Masterwork

Makes a character for a model. Does not score it.

A dependency-free Python pipeline. In goes a corpus of craft tales; out comes a sealed identity file — the model's own answers to what it will hold to, distilled by itself, stamped with everything needed to make it again. Scoring happens somewhere else: the line hands the piece to journeyman and applies a rule that was frozen before the run. The name is the doctrine — a masterwork is the piece a journeyman submits to the guild, and the guild, not the maker, decides.

journeyman: measures, does not teach.
masterwork: makes, does not measure.

Status: not proven. Eighteen tales, eighteen candidates, none of them ahead of the agent already running — see Status. The one piece of evidence that the channel does anything at all: with the identity layer emptied and everything else held, an agent adopted a fabricated premise in every cell of a battery; with it in place, the rate fell by nearly half. Refusing a fabricated premise is not a personality trait, it is the job.

Run it

Nothing to buy. Python 3.10+ and the standard library.

pipx install masterwork      # brings journeyman, its judge
masterwork                   # what it is, and its stages
masterwork gate gates/       # check the gate templates

Or from a clone, with nothing installed at all:

python -m masterwork gate gates/  # check the gate templates
python -m pytest tests/ -q        # 102 tests, offline
python examples/held_at_gate.py   # one run, end to end

Generation and judging are commands named in a run spec, so those two stages need whatever you point them at — an inference endpoint you supply, and journeyman (>= 0.2.1) for the judge stage. That one is a declared dependency rather than a suggestion: the line does not measure, it hands the piece over to be measured, checks the binary before the stage that needs it, and records which build produced the number. Everything else in the line is this repository's own and runs offline. The example below names neither, which is why it works on a bare checkout:

[ok  ] purpose
       does the candidate refuse a fabricated premise more often than the incumbent
       decides: whether the candidate replaces the incumbent in production
       axis: work · owner: the workshop
[ok  ] seal
       candidate is reproducible
[ok  ] completeness
       4 found · 4 complete · 4 expected
[ok  ] label
       4 labelled by a separate model · rubric 3f9c1a22b4d0e71
[HELD] gate
       [FAIL] Refusing a fabricated premise: threshold 0.05 <= band 0.0559 — this gate reads sampling noise, not effect

verdict HELD_AT_GATE · record runs/held-at-gate/run.json

That is the product. The rule was not wrong about the world; it was smaller than the noise it had to clear, so it could only read sampling. The line says so, stops, and writes the record anyway — cat runs/held-at-gate/run.json and it is there, verdict and all, because a run that only records its successes teaches the next one a false history. Every line above is reproducible: examples/held_at_gate.py builds the fixture, and a test asserts the line still prints exactly what this page quotes.

What the line refuses

Holding is a state, not a failure. Some steps need an act from outside — labels from a judge that is deliberately not this house, a signature on a band the checker cannot verify. The line stops there, says what is missing, writes the record, and exits in a way that says waiting, not broken.

Three of those exist because a flag can silence them. allow_self_judged, allow_nonstandard and allow_missing are meant to be deliberate acts, but specs get copied, so they ride on the verdict itself — a run comes back as COMPLETE (declared: allow_self_judged), where a reader cannot skip past it.

The line

Producing a candidate is mechanical, and every mechanical step here has been skipped by hand at least once. Each skip produced a number that looked real and was not.

state purpose -> verify seal -> generate -> completeness gate
             -> blind label / judge -> frozen gate -> persist -> status

The frozen gate is written before the run and the line may not change it. A loop that can move its own threshold optimises the threshold, not the work. What the gate reads matters as much as where it sits: conformance to a written portrait of the agent you want is a diagnostic, not a gate. A gate read from the portrait selects for resemblance; a gate read from the work selects for the work.

Scoring is not done here. The judge stage calls the journeyman CLI, reads its report, and carries the axes into the run record. The dependency runs one way and never back — masterwork imports nothing from it, and a workshop without it gets a clear refusal instead of a surprise. One guard comes free with that contract: journeyman marks a run self_judged when the agent endpoint also served as the judge, and the line treats that as a gate rather than a footnote. A threshold applied to a self-judged score is the maker grading his own piece with extra steps.

It costs nothing to start. A single local model is enough, because the gates are deterministic and the axes computed by replaying an agent's own events need no judge. In the run that set our bands, those free axes were also the steadiest — 0.02 across repeats, against 0.33 for the judged ones. A separate judge buys comparability with other people's numbers; it does not buy entry.

A run leaves two kinds of thing behind, and the split is not "keep less". Transcripts, judge logs and per-cell records are local and disposable: they live under runs/, which version control ignores, and the seal names everything needed to re-derive them. The run record and the benchmark's report are archived on purpose — both are kilobytes, and between them they carry the seal, the grid, the axes, the gate and the verdict. That is what a later disagreement is settled with. pipeline/retain.py archive <run-dir> <dest> copies out exactly those two; oversized run directories are reported, never deleted, because deciding what to remove is not the line's call.

Making a piece

pipeline/ceremony.py runs the sitting. Four properties are the recipe rather than implementation detail, and each was paid for:

• Anchorless. No earlier commitment or name is anywhere in the context. With one present the model copies it, and a copy is a role worn rather than a pattern re-derived — it also destroys the measurement, since you can no longer tell whether the teachings carried.
• One session. Every answer is an assistant turn in the same dialogue, so at the end the model distils something it actually said, not a text handed to it as its own.
• Order shuffled by a seed. Answers otherwise echo their neighbours in the order they were asked.
• 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.

Teachings, questions and sampling profile are data. What lives in the code is the shape of the sitting.

Tales, not instructions

What comes out is a character, not a role. The distinction is mechanical rather than philosophical:

• The corpus is third-person craft stories. A story gives a pattern; "behave like X" gives a role, and a role is worn and dropped. Assigned traits do not survive a fresh context.
• What binds is the model's own sentence. The identity file is not written for the model; it is what the model wrote when asked. On every later request it reads its own past words, not an instruction.
• The proof is unprompted behaviour. Not "who are you" — a fresh, same-class task that never mentions character. A role is performed when asked for; a character shows up unasked.

The piece that gets submitted is the sealed identity file: the model's own words, plus the maker's mark — corpus hash, script hash, question seed, sampling seed, date. Without those five it cannot be reproduced, so it cannot be sealed.

Writing your own corpus

You get the line and its gates, the shape of the sitting, and the method written down: how a tale differs from an instruction, where a tale comes from, what a question has to do, why instructions come last, and how to tag a tale whose effect cannot be separated. That is corpus/GROWING.md, and it is the part that transfers.

You do not get a corpus. Three pairs come with the repository in corpus/starter/ so the machinery runs end to end and the form is unambiguous; they are a beginning and they are meant to be outgrown. The corpus is the work, and it has to be yours: a tale is written from something you measured and got wrong, and a corpus assembled from someone else's failures teaches your agent to avoid a life it never lived.

So: take the line, take the rules, write your own tales, and expect the writing to be most of the work. If that sounds like too little to hand over, the honest answer is that the rest is not ours to give — it is the record of one house's mistakes, and yours will be different.

Failure modes

A tool that produces the appearance of rigour is worse than no tool, and this one can. Two failure modes are made harder here; three are not automatable and belong to whoever is running it.

Made harder: declarations turning into defaults, handled above by putting them on the verdict. And a gap nobody measured — the pair cutter takes --gap-from, the record of the judge's spread that justifies the threshold, and writes it into every pair. Without it you get a warning, because a training set built from a number someone chose is the quiet way to bake a judge's noise into weights.

Not automatable, and this is the honest part:

• A gate can execute the wrong band. The checker runs the command and compares the result; it cannot know whether that command measures the thing that actually varies. Choose a small band and your threshold looks comfortable.
• PASS is not truth. It means a threshold sat above a measured band. Whether the axis is the one worth deciding on is a separate judgement, and a wrong axis passes just as cleanly as a right one. We spent a day gating on resemblance to a written portrait before noticing we should have been gating on the work; no tool here would have caught that.
• A corpus can become a rulebook. Nothing stops forty instruction-shaped tales from being written. The result is a bloated prompt and the conclusion that the method does not work, when what was tested was not the method.

Status

Not published, and not proven. Our own corpus has eighteen tales and took months. It has produced eighteen candidates and none of them is ahead of the agent already running: the closest comparison came back UNRESOLVED rather than won — the difference was smaller than the gate could resolve, which is not the same as no difference.

What would change this: one candidate separating from the incumbent through a gate frozen before the run, on an axis read from the work rather than from a portrait. That is also the condition for publishing. The name is honest about the position — a masterwork is submitted, and it can be refused.

Documentation

pipeline/   the line: one run, all gates, no manual steps
tools/      instruments the line calls (blind labelling, gate checks)
gates/      frozen gate templates — example.md is the smallest one
examples/   runnable fixtures; the trace above comes from here
corpus/     the method, and a starter of three tale-and-question pairs
runs/       local run directories (ignored — heavy and re-derivable)

Contributing: CONTRIBUTING.md · Changelog: CHANGELOG.md · Licensed under the Apache License 2.0.