# Zincrom Substantive Claim-to-Evidence Matrix

**Evaluation envelope:** Single Oracle OCI A1 instance (4 Arm Neoverse-N1 cores, 24 GiB RAM)  
**Evidence cutoff:** 2026-07-25  
**Canonical lineage:** `llama.cpp` commit `6e9007ae6` (build 9641 lineage)

| Claim ID | Substantive claim | Status | Evidence path and primary metric | Scope and limitations | Draft locations |
|---|---|---|---|---|---|
| `CLM-ROUT-01` | OLMoE router and top-K selection can be separated from expert execution via the NIF boundary. | Supported | `results/track2_test1_verdict.md`; `docs/CURRENT_EVIDENCE.md` | Single-token and batch routing graph materialized at `ffn_moe_topk`. | Paper §2; HTML Chapter 3 |
| `CLM-CORR-01` | Split OLMoE execution achieves exact numerical equivalence with pinned `llama.cpp` for single tokens across all layers. | Supported | `results/external-equivalence-20260725T031655Z/summary.md` (48 cases: tokens 50279, 42, 1000 across layers 0–15; weight delta 0.0; output delta 0.0) | OLMoE Q4_K_M on commit `6e9007ae6`. Excludes autoregressive multi-step decode. | Paper §3.1; HTML Chapter 3 |
| `CLM-CORR-02` | Split OLMoE execution achieves exact numerical equivalence for causal batches 2, 4, 8, and 16 across all 16 layers; batch 32 fails the strict gate. | Supported (batches 2–16); rejected (batch 32) | `results/multitoken-equivalence-20260725T031743Z/summary.md` (80 cases; batches 2–16 maximum delta 0.0; batch-32 layer-2 output delta `1.1444e-5` > `1.0e-5`) | Batch 32 maintains exact expert IDs but exceeds the output threshold at layer 2. Autoregressive multi-step token equivalence remains open. | Paper §3.1; HTML Chapter 3 |
| `CLM-OVHD-01` | BEAM coordination latency is small relative to selected-expert compute in the isolated split-path harness. | Supported, narrow | `results/sprint3-overhead-20260724T012420Z/summary.md` (1,000 iterations: coordination p99 74 µs vs compute p99 2,205.553 µs; ratio 3.36%) | Workers execute confirmation/dispatch only; no weight I/O or math in the BEAM process. | Paper §3.2; HTML Chapter 3; `hf_model_card.md` §1 |
| `CLM-DISP-01` | BEAM no-op worker dispatch overhead measured 37 µs at p99. | Supported, narrow | `docs/CURRENT_EVIDENCE.md` (8-worker no-op dispatch p99 37 µs) | Workers perform no I/O, allocation, or arithmetic. | Not asserted in publication drafts; evidence authority only. |
| `CLM-BASE-01` | The historical 8B OEF collapse to 0.270 was an artifact of denominator mismatch. | Superseded | `docs/CURRENT_EVIDENCE.md` (4-worker contention efficiency: 3B 0.967, 8B 0.963, OLMoE 0.964) | Four Neoverse-N1 cores, one thread per pinned worker versus four-thread isolated baseline. Bandwidth saturation was not proven. | Paper §2.2; HTML Chapter 2 |
| `CLM-MEM-01` | Default OS mmap fault-around inflates active MoE parameter residency above the theoretical 12.5%. | Supported | `docs/CURRENT_EVIDENCE.md` (`T021609Z`: logical 12.5%, PTE-present 27.7%, RSS delta 1.089 GB vs 3.921 GB all-expert control) | Linux default readahead/fault-around behavior on ARM64 page tables. | Paper §4.1; HTML Chapter 4 |
| `CLM-MEM-02` | `MADV_RANDOM` restores near-logical active-expert RSS scaling for isolated requests. | Supported, narrow | `docs/CURRENT_EVIDENCE.md` (`T022138Z`: PTE-present 12.88%, RSS delta 12.98% / 506 MB vs 3.899 GB) | Single hidden state per layer. Does not prevent multi-token working-set accumulation across decode rounds. | Paper §4.1; HTML Chapter 4 |
| `CLM-LIFE-01` | Token-boundary expert eviction is sufficient to bound concurrent multi-token residency. | Rejected | `docs/CURRENT_EVIDENCE.md` (PTE presence 28.0%/46.2%/61.4% for 1/2/4 requests vs 20%/35%/55% limits) | Eviction occurs too late and at coarse granularity. | Paper §4.2; HTML Chapter 5 |
| `CLM-LIFE-02` | Layer-completion eviction bounds concurrent expert residency with measured latency overhead. | Supported, narrow | `docs/CURRENT_EVIDENCE.md` (PTE presence 3.53%/4.67%/6.11%; staggered churn at 16/64/256 tokens retains 85.4%–90.8% throughput, p99 overhead 1.17×–1.23×, peak RSS delta 340–364 MB vs 4.04 GB control) | Synchronized barrier across concurrent contexts. Tested through a 256-token retained KV horizon on the four-core host. | Paper §4.2–4.3; HTML Chapters 5–6 |
| `CLM-PREF-01` | Same-host oracle lookahead prefetch hides expert-page readiness behind current compute. | Rejected | `docs/CURRENT_EVIDENCE.md` (`T014000Z`, `T014557Z`, `T015149Z`, `T020219Z`, `T021010Z`: latency failed in 7/12, 11/12, 7/12, and 5/12 conditions; batch-8 three-repeat gate failed at layer 14→15) | Tested with user page touch, `MADV_WILLNEED`, thread reservation, and a core-pinned helper process. | Paper §6.1; HTML Chapter 7 |
| `CLM-NET-01` | Redpanda streaming supplies real 31.1 MiB expert shards within the per-layer compute window. | Rejected | `docs/CURRENT_EVIDENCE.md`; `scripts/k8a_redpanda_transport.py` (fetch p99 86.96–138.68 ms vs compute p99 7.2–43.4 ms; 0/9 cases passed; 64 MiB control p99 260.95–269.44 ms; supersedes `results/TRACK2_TEST2_REDPANDA_RESULT.md`) | Single-host Redpanda broker with native and persistent brod clients. | Paper §6.2; HTML Chapter 8 |
| `CLM-SCHED-01` | Active decode concurrency should be capped at four on the tested four-core ARM host. | Supported, narrow | `docs/CURRENT_EVIDENCE.md` (peak-8 vs peak-4: throughput +21.1%, p99 latency +73.1% > 50% limit; FIFO cap-4 reduces service p99 35.7% and RSS 34.8%) | Specific to four Neoverse-N1 cores with one thread per decoder context. | Paper §5; HTML Chapter 9 |
| `CLM-SCHED-02` | FIFO scheduling is fair for deep mixed-length queues under MoE execution. | Rejected | `docs/CURRENT_EVIDENCE.md` (16-request, 2,104-token run: slowdown Jain fairness 0.739 < 0.80 minimum) | Mixed request lengths of 56–224 tokens. | Paper §5; HTML Chapter 9 |
| `CLM-SCHED-03` | Aging size-aware scheduling restores deep-queue fairness without memory regression. | Supported, narrow | `docs/CURRENT_EVIDENCE.md` (priority `budget - waited`: fairness 0.857–0.901, wait p99 65.016 s vs 71.584 s FIFO, throughput 19.26 token/s, peak RSS 351 MB) | Tested with declared token lengths. Online estimation error remains unvalidated. | Paper §5; HTML Chapter 9 |
| `CLM-LRU-01` | Routed-expert LRU cache capacity is universally predictable from top-K routing count. | Rejected | `docs/CURRENT_EVIDENCE.md` (OLMoE [8/64] favored LRU-8; Granite [8/40] favored LRU-4; Qwen [4/64+4] favored LRU-4) | OLMoE and Granite both route eight experts but have different optimal caches. Requires an explicit allowlist. | Paper §5.2; HTML Chapter 9; `hf_model_card.md` §1 |
| `CLM-LRU-02` | Directional LRU throughput gain correlates with avoided cold-fault cost rather than raw hit rate. | Supported directionally | `docs/CURRENT_EVIDENCE.md` (OLMoE hit 43.1% / gain 1.8%; Granite hit 27.2% / gain 2.6%; Qwen hit 17.1% / gain 4.2%; cost-weighted ordering 17.8 µs < 41.9 µs < 85.7 µs matches gain ordering) | Three evaluated models only; proxy based on median cold-fault times. Non-causal. | Paper §5.2; HTML Chapter 9 |
| `CLM-SUSP-01` | F16 sequence-state suspension produces bit-exact context restoration; Q8 alters logits. | Supported | `docs/CURRENT_EVIDENCE.md` (`results/kv-suspension-probe-20260724`: F16 zero logit delta; Q8 changed the top token in 2/4 checkpoints with mean absolute error 0.099–0.130) | Tested on Qwen decoder context across 16–224-token horizons. | Paper §5.3; HTML Chapter 10 |
| `CLM-SUSP-02` | Fixed quantum Q24/Q28 suspension is a universal sub-1-GB memory policy. | Rejected as universal | `docs/CURRENT_EVIDENCE.md` (Q24 reached 1.279 GB on 224-token requests and 1.059 GB on a 24-request burst; Q28 reached 1.241 GB on max-prefix and 1.056 GB on queue-24) | Supported only for bounded 16-request workloads; memory remains workload-dependent. | Paper §5.3; HTML Chapter 10 |
| `CLM-FFT-01` | Causal FFT convolution trains at toy scale. | Supported, narrow | `docs/CURRENT_EVIDENCE.md` (three-seed Shakespeare character model: approximately 5.9% lower validation loss vs attention baseline) | Character-level toy model only. | Not asserted in publication drafts; evidence authority only. |
| `CLM-FFT-02` | FFT spectral magnitude provides semantic routing in activation space. | Inconclusive | `docs/CURRENT_EVIDENCE.md` (initial test crossed kill threshold; coordinate order arbitrary and phase discarded) | Non-empirical/unverified mechanism. | Paper §7.3; not asserted in HTML article |
| `CLM-COG-01` | A few-million-parameter frozen cognitive core is separable from world knowledge. | Untested | `docs/CURRENT_EVIDENCE.md`; `docs/publication/RELEASE_SCOPE.md` | Conceptual research hypothesis only; no valid empirical experiment was executed. | Paper §7.3; not asserted in HTML article |