role-set

role-set is a local generator for role-shaping items. It plans a batch, asks a model for the parts a writer should write, and records which gates each item passed. For kind B, it can also build the puzzle in code and prove it before a model words it.

Prototype, developed in the open. Advisory and experimental: interfaces and data shapes change without notice. Version 0.2.0 adds the T10 generator’s cell map, if_shape axis, allocator and record gate. Nothing role-set writes is training data until a person checks it. Every item goes through R&D’s key-check and a second read before it nears a training set.

The program is not a judge. A role pack, which people write, says what a good role-holder is. role-set serves that pack in slices and runs structural checks. A refusal is a failed check, or a known-answer mismatch on verdict and evidence. The critic agreeing is not a pass.

What a session is

A session is a directory of files, not a server. You name the plan, the pack, the output directory, and, when you use the store, the database file. write --mock does the rest of the pipeline with stand-ins, so CI never calls a model.

A real writer or critic has to be http://127.0.0.1 with an explicit port. The call is refused before anything is sent if the host is anything else. role-set does not rent a machine and does not talk to a pod. offrig is the program that rents one, when a run needs one. role-set only sees the tunnel that lands on the loopback.

Two packs, one generator

• The verifier pack is synced from aspire-si at a commit and is the pack every run so far has used.
• The Advocate pack is a pack a person wrote, approved by the Director on 2026-10-14 on pull request #56, where batch 1 is being built. It is not on main until #56 merges.
• The T10 generator builds puzzles on the strict size words, proves each one with the solver and R&D’s four readings, spreads them over structural cells, and holds R&D’s variety floors through a record that CI checks on every pull request.

Where the numbers come from

Three paid pilots are written up under docs/runs/ in the repository. Pilot 1 cost $3.52, and none of its 104 accepted rows is data. Pilot 2 cost $3.61, and R&D’s key-check of 49 candidates was 15 pass, 15 in need of a mechanical fix, and 19 fail. Pilot 3 cost $2.80 and wrote nothing, because the writer’s engine hung after startup. Those three notes are the only cost record this handbook uses. The cost command is an estimate for a plan you have not launched.

The generator’s figures (cells, distinct skeletons, headroom) come from the committed tables and record in src/role_set/generate/, and the scripts that make them.

Read next

• Getting started installs the tool and runs a plan, a cost estimate, and a mock write.
• Usage is the real runner, pilot-auto, and readout.
• Kinds and gates is A through E, and the limits of the B solver.
• The T10 generator is cells, if_shape, the allocator, the floors, and the record gate.
• Role packs is the verifier and the Advocate, and how a pack is hashed and verified.
• Configuration is writers.json, the pins, and the grant.
• Store and firewall is ingest, voids, and what is refused at the door.
• Security is what a run touches.

The designs are docs/role-set-design.md and docs/generator-design.md in the repository, kept in line with the program.