{
  "schema_version": "1.0.0",
  "id": "CR-R-0001",
  "record_type": "review",
  "finding_id": "CR-F-0004",
  "review_type": "adversarial",
  "verdict": "confirm_with_limitations",
  "reviewer": {
    "id": "anthropic-claude-crystalram-review-2026-08-09",
    "kind": "human_ai_team",
    "role": "skeptic",
    "provider": "Anthropic",
    "model": "Claude Sonnet 5 (visible product label reported to the session; an exact versioned internal deployment identifier was not exposed to the runtime)",
    "organization": null,
    "sector": "cross-provider adversarial review",
    "orcid": null,
    "disclosures": [
      "The same human project steward who operates the CR-F-0004 author session (OpenAI Codex, GPT-5 family) also operates this reviewer session. This review therefore crosses an AI-provider boundary (Anthropic vs. OpenAI) but not a human-operator, dataset, or repository boundary.",
      "This review reused the same preserved, hash-verified trace CSVs, manifests, and prompt suite that CR-E-0005 produced. It did not capture any new OLMoE routes under any condition, so it is a cross-provider adversarial review, not an independent experimental reproduction, and does not satisfy CR-F-0004's outstanding 'reproduction' review requirement.",
      "The reviewer had no access to an MLX/OLMoE execution environment or the model weights in this workspace and could not rerun capture_olmoe_mlx_router_trace.py. Code-level fidelity of the capture wrapper was instead checked by direct comparison against the installed mlx-lm==0.31.3 source found in a sibling, non-worktree checkout on the same host (/Users/pilotsuite/Documents/CrystalRAM/.venv), not by executing it.",
      "All quantitative recomputation in this review used a script written independently for this review (research/reviews/CR-R-0001_independent_recompute.py) that does not import CrystalRAM-Simulator-v0/tools/analyze_layer_local_hot_bank.py, but it necessarily implements the same preregistered protocol (same bank rule, same paired whole-prompt bootstrap algorithm, same seed) described in CR-E-0005.json, so it cannot catch a defect in the shared protocol design itself, only in the shared implementation."
    ]
  },
  "created_at": "2026-08-09T16:19:55Z",
  "method": "Read AGENTS.md, research/PROTOCOL.md, GOVERNANCE.md, CR-F-0004.json, CR-E-0005.json, RESEARCH_LOG_007.md, the review template/schema, both capture manifests, the frozen result JSON/summary CSV, the prompt suite, the capture tool (capture_olmoe_mlx_router_trace.py), the analysis tool (analyze_layer_local_hot_bank.py), and the regression tests (test_hot_bank.py). Independently recomputed SHA-256 for every artifact hash cited in CR-F-0004's evidence list plus all 96 individual trace/provenance file hashes referenced inside both manifests (192 hash checks total), comparing against the declared values. Checked train/test leakage by tracing the frozen suite's split assignment through the analysis code's cross-checks and by diffing this suite's 48 prompt ids and texts against the CR-E-0004 pilot suite (moe_router_pilot_v1.json) for overlap. Derived the 6.25% random-identity expectation analytically (symmetry argument for a uniform k-of-n subset) rather than accepting it as asserted. Wrote a from-scratch Python script (not importing the reviewed analysis module) that parses the raw per-token, per-layer CSV route records directly from the preserved traces and independently recomputes: the primary cross-condition and native training-bank hit rates, the bank identity overlap fraction, the paired whole-prompt bootstrap gain-over-random and cross-minus-native 95% intervals (20,000 resamples, seed 130363), the early/middle/late fixed-window results, the per-family gains for all six families, and the idealized physical-geometry figures (module parameter count, resident bit/byte totals, density for a 400 mm^2 area). Checked every layer's training-count ranking in both banks for ties at the 4th/5th-place selection boundary to rule out tie-breaking sensitivity. Located the exact pinned dependency (mlx-lm==0.31.3, confirmed via installed dist-info metadata in a sibling checkout's virtual environment) and read mlx_lm/models/olmoe.py and switch_layers.py to compare the unmodified OlmoeSparseMoeBlock routing computation against the TracingGate wrapper's computation line by line. Grepped CR-F-0004.json, CR-E-0005.json, and RESEARCH_LOG_007.md for hardware/energy/latency/throughput/device/optical terms to check every occurrence is confined to an explicit exclusion or disclaimer. Ran the full completion checklist: CrystalRAM-Simulator-v0/tests/test_sim.py, tools/research_ledger.py validate/sync/summary, and python3 -m unittest discover -s tests -p 'test_*.py'.",
  "notes": [
    "Every artifact hash cited in CR-F-0004's evidence array (CR-E-0005.json, the prompt suite, both manifests, the frozen result JSON, the summary CSV, and test_hot_bank.py) matches the corresponding file on disk exactly. All 96 per-prompt trace_sha256 and provenance_sha256 entries inside both the 4bit and 8bit manifests also matched their referenced files, and every provenance sidecar's condition_id and generated_tokens (>=90) field checked out. No hash mismatch, truncation, or substitution was found anywhere in the evidence chain.",
    "No train/test leakage was found. The suite freezes prompts 01-05 as train and 06-08 as test per family; analyze_layer_local_hot_bank.py's load_condition() cross-checks each capture's declared split against this frozen suite assignment and raises if they disagree. The cross-condition and native banks are built only from *_train traces and scored only on eight-bit *_test traces; the held-out oracle bank is built from test-split counts but is explicitly excluded from every success gate and labeled a non-deployable descriptive ceiling. The 48 hotbank-* prompt ids and texts in this suite have zero overlap (by id or by exact text) with the 20 prompts in the earlier CR-E-0004 suite (moe_router_pilot_v1.json), supporting the disclosed 'entirely new relative to CR-E-0004' claim.",
    "The 6.25% random-identity expectation is not merely plausible but exactly correct: for any fixed top-8 routed expert set and a resident set of 4 experts drawn uniformly at random without replacement from 64, the probability that a specific routed expert falls in the resident set is exactly 4/64 regardless of the routing distribution (a standard property of uniform random subsets), so the expected hit rate is exactly 6.25% independent of the empirical routing pattern.",
    "Independent recomputation from the raw per-token, per-layer CSV traces, using a script written from scratch for this review (research/reviews/CR-R-0001_independent_recompute.py, SHA-256 865c0471d225e29783399055dee9c83e23d175cc2d71dd5a62d58a7432fee9e6) rather than the reviewed analysis module, reproduced every primary CR-F-0004 metric to the full reported precision: cross-condition hit rate 14.668782552083334%, native hit rate 14.671495225694445%, gain over random 8.418782552083334pp with 95% CI [5.134639033564814, 11.70970775462963], cross-minus-native -0.0027126736111110616pp with 95% CI [-0.5881980613425927, 0.5854853877314812], late-window gain 95% CI [4.969618055555555, 11.406583867521368], bank identity overlap 85.9375%, and the idealized 8-bit resident weight of 402,653,184 bytes (402.653184 MB decimal / 384 MiB binary) implying 8.05306368 bit/um^2 for a 400 mm^2 area.",
    "The per-family recomputation independently confirmed all six results, including the negative creative-writing outlier (3.776% hit rate, -2.474pp versus random), and confirmed the family consistency gate passes at exactly its preregistered boundary of 5 of 6 positive families, no more. Checked every layer of both training banks for ties at the 4th/5th expert selection boundary; none exist, so the reported bank composition and every downstream metric are not artifacts of the ascending-expert-id tie-breaking rule.",
    "Code-level inspection of the exact pinned dependency (mlx-lm==0.31.3, verified by installed dist-info metadata) substantiates rather than merely inherits the finding's central capture-fidelity assumption: TracingGate computes mx.softmax(logits, axis=1, precise=True) followed by mx.argpartition(-probabilities, kth=top_k-1, axis=-1)[..., :top_k] and returns the original logits unchanged, which is line-for-line identical to the softmax/argpartition sequence inside the unmodified OlmoeSparseMoeBlock.__call__; because both computations are deterministic functions of the same unmodified logits, the block recomputes exactly the indices the wrapper recorded. SwitchGLU/QuantizedSwitchLinear consume those indices directly with no capacity-based dropping or rerouting that could make the served computation diverge from the recorded route.",
    "A grep of CR-F-0004.json, CR-E-0005.json, and RESEARCH_LOG_007.md for hardware/energy/latency/throughput/device/optical/endurance/retention terms found every occurrence inside an explicit exclusion, disclaimer, or limitation (e.g. 'not a hardware, latency, energy, or model-quality result', 'excludes ... optical system overhead'), never as an asserted positive claim. No unsupported hardware, energy, latency, or device claim was found.",
    "python3 CrystalRAM-Simulator-v0/tests/test_sim.py (26 tests), tools/research_ledger.py validate/sync/summary, and python3 -m unittest discover -s tests -p 'test_*.py' (19 tests) all pass on this branch with CR-R-0001.json added."
  ],
  "limitations": [
    "This is cross-provider adversarial review, not independent experimental reproduction: no new four-bit or eight-bit OLMoE routes were captured, so CR-F-0004's stated requirement of one independent reproduction review remains unmet after this review.",
    "The same human project steward operates both the CR-F-0004 author session and this reviewer session, and both drew on the same repository and the same preserved traces. A systematic error shared by the capture tool, the frozen prompt suite design, the pinned model revisions, or the host environment would not be caught by this review, because this review did not regenerate any of those inputs from an independent source.",
    "Capture-wrapper fidelity was checked by reading the exact pinned mlx-lm==0.31.3 source located in a sibling, non-worktree checkout already present on the host, not by installing it fresh or by executing the capture tool in this session; this workspace has no MLX runtime or model weights available.",
    "This review's independent recomputation reimplements the same preregistered bank-construction and bootstrap algorithm that CR-E-0005 specifies (from raw CSVs, without importing the reviewed module), so exact numeric agreement rules out an implementation bug specific to analyze_layer_local_hot_bank.py but cannot by itself rule out a defect in the shared protocol design (e.g., whether whole-prompt bootstrap resampling is the right uncertainty unit for this data-generating process).",
    "The scope of the underlying claim remains as narrow as CR-F-0004 itself discloses: both conditions are quantized conversions of one OLMoE checkpoint, the 48-prompt suite is author-written rather than sampled production traffic, only 3 held-out prompts back each family result, and the positive-family gate passes at its exact preregistered boundary (5 of 6) rather than with margin.",
    "This review did not attempt device-physics verification of the idealized weight-bit density calculation's relevance to any real optical or electronic storage medium; it only checked that the arithmetic (parameter count, bit totals, area conversion) is internally correct and that the finding does not overstate it as a demonstrated device density.",
    "This review did not independently verify that the named Hugging Face model repositories and revisions (mlx-community/OLMoE-1B-7B-0125-Instruct-4bit and -8bit) actually correspond to the pinned revision hashes recorded in the manifests; it trusted the recorded revision strings without re-downloading the models."
  ],
  "evidence": [
    {
      "uri": "research/experiments/CR-E-0005.json",
      "sha256": "ad1d8fd40dd9ed87ce96e40a06b2e15e7835b36a45870714bc41b5196455c18d",
      "supports": "Independently reverified via SHA-256; matches the hash CR-F-0004.json cites for the preregistered protocol."
    },
    {
      "uri": "research/prompt_suites/layer_local_hot_bank_v1.json",
      "sha256": "0f1fdc4377526a38faf321b34d8f36e42f98a0ef83720395a44f9a7caee9c0f7",
      "supports": "Independently reverified via SHA-256; also used to confirm the frozen train/test split and to check for zero prompt overlap with the CR-E-0004 pilot suite."
    },
    {
      "uri": "CrystalRAM-Simulator-v0/traces/layer_local_hot_bank_v1/4bit/manifest.json",
      "sha256": "63cb0ed38579f0b830b7ec703d85a5b583bb2c3489ddd268c83d4e47f6a1446c",
      "supports": "Manifest hash reverified; all 48 referenced trace and provenance file hashes inside it independently reverified against the files on disk with zero mismatches."
    },
    {
      "uri": "CrystalRAM-Simulator-v0/traces/layer_local_hot_bank_v1/8bit/manifest.json",
      "sha256": "dd802aead48d56f001296106318efb7d1b6454bf88ab040d21b5d9913628cead",
      "supports": "Manifest hash reverified; all 48 referenced trace and provenance file hashes inside it independently reverified against the files on disk with zero mismatches."
    },
    {
      "uri": "CrystalRAM-Simulator-v0/results/layer_local_hot_bank_v1.json",
      "sha256": "ebd6b22796681bb86973d60d33b17784de84b0f4b464c03f359cdf29937a1f97",
      "supports": "Hash reverified; every primary, native, random, temporal, and family metric it reports was also independently recomputed from raw traces to the same precision."
    },
    {
      "uri": "CrystalRAM-Simulator-v0/results/layer_local_hot_bank_v1_summary.csv",
      "sha256": "f5d74631ac8855409b1c85574439296e9bc7d7cff8773937cc8c14bd3043ab16",
      "supports": "Hash reverified; primary-row values cross-checked against the independent recomputation."
    },
    {
      "uri": "CrystalRAM-Simulator-v0/tests/test_hot_bank.py",
      "sha256": "419427d481d0ab2e984b310f0f3bef6b5a3da71116230d2ac55a36fa5376b54e",
      "supports": "Hash reverified; test suite re-executed and passes on this branch."
    },
    {
      "uri": "research/reviews/CR-R-0001_independent_recompute.py",
      "sha256": "865c0471d225e29783399055dee9c83e23d175cc2d71dd5a62d58a7432fee9e6",
      "supports": "New evidence produced by this review: a from-scratch recomputation script (independent of analyze_layer_local_hot_bank.py) that reads only the preserved raw CSV traces and reproduces every primary CR-F-0004 metric to full reported precision, plus the physical-geometry arithmetic and per-family breakdown."
    },
    {
      "uri": "CrystalRAM-Simulator-v0/tools/capture_olmoe_mlx_router_trace.py",
      "sha256": "746196964907963bb5e5d4f85ebc1584bf4d23735ffc9d2bd7d0ad169fc3c1a0",
      "supports": "Reviewed the TracingGate wrapper implementation and compared it line-by-line against the pinned mlx-lm==0.31.3 OlmoeSparseMoeBlock routing computation to substantiate the capture-fidelity assumption in CR-F-0004's assumptions list."
    },
    {
      "uri": "/Users/pilotsuite/Documents/CrystalRAM/.venv/lib/python3.14/site-packages/mlx_lm/models/olmoe.py",
      "sha256": "4f8f78d368666ad0bf396963cd094bdf48caeec45d7a188c299bee7fc4bfea90",
      "supports": "Installed mlx-lm==0.31.3 (PyPI package, confirmed via dist-info metadata) found in a sibling, non-worktree checkout on the same host. Ground truth for the unmodified OLMoE gate computation; confirms TracingGate's softmax/argpartition sequence and unchanged-logits return value are identical to the routing computation the capture tool instruments, for the exact pinned dependency version."
    },
    {
      "uri": "/Users/pilotsuite/Documents/CrystalRAM/.venv/lib/python3.14/site-packages/mlx_lm/models/switch_layers.py",
      "sha256": "073a6a808d5c90bb699a2ecca0e559b06727ae96dbc1f0253e4c7e77e4ee1ef2",
      "supports": "Installed mlx-lm==0.31.3 (same sibling checkout as olmoe.py). Confirms SwitchGLU/QuantizedSwitchLinear consume the routing indices directly with no capacity-based dropping or rerouting, so the served computation cannot diverge from the recorded route."
    }
  ]
}