mixle.experimental.graduation module

Graduation bookkeeping for mechanisms living in mixle.experimental.

Policy (see mixle/experimental/README.md): a mechanism graduates out of experimental/ into the stable package when it (a) beats the E1 baseline on the E7 long-context referee suite at matched FLOPs, and (b) has misfit/truncation receipts – measured error-characterization artifacts, not just “it works”.

Neither the E1 baseline nor the E7 referee suite exists yet at the time this scaffold was written; this module only encodes the shape of the bookkeeping those later items will populate and check against. Nothing here enforces the rule today – ExperimentalMechanism.is_eligible() is pure bookkeeping over whatever receipts a mechanism happens to have attached.

class ExperimentalMechanism(name, graduated=False, baseline_receipt=None, misfit_receipt=None)[source]

Bases: object

One entry in the experimental-mechanism graduation ledger.

  • name: short identifier, e.g. "chunked_recurrent_spine" (E1) or "sketch_state_attention" (E3).

  • graduated: whether this mechanism has already been promoted into the stable package.

  • baseline_receipt: the matched-FLOPs comparison against the E1 baseline on the E7 suite, once both exist – e.g. {"metric": "bpb", "mechanism": 1.02, "baseline": 1.05, "flops": 3.1e20}. None until measured.

  • misfit_receipt: honest error-characterization artifacts for the mechanism’s state structure – e.g. sketch collision rate, tree truncation error, moment-closure residual. None until measured.

Parameters:

  • name (str)

  • graduated (bool)

  • baseline_receipt (dict | None)

  • misfit_receipt (dict | None)

is_eligible()[source]

Whether both receipts required by the graduation rule are present.

This does not check graduated – it answers “could this graduate”, not “has it”.

Return type:

bool

REGISTRY = _GraduationRegistry(_mechanisms={})

The process-wide graduation ledger. Track-E items register their ExperimentalMechanism here.