Classification: layer=E (verification) + H (weak-reader); mechanism=scaffold H.

Intent: Ship a small runnable model of a subsystem's behavior the reader can EXECUTE to understand or predict — offloading the execution-prediction that weak models fail at, without running the whole system. Go's test culture already lives here: small interfaces make hand-rolled in-memory fakes one-screen literals, and `httptest` is a stdlib network simulator.

Also Known As: reference implementation; in-memory fake; executable spec; oracle model; test double; `httptest` server; steppable model.

Applicability / Recognition: Apply when — a subsystem has non-obvious dynamics (a reconcile loop, a state machine, a retry/backoff protocol); understanding requires mentally simulating execution; an external dependency (HTTP, a store, a queue) must be reasoned about offline. *Detector seed:* a subsystem whose behavior is documented in prose-describing-execution, with no runnable model or fake → recognition fires (execution-prediction is weak models' weakest point — DR2-019, CRUXEval ~63% even for strong models).

Motivation: A weak agent must modify the reconciler's convergence loop (diff → actions → apply → re-diff). It cannot mentally simulate whether a partial apply converges or oscillates. A steppable in-memory world — `sim.World` with `Step()` returning the applied actions and the next state — replaces mental simulation with execution: feed a desired/actual pair, watch convergence, print the trace. The EsoLang library shipped exactly this idea (a local simulator) and it carried the weak-agent gain.

Structure & Participants: *Reference model* (runnable, small, steppable — `Step()`/`State()` inspection surface) · *In-memory fake* (a literal implementation of the seam's narrow interface — Go's native double) · *`httptest` server* (the stdlib simulator for HTTP boundaries) · *Conformance test* (model vs production agree on representative inputs).

Collaborations: Provides the comparator for Class D oracles (the model IS the expected-behavior source); backs Class C contracts; pairs with Class G (the model's usage is Example-demonstrated). Capability injection (§2) is what makes fakes drop-in — a cell taking `seams.Store` accepts the ten-line map-backed fake with no mocking framework.

Goals / Non-Goals: *Goals:* make non-obvious dynamics executable, not just described. *Non-Goals:* NOT a second production implementation (a reference model, kept simple); NOT for trivially-obvious subsystems; NOT reflection-based mock generation (gomock-class module-graph interception is the §7 posture — literal fakes are cheaper and honest).

Consequences: (+) the reader runs instead of simulates; (+) doubles as a Class D comparator and the test fixture; (+) zero third-party cost — interfaces + httptest are stdlib culture. (−) a model is code to keep in sync — conformance-test it against production; (−) over-modeling wastes effort — only non-obvious dynamics.

Alternatives: prose describing behavior (weak readers can't execute prose); reading the production code directly (the thing too complex to simulate). The model is the offload.

Risks & Assumptions: assumes the subsystem's behavior is modelable simply; a model that drifts from production misleads — conformance-test it. *Sunset:* if the production code becomes simple enough to read directly, the model retires.

Evidence & Transfer-strength: R2C-008 (simulator in the transformative library, benchmark), DR2-019 (execution-prediction weakness, benchmark). Class: benchmark. Tag: **[E-strong]**.