Joint multi-slot evolution — search the whole decision plane

Module: agentdescent.joint_meta · Issue: #388 (P7)

Every other rung of the meta stack evolves one slot: meta_run / continuous_meta_run fix a slot and search its value, meta_population_run runs several rules of one slot as islands, and portfolio_report only moves budget between slots after the fact. joint_meta_run searches several slots together: the artifact is a bundle {slot: value}, each outer rollout installs all the slots and runs the inner search, and the reward is the inner held-out curve. An effect that only appears when two rules change together — invisible to two independently evolved rules — is visible here.

from agentdescent import (SourceSlot, joint_meta_run, joint_problem,
                          priority_selection)

specs = {
    "selection": priority_selection(),
    "task_sampler": SourceSlot(initial_value=task_sampler_seed),
}
train = {"algotune": joint_problem(tasks, reward, rounds=4, n_workers=2)}
val = {"algotune-val": joint_problem(val_tasks, reward, rounds=4, n_workers=2)}

report, result = joint_meta_run(
    train, val, slots=list(specs), specs=specs,
    propose=offline_propose, seeds=range(4), val_seeds=range(4, 8),
    isolate=True)
print(report.verdict, report.joint_gain)     # e.g. adopt, +0.4
print(report.slot_gains)                     # per-slot marginal held-out gain
print(report.isolated)                       # slots rolled back as untrusted

The artifact: a bundle

PoliciesBundle is both a Strategy (for the outer evolve()) and a spec-like object (for validation and the reflector). The state is {slot: rendered value} and the key space is the slot names, so the aggregator's existing rules apply with no new machinery:

• a proposal names one or more slots as === <slot> === blocks;
• edits to different slots union-merge (a coherent multi-slot change);
• two edits to the same slot contradict and are resolved on held-out score;
• a bare proposal (no header) is accepted only when exactly one slot's gate compiles it — otherwise it is ambiguous and refused.

Isolating untrusted rules

joint_meta_validate attributes the joint change per slot: each slot's marginal held-out gain is its edit against the seed bundle with every other slot held at the seed. A slot whose marginal gain is below min_gain is untrusted — the joint gain may be positive on the others' backs, but this slot did not pay for itself on held-out problems.

• isolate=True (default): the untrusted slot is rolled back to its seed value and the corrected bundle re-validated — adopted only if the corrected gain still clears min_gain.
• isolate=False: any untrusted slot rejects the whole bundle (fail-closed).

The archive is unchanged

An adopted bundle writes one ordinary meta-rule record per slot (with its own problem_descriptor), so _archived_rules, exam_archive and continuous_meta_run keep seeding and examining it with no special case. Pass store= and the next run seeds its bundle from the archive's best per-slot rules instead of each spec's raw initial value.

Isolation contract

As with meta_run, val_problems and val_seeds must be isolated from the training set: overlapping problem names and overlapping seed sets both raise.