# Pre-registration — WP31 PRISME pilot (DRAFT v0.1, 2026-10-29; to be FROZEN before any `test_*` split is read)

Status: DRAFT. No file named `test*.jsonl` / `test.bin` has been opened by any evaluation script so far
(`protocol/locked_test_manifest.json` records the split hashes and the unsealing timestamp once frozen).

## 1. Primary benchmark suite
* **Synthetic compositional suite v1** (`data/splits/synthetic_v1`, generator `data/generators/synthetic.py`,
  manifest sha256 in `manifest.json`): families arith, bind, fsm, excl, sched (train/dev) + compose (held out).
  Task unit = one item (prompt + answer). Exclusion: items whose full text duplicates an earlier split (dropped at
  generation time; counts in manifest). Regimes: `*_iid` (train templates & depths, new seed), `*_template`
  (held-out paraphrase templates 3–4), `*_depth` (depths 5–6 > train max 4), `*_compose` (held-out family).
* **General language**: FineWeb-Edu held-out `test.bin` (2.88M tokens, locked; `val.bin` is the development set).
* Independent benchmarks (GSM8K, EvalPlus) are OUT of the primary scope of the 103M pilot (a ~200M-token
  backbone cannot solve them above chance); they are reported only if a run is executed, as secondary.

## 2. Primary metrics
* Quality (reasoning): exact-match accuracy of free greedy generation on the synthetic test regimes
  (pooled `test_iid+test_template+test_depth`, each regime also reported). Quality (language): held-out
  cross-entropy on `test.bin` at each budget.
* Cost: (primary) measured median end-to-end decode latency per token at batch 1 on the reference device, with
  the backbone-only latency as the unit; (secondary) analytic ops/token (`prisme/flops.py`).
* Joint superiority: a configuration dominates if quality is higher at ≤ the comparator's latency, or latency is
  lower at non-inferior quality (paired CI excludes zero).

## 3. Baselines and tuning budgets
Same backbone checkpoint, same data exposure, same mixed-budget training recipe and step count:
(a) frozen backbone (budget 0); (b) parameter-matched plain decoder trained from scratch with the backbone recipe;
(c) dense 16-slot editor; (d) PRISME editor with random / uniform / fixed-pair routing at budgets 1, 2, 4, 8;
(e) entropy and norm-change routers (thresholds tuned on calibration data); (f) global-recurrence adapter
(Huginn/ELR-inspired, parameter-matched); (g) learned value controller (λ, κ tuned on calibration rollouts only);
(h) exhaustive one-step oracle and two-step oracle (diagnostic only, never plotted as attainable).
Hyperparameter budget per adapter: one learning rate (1e-3 for adapters, chosen in Phase 4) and the shared
schedule; no per-baseline tuning beyond the router thresholds.

## 4. Minimum meaningful effect
* Reasoning: +3 percentage points pooled exact match at comparable latency (≤ +10 % of the comparator's), or
  −20 % median latency at non-inferior accuracy (lower CI bound > −1 pp).
* Language: zero-budget cross-entropy within 1 % (relative) of the frozen backbone (H5 preservation).
* Routing (H2/H3): learned controller loss at matched mean extra-ops within the CI of the best fixed/random
  schedule is a NULL for H3.

## 5. Seeds
Editor/controller stage: 5 independent seeds (1–5) at the Phase 4 scale; at 103M as many as the cluster
permits within the session (target 3–5, actual count reported). The backbone is a single seed (declared limitation).

## 6. Uncertainty
Paired bootstrap over task items (10,000 resamples) for accuracy differences within a seed; across seeds,
mean ± SD and a t-interval over seed-level differences. Tokens of a single run are never treated as
independent replications of training.

## 7. Stopping / resource rules
Pilot compute cap: ≤ 2 nodes × 8 h for Stage A, ≤ 6 nodes × 6 h for stages B–D; if a stage cannot complete,
the phase is reported as incomplete and downstream claims are withheld.

## 8. Deviations
Any change after freezing is recorded in `protocol/deviations.md` with justification and timestamp.