REAP-LOOP

Status: work-in-progress research. Numbers here are PRELIMINARY and several parts are unvalidated. This repository is a research artifact, not a product. Read the caveats section before quoting any figure.

REAP-LOOP is a reversible gate-bias that live-calibrates the expert working-set of a Mixture-of-Experts (MoE) model during the session itself.

Instead of pruning a MoE offline for a fixed domain, REAP-LOOP:

1. Starts at the full model (K0 — no experts masked), so the very first tokens run at full quality.
2. Watches which experts the running session actually uses (via the router's own signals — no extra model, no separate predictor).
3. Progressively narrows the resident working-set toward an aggressive keep-level (e.g. keep-9%), biasing the gate away from experts the session is not using.
4. Reverses when it must: the bias is a mask over the gate, not surgery on the weights. Expert IDs are preserved, so the model can widen back out (a "breath") the moment the router shows drift.

Because the working-set is learned from the live session, REAP-LOOP adapts to any task. It is not domain-pretrained and it does not require a fine-tune. A catalogue of pre-learned working-sets (per output-type) is an optional pre-warm — a way to skip the warm-up cost on a familiar workload — never a prerequisite.

The direction this points at is PACE — a Perceptive Adaptive Control Engine: a self-calibrating controller inside the inference engine: keep-level = f(routing-drift). Drift low → tighten; drift rising → widen proportionally. Dynamic MoE sparsification framed as a control problem, with the router itself as the drift sensor.

---

Built on ds4 / DwarfStar (credit in evidence)

REAP-LOOP is built on top of ds4 / DwarfStar by Salvatore Sanfilippo (antirez) — the MoE-streaming inference engine is the foundation this work stands on. ds4/DwarfStar is MIT-licensed; all credit for the underlying engine, the SSD/host-side expert streaming, and the CUDA/Metal MoE paths belongs to it.

Everything in patches/ is a set of diffs against ds4 — this project does not reimplement the engine, it extends it. A verbatim copy of the upstream README and the relevant paper are kept under docs/references/ for attribution and context.

---

Companion mechanisms

• SPEX (Speculative Prefetch of EXperts) — predicts which experts to page in next, hiding expert-load latency behind compute. SPEX predicts the loading, never the gating: the real router always decides, so accuracy is preserved by construction. See docs/SPEX_spec.md, docs/SPEX_LOOP.md, and src/msc/spex/.
• WRAP — Working-set Resident Aggregate Prefetch (bulk page-in) — a host-side bulk expert page-in path that batches the transfer of a working-set into resident memory in one shot, rather than fetching expert-by-expert on demand.
• Together with REAP-LOOP these are three composable levers: warm + cache + REAP-loop (+ SPEX). Because REAP-LOOP acts in expert-ID space (a gate bias, not surgery), the SPEX/prefetch metadata stays valid across mask changes — the levers compose without re-dumping.

---

Repository layout

---

Caveats (read this)

• The numbers are PRELIMINARY. Much of the measured speed-up is confounded by cold-vs-warm cache effects and has not been isolated with a clean A/B/A protocol. Treat every t/s and every accuracy-drop figure as indicative, not final.
• Many parts are WIP and unvalidated. PACE, the self-calibrating controller (drift-driven keep-level), is a research direction, not a shipped algorithm; the fixed keep-level schedules in the ledger are the datapoints that would map the response curve, not the final controller.
• The prior-art analysis is deliberately conservative. No single building block is claimed as new. What is defensible is the reversible closed-loop composition (mask actuator, zero-violation), a paired rand/reap contrast (1.345x, CI[1.270, 1.423], same GPU, confound-clean), and static keep-23 as a speed diagnostic (11-17 t/s on a pod 3090, not generalized). The previously claimed "causal asymmetry in pruning" is RETRACTED — it did not replicate under multiseed (N=3) and was a mask-inert + n=1 artifact. For the current, canonical status of every claim, see docs/CLAIMS_CURRENT.md (single source of truth), plus docs/REAP_LOOP_NOVELTY.md and docs/PRIOR_ART.md.

License

MIT — see LICENSE. Copyright 2026 REAP-LOOP contributors. The upstream ds4/DwarfStar engine that this work extends is MIT-licensed by its author; see docs/references/ for the upstream material.