1# WP30 — Research questions and falsifiable hypotheses (registered before any held-out outcome was examined)23Registered: 2026-10-29 (Phase 1). Amendments only as dated entries at the bottom.4Test-split outcomes had **not** been examined at registration (no test split existed yet; only the E0 smoke run on a5hand-written 3-sentence memo had been executed, see `runs/logs/`).67## Research questions89- **RQ1 (compression frontier)** How do text shortening (A2–A4, A8), packed KV quantization (B2), token eviction (B3–B6) and10 learned compaction (B7, if run) trade task quality against *measured* serialized bytes and resident cache bytes?11- **RQ2 (transfer compatibility)** How much performance is preserved by same-checkpoint reuse (B1, cross-process, cross-node)12 versus learned mappings between distinct checkpoints (Track C), relative to the receiver's own full-context answer?13- **RQ3 (resource accounting)** Do rankings of methods under a token budget change when the budget is bytes, end-to-end latency,14 peak memory or amortized cost over N reuses?15- **RQ4 (repeated handoffs)** How do correct-fact retention, unsupported additions and task accuracy evolve after 1, 2, 4, 8 handoffs,16 for prose summaries vs structured evidence, in closed vs repairable chains?17- **RQ5 (evidence)** Under an equal byte budget, does structured evidence with source spans (A5) retain exact facts (identifiers,18 numbers, dates, negations) better than an abstractive summary (A4)? Does metered retrieval (A6) repair losses, and at what cost?19- **RQ6 (locality)** When is running the next query where the cache resides cheaper than moving the cache or recomputing from text,20 as a function of context length, payload bytes, measured bandwidth and queue delay?21- **RQ7 (adaptive policy)** Can a router fitted on development data improve the held-out quality/cost frontier over the best22 development-selected static method, once routing overhead and fallbacks are counted?2324## Hypotheses (each may be rejected)2526- **H1** Lossy compression degrades exact identifiers, decimal numbers, negation and temporal order more than paraphrasable content:27 on the synthetic private-fact family, the accuracy drop from A1 to A4 at a 512-token budget is larger (≥ 10 percentage points)28 for `exact_copy`, `negation`, and `revision` question types than for `single_fact` questions.29- **H2** Packed KV quantization is model-dependent: at 4 bits/group 64, Llama-3.2-3B retains ≥ 95 % of its B0 accuracy while30 Qwen2.5-3B loses ≥ 20 percentage points (motivated by E0 smoke: Qwen2.5-1.5B produced garbage at 4-bit in both the native31 mlx_lm path and ours; Llama-3.2-3B did not). 8-bit retains ≥ 97 % for both.32- **H3** Sink+recent eviction (B3) at ≤ 1/4 retention is worse than a *question-aware* attention-informed selection (B5-aware)33 at equal retained positions, but **not** reliably better than random selection (B4) on questions whose answer lies in the34 middle of the document.35- **H4** Serialized cache bytes exceed raw UTF-8 document bytes by ≥ 2 orders of magnitude for 3B models; hence B1 transfer only36 wins on total latency when (a) the receiver's prefill cost exceeds transport time, which on the measured LAN requires contexts37 above a break-even length we will estimate (prediction: several thousand tokens), or (b) the cache is reused ≥ N* times.38- **H5** Repeated prose handoffs lose facts monotonically; after 4 handoffs the retained-correct-fact rate of A4 chains falls below39 70 % while A5 (structured, explicit unknowns) stays above 85 %; unsupported additions appear mainly in prose chains.40- **H6** Metered retrieval (A6, repairable chains) restores most of the lost exact facts but adds fetched bytes comparable to41 the message itself.42- **H7** Question-aware compression dominates question-blind compression at equal budget for *all* lossy methods; the gap is43 largest for attention-based selection (which is query-dependent by construction).44- **H8** A lossless transport codec (A7, zstd) on text yields ≥ 2× byte savings with < 5 ms overhead but **no** change in receiver45 latency or quality; on bf16 KV tensors the codec yields < 1.3× (high-entropy payload).46- **H9** The adaptive router (E8) will not reliably beat the best development-selected static method on held-out data once47 overhead is included (we expect a null or small, uncertain gain).48- **H10** Cross-family cache transfer without a trained mapper is not possible; with a mapper trained under our budget (if feasible),49 quality is below the receiver's own A4 summary baseline (we expect learned transfer to lose to text at our scale).5051## Primary endpoints (frozen at Phase 7 together with the matrix)5253Per task family: exact-match / F1 (LongBench normalization) or numeric correctness with tolerance; operational success rate;54serialized payload bytes; end-to-end wall time (defined boundaries, §6.1 of the brief); peak MLX memory.5556## Primary confirmatory comparison family (Holm-corrected; everything else exploratory)57581. A5 vs A4 at matched byte budget (synthetic, question-blind) — accuracy.592. B2-8bit vs B0 — accuracy (both models).603. B2-4bit vs B0 — accuracy (Qwen vs Llama interaction stated as H2).614. B5-aware vs B3 at 1/4 retention — accuracy.625. A4 chain vs A5 chain at 4 handoffs — fact retention.636. Router vs best static — held-out accuracy at equal or lower cost.6465## Amendments6667(none yet)6869- **A-001 (2026-10-30, before freeze, after pilot on dev only).** Core matrix bounded for run time: 40 test documents per family (synthetic and70 exec = all; public = first 40 in manifest order), 5 questions per synthetic document, question-aware A4/A5 limited to the first 3 questions71 of synthetic documents only; one sender budget (512 tokens) and one retention ratio (1/4) in the registered comparisons; B2 at 8 and 4 bits72 (2-bit dropped from the core: pilot showed total failure on Qwen and no additional information). Campaign sharded by model pair across two73 machines (D-015). Track C uses the ridge mapper v2 (D-016). None of these changes used test outcomes.74- **A-002 (2026-10-30).** Pilot observation recorded without changing hypotheses: on dev, A5 (structured) was *below* A4 (summary) at the same75 budget; H1/T1 are kept as registered and may be rejected on test.76