{
  "schema_version": "1.0.0",
  "id": "CR-F-0004",
  "record_type": "finding",
  "title": "Fresh cross-quantization OLMoE traces support a 64-module Layer-Local Hot Bank",
  "claim": "On eighteen held-out prompts from six authored task families, a fixed four-expert-per-layer bank learned only from four-bit OLMoE training routes captured 14.669% of eight-bit generated-phase expert selections versus the exact 6.25% random-identity expectation, and was non-inferior to an eight-bit-native training-hot bank; this establishes a narrow cross-quantization expert-identity locality result, not a hardware, latency, energy, or model-quality result.",
  "evidence_class": "empirical_model_trace",
  "status": "reported",
  "confidence": {
    "score": 0.84,
    "level": "high",
    "rationale": "The narrow claim passed every preregistered gate on a fresh balanced suite, two immutable quantization conditions, training-only placement, paired whole-prompt uncertainty, fixed temporal windows, complete all-layer traces, and hash-bound artifacts. Confidence does not extend to other checkpoints or architectures, production traffic, or device-level benefit, and no independent reproduction has yet completed."
  },
  "created_at": "2026-08-09T07:20:59Z",
  "updated_at": "2026-08-09T07:20:59Z",
  "contributors": [
    {
      "id": "crystalram-founders-session-2026-08-08",
      "kind": "human_ai_team",
      "role": "simulation_scientist",
      "provider": "OpenAI",
      "model": "Codex (GPT-5 family; exact deployment version unavailable to runtime)",
      "organization": null,
      "sector": "cross-sector independent research",
      "orcid": null,
      "disclosures": [
        "The prompt suite, model revisions, split, bank budget, random baseline, non-inferiority margin, fixed windows, family gate, bootstrap method, and code were committed before any CR-E-0005 aggregate result was computed.",
        "Both quantization conditions are MLX conversions of the same source checkpoint, so this is not cross-checkpoint or cross-architecture validation.",
        "The creative-writing family did not beat the random-identity expectation; the preregistered consistency gate passed at its exact boundary of five positive families out of six.",
        "No independent reproduction, external code review, adversarial review, or device-physics review had completed when this finding was reported."
      ]
    }
  ],
  "method": {
    "summary": "Capture deterministic greedy top-eight router selections for all sixteen layers and 48 entirely fresh prompts under pinned four-bit and eight-bit OLMoE conversions. Learn four resident expert identities per layer from thirty four-bit training prompts, freeze those 64 identities, and score them on eighteen eight-bit held-out prompts. Compare against the exact equal-capacity random-identity expectation and a bank learned only from the thirty eight-bit training prompts using 20,000 paired whole-prompt bootstrap samples. Preserve fixed generated-token windows, all six families, secondary budgets, quantization-stability diagnostics, and physical lower-bound geometry.",
    "code_revision": "af18f250b4b7eae4de8d8789b884aa096d7f8c38",
    "configuration": "research/experiments/CR-E-0005.json",
    "data": "CrystalRAM-Simulator-v0/traces/layer_local_hot_bank_v1",
    "random_seed": 130363,
    "reproduce": [
      ".venv/bin/python CrystalRAM-Simulator-v0/tools/capture_olmoe_mlx_router_trace.py --model mlx-community/OLMoE-1B-7B-0125-Instruct-4bit --revision e23844197887b031e7ddddbb0b8959c5a6853a7b --suite research/prompt_suites/layer_local_hot_bank_v1.json --condition-id 4bit --manifest-id layer-local-hot-bank-v1-4bit --output-dir CrystalRAM-Simulator-v0/traces/layer_local_hot_bank_v1/4bit",
      ".venv/bin/python CrystalRAM-Simulator-v0/tools/capture_olmoe_mlx_router_trace.py --model mlx-community/OLMoE-1B-7B-0125-Instruct-8bit --revision 7055a795fc029a51108f881d4a118c6f17deb59f --suite research/prompt_suites/layer_local_hot_bank_v1.json --condition-id 8bit --manifest-id layer-local-hot-bank-v1-8bit --output-dir CrystalRAM-Simulator-v0/traces/layer_local_hot_bank_v1/8bit",
      "python3 CrystalRAM-Simulator-v0/tools/analyze_layer_local_hot_bank.py",
      "python3 CrystalRAM-Simulator-v0/tests/test_hot_bank.py",
      "python3 tools/research_ledger.py validate"
    ]
  },
  "results": [
    {
      "metric": "primary_cross_condition_heldout_hit_rate",
      "value": 14.668782552083334,
      "unit": "%",
      "uncertainty": null,
      "scope": "Four-bit-trained bank with four identities per layer scored across all 96 generated tokens of eighteen eight-bit held-out prompts"
    },
    {
      "metric": "primary_random_identity_expected_hit_rate",
      "value": 6.25,
      "unit": "%",
      "uncertainty": null,
      "scope": "Exact expectation for four uniformly random resident identities among 64 experts in every layer"
    },
    {
      "metric": "primary_cross_condition_gain_over_random_expectation",
      "value": 8.418782552083334,
      "unit": "percentage_points",
      "uncertainty": {
        "method": "paired_prompt_percentile_bootstrap_against_exact_random_identity_expectation",
        "confidence": 0.95,
        "samples": 20000,
        "seed": 130363,
        "low": 5.134639033564814,
        "high": 11.70970775462963
      },
      "scope": "Eighteen eight-bit held-out prompts; whole prompts are the resampling unit"
    },
    {
      "metric": "primary_eight_bit_native_training_hot_hit_rate",
      "value": 14.671495225694445,
      "unit": "%",
      "uncertainty": null,
      "scope": "Same 64-module capacity learned only from thirty eight-bit training prompts and scored on the same held-out prompts"
    },
    {
      "metric": "primary_cross_condition_minus_native_hit_rate",
      "value": -0.0027126736111110616,
      "unit": "percentage_points",
      "uncertainty": {
        "method": "paired_prompt_percentile_bootstrap",
        "confidence": 0.95,
        "samples": 20000,
        "seed": 130363,
        "low": -0.5881980613425927,
        "high": 0.5854853877314812
      },
      "scope": "Preregistered non-inferiority margin was -2.0 percentage points"
    },
    {
      "metric": "late_window_cross_condition_gain_over_random_expectation",
      "value": 8.23150373931624,
      "unit": "percentage_points",
      "uncertainty": {
        "method": "paired_prompt_percentile_bootstrap_against_exact_random_identity_expectation",
        "confidence": 0.95,
        "samples": 20000,
        "seed": 130363,
        "low": 4.969618055555555,
        "high": 11.406583867521368
      },
      "scope": "Fixed generated-token window 64 through 89 on all eighteen held-out prompts"
    },
    {
      "metric": "families_with_positive_mean_gain_over_random",
      "value": 5,
      "unit": "families_of_6",
      "uncertainty": null,
      "scope": "Coding, planning, quantitative reasoning, science explanation, and structured transformation were positive; creative writing was -2.474 percentage points"
    },
    {
      "metric": "four_bit_vs_eight_bit_training_bank_identity_overlap",
      "value": 85.9375,
      "unit": "%",
      "uncertainty": null,
      "scope": "Fraction of 64 resident identities shared by independently learned four-bit and eight-bit training-hot banks"
    },
    {
      "metric": "prompt_prefill_selected_expert_overlap",
      "value": 93.2699146514936,
      "unit": "%",
      "uncertainty": null,
      "scope": "Mean selected-expert overlap for 44,992 layer-token route records with identical prompt-prefill token IDs across quantization conditions"
    },
    {
      "metric": "four_bit_training_to_eight_bit_heldout_routing_total_variation",
      "value": 5.105432581018518,
      "unit": "%",
      "uncertainty": null,
      "scope": "Mean expert-popularity total variation across sixteen layers"
    },
    {
      "metric": "idealized_eight_bit_64_module_resident_weight",
      "value": 402.653184,
      "unit": "MB",
      "uncertainty": null,
      "scope": "Weight bits only; excludes quantization metadata and all optical system overhead"
    },
    {
      "metric": "idealized_eight_bit_64_module_density_required_for_400mm2",
      "value": 8.05306368,
      "unit": "bit/um2",
      "uncertainty": null,
      "scope": "Weight-bit lower bound excluding addressing, conversion, control, redundancy, and packaging"
    }
  ],
  "assumptions": [
    "The MLX gate wrapper records the same top-eight expert identity set consumed by the unmodified OLMoE routing computation while returning the original gate logits unchanged.",
    "Each selected expert identity counts as one module access; gate weights and expert compute cost are not used to weight hits.",
    "The analytical random baseline chooses four resident identities uniformly without replacement among 64 experts independently within every layer.",
    "A deployment using identities learned from four-bit routes stores the selected experts' eight-bit weights when serving the eight-bit model.",
    "One idealized expert module contains 3 times 2048 times 1024 parameters; physical equivalents omit all non-weight costs."
  ],
  "limitations": [
    "Both models are quantized conversions of one OLMoE checkpoint, so the result does not establish cross-checkpoint, cross-training-run, cross-architecture, or full-precision behavior.",
    "The 48-prompt suite was authored for this experiment and is not sampled production traffic; only three held-out prompts contribute to each family result.",
    "Greedy generations stop at 96 tokens, and the decisive persistence window ends at token 89, leaving long contexts, multi-turn conversations, batches, sampling, and online distribution change untested.",
    "Creative writing was the sole negative family, with a 3.776% hit rate against the 6.25% expectation, so workload-aware or adaptive placement may be necessary.",
    "Expert-identity hits are locality measurements, not measured HBM bytes avoided, latency, throughput, energy, model quality, optical loss, conversion cost, programming rate, eviction behavior, or cell endurance.",
    "The density calculation is an ideal weight-bit lower bound, not a demonstrated optical storage density or a complete device-area estimate.",
    "No independent reproduction or external review has yet tested the capture wrapper, analysis implementation, prompt sensitivity, or hardware interpretation."
  ],
  "evidence": [
    {
      "kind": "configuration",
      "uri": "research/experiments/CR-E-0005.json",
      "sha256": "ad1d8fd40dd9ed87ce96e40a06b2e15e7835b36a45870714bc41b5196455c18d",
      "title": "Preregistered cross-quantization Layer-Local Hot Bank experiment",
      "supports": "The frozen question, exact model revisions, success and failure gates, train and held-out split, primary budget, temporal windows, and uncertainty method.",
      "accessed_at": null
    },
    {
      "kind": "configuration",
      "uri": "research/prompt_suites/layer_local_hot_bank_v1.json",
      "sha256": "0f1fdc4377526a38faf321b34d8f36e42f98a0ef83720395a44f9a7caee9c0f7",
      "title": "Fresh balanced 48-prompt confirmation suite",
      "supports": "The complete prompt text, six families, immutable five-training and three-held-out assignments per family, and decoding length.",
      "accessed_at": null
    },
    {
      "kind": "dataset",
      "uri": "CrystalRAM-Simulator-v0/traces/layer_local_hot_bank_v1/4bit/manifest.json",
      "sha256": "63cb0ed38579f0b830b7ec703d85a5b583bb2c3489ddd268c83d4e47f6a1446c",
      "title": "Four-bit all-layer capture manifest",
      "supports": "Forty-eight verified four-bit prompt traces, 96 bound artifact hashes, the exact model and code revisions, split labels, and minimum-token validity.",
      "accessed_at": null
    },
    {
      "kind": "dataset",
      "uri": "CrystalRAM-Simulator-v0/traces/layer_local_hot_bank_v1/8bit/manifest.json",
      "sha256": "dd802aead48d56f001296106318efb7d1b6454bf88ab040d21b5d9913628cead",
      "title": "Eight-bit all-layer capture manifest",
      "supports": "Forty-eight verified eight-bit prompt traces, 96 bound artifact hashes, the exact model and code revisions, split labels, and minimum-token validity.",
      "accessed_at": null
    },
    {
      "kind": "artifact",
      "uri": "CrystalRAM-Simulator-v0/results/layer_local_hot_bank_v1.json",
      "sha256": "ebd6b22796681bb86973d60d33b17784de84b0f4b464c03f359cdf29937a1f97",
      "title": "Frozen cross-quantization Layer-Local Hot Bank analysis",
      "supports": "The primary, native, random, temporal, family, layer, quantization-stability, secondary-budget, and physical-equivalent results.",
      "accessed_at": null
    },
    {
      "kind": "artifact",
      "uri": "CrystalRAM-Simulator-v0/results/layer_local_hot_bank_v1_summary.csv",
      "sha256": "f5d74631ac8855409b1c85574439296e9bc7d7cff8773937cc8c14bd3043ab16",
      "title": "Layer-Local Hot Bank budget summary",
      "supports": "A compact primary and preregistered secondary budget comparison with intervals and placement overlap.",
      "accessed_at": null
    },
    {
      "kind": "test",
      "uri": "CrystalRAM-Simulator-v0/tests/test_hot_bank.py",
      "sha256": "419427d481d0ab2e984b310f0f3bef6b5a3da71116230d2ac55a36fa5376b54e",
      "title": "Layer-Local Hot Bank regression and synthetic fixture tests",
      "supports": "Equal per-layer placement, fixed token windows, prefill agreement, deterministic non-inferiority bootstrap, and end-to-end confirmatory analysis behavior.",
      "accessed_at": null
    },
    {
      "kind": "source",
      "uri": "https://huggingface.co/mlx-community/OLMoE-1B-7B-0125-Instruct-4bit",
      "sha256": null,
      "title": "Pinned four-bit MLX OLMoE conversion",
      "supports": "The public model repository and four-bit conversion used for training-bank route capture; the immutable revision is bound in the manifest.",
      "accessed_at": "2026-08-09T07:20:59Z"
    },
    {
      "kind": "source",
      "uri": "https://huggingface.co/mlx-community/OLMoE-1B-7B-0125-Instruct-8bit",
      "sha256": null,
      "title": "Pinned eight-bit MLX OLMoE conversion",
      "supports": "The public model repository and eight-bit conversion used for the native comparator and decisive held-out route capture; the immutable revision is bound in the manifest.",
      "accessed_at": "2026-08-09T07:20:59Z"
    }
  ],
  "relationships": [
    {
      "type": "derived_from",
      "target_id": "CR-E-0005",
      "note": "This finding reports every preregistered confirmation gate and preserves the complete positive, null, and negative results."
    },
    {
      "type": "supports",
      "target_id": "CR-F-0003",
      "note": "Fresh prompts and an eight-bit held-out condition support the earlier post-result lead that training-hot expert identities can populate a useful regular per-layer bank."
    }
  ],
  "review_requirements": {
    "independent_reproductions": 1,
    "required_review_types": [
      "reproduction",
      "code",
      "adversarial",
      "device_physics"
    ]
  }
}