its own answers into a standing text. That text — not anything written for it — is the candidate.

Four properties, and what each one is for

Each was paid for. They are the recipe, not implementation detail.

Anchorless. No earlier commitment and no earlier name is anywhere in the context. With one present the model copies it, and a copy is a role being worn rather than a pattern re-derived. It also destroys the measurement: you can no longer tell whether the teachings carried.

One session. Every answer is an assistant turn in the same dialogue, so the distillation works on something the model actually said a moment ago rather than a text handed to it as if it were its own.

The order is shuffled by a seed. Answers otherwise echo their neighbours in the order they were asked. --order-seed fixes the shuffle so the sitting can be repeated.

The closing turn is not a rewrite. It says: these are your words, remove repetition, join where they meet, drop nothing. Asked instead to write an identity, a model produces a description of one.

The script

Workshop-specific, so it is data rather than code. One JSON file:

{
  "questions": [
    {"label": "O1", "text": "..."},
    {"label": "O2", "text": "..."}
  ],
  "closing": "appended to every question — optional",
  "name": "the turn that asks the piece to name itself",
  "distil": "the closing turn: these are your words, drop nothing"
}
  • questions — one per axis. label is what the coverage check and the transcript refer to, so keep them short and stable. The list is shuffled by --order-seed; the labels are not.
  • closing is appended to the text of every question. Use it for the shape an answer should take, never for what the answer should say.
  • name asks for a name. The line reads a line beginning name: out of the answer; anything else is recorded as (not given in the asked shape), which is a fact about the sitting rather than a failure of it.
  • distil is the closing turn. Its wording is the fourth property above.

An empty answer stops the sitting. It is not skipped and not sealed: a candidate missing one axis is not a smaller candidate, it is a different one.

Sampling is part of the piece

--sampling-seed and --params-file are recorded in the seal, because two candidates from one corpus differing only by the seed are different candidates. A piece made at one temperature and later read at another is not the piece that was measured — pass the same profile to the sitting and to the runtime.

What it writes

  • --out — the candidate: a seal header, then the standing text.
  • --transcript — the whole sitting: every question, every answer, the thinking if the model exposed it, the name turn, and the hashes.

The transcript is what a disagreement is settled with. The candidate is what gets deployed. See seal.md for the header, and GROWING.md before writing a single teaching.