Teacher-context variants

Each variant is one way to build the teacher's refined trajectory r — a self-contained folder that overrides only what differs from the live default. The shared machinery (HTTP transport, MC rollouts, divergence metrics, the batched training loop) lives in the core and is never copied, so adding a variant never adds an if variant == ... branch.

Layout

agents/variants/
  base.py            TeacherVariant — the interface + shared defaults + t* helpers
  __init__.py        registry: load(name) / available()
  <variant>/
    utils.py         a TeacherVariant subclass (+ VARIANT = TheSubclass)
    prompts.yaml     strategy + maker_system_prompt (+ optional teacher_user wrapper)
    README.md        what it is, the bet, the pre-registered prediction

The interface (what a variant may override)

method	default	overridden by
`resolve_prefix(problem, solution, rollout, maker)`	maker rewrite, fall back to reference	P4 (answer-only, maker-free), P6 (captures diagnosis quote)
`accepts(prefix, reference)`	`is_acceptable_prefix` (len ≥ 80 ∧ boxed-match)	P2 / P4 (answer-only bypass: boxed-match only)
`build_ctx(tok, client, problem, prefix, cfg)`	user turn + `<think>` analysis + close + transition	P5 (reactive wrapper), P4 (no thinking)
`t_star(prefix, rollout, tok)`	`difflib` first non-matching block	P3 (`lcp`), P6 (quote-match)
`thinking`	`True`	P4 (`False`)

Everything else — score_topk, mc_branch, select_divergent, M1–M3, the trainer loss — is shared and calls these methods.

Two seams this relies on (both in `agents/maker.py::build_prefix`, backward-compatible)

prompts= — a variant supplies its own {strategy, maker_system_prompt, maker_user_template} instead of a file swap.
accept= — a variant supplies its own gate (e.g. answer-only bypass).

Usage

from agents.variants import load, available
v = load("p1_surgical")                 # available() -> ['p1_surgical', 'p2_...', ...]
prefix = v.resolve_prefix(problem, solution, rollout, maker_client)
ctx    = v.build_ctx(tok, client, problem, prefix, cfg)
tstar  = v.t_star(prefix, rollout, tok)

Scope

Variants are the prompt-space ranking substrate (probe / Stage-1). The loss-side mask and per-sample gate (the reactive-constructor spec) are deferred; when they land, build_ctx's declarations feed the trainer but generation stays shared.

Adding one

Copy the closest folder, edit prompts.yaml + the few overridden methods in utils.py, write the README.md. No core edits.