CLAIMS.json

{
  "schema_version": "1.0.0",
  "claim_status": "Canonical ledger for current headline values and their receipt bindings. Public surfaces are checked against this file and the underlying receipts. Updated: 2026-09-06.",
  "baselines": {
    "f32_raw": {
      "description": "Eight independent contexts of 800 shared + 28 unique tokens, with 4-byte f32 K/V elements. Current N=8 size denominator: 2,604,662,784 bytes.",
      "bytes_per_element": 4
    },
    "fp16_equivalent": {
      "description": "The same independent-context geometry with 2-byte fp16 K/V elements. Ratios are exactly half the f32-raw ratios.",
      "bytes_per_element": 2
    },
    "bf16_equivalent": {
      "description": "The same independent-context geometry with 2-byte bf16 K/V elements. Same as fp16-equivalent for ratio purposes.",
      "bytes_per_element": 2
    }
  },
  "claim_classes": {
    "PPL_validated": "Deprecated status label retained for schema compatibility. No current claim may use it until a cache-aligned held-out-continuation benchmark passes.",
    "size_validated_shared_prefix_ppl_degraded": "The size denominator and compressed byte totals are receipt-validated. A separate cache-aligned shared-prefix continuation check produced materially worse PPL than the oracle and did not consume shell K/V; no PPL-neutrality or shell-quality claim is admitted.",
    "historical_unadmitted": "Historical receipt retained without publication admission because its PPL method predated the cache-aligned held-out-continuation contract.",
    "size_only": "Measured byte count for the named config. No PPL bench attached. These numbers are useful for N-scaling trends but are not load-bearing claims.",
    "wire_lossless": "The codec's wire format (FB2 / TQB1) is bit-exact: serializing a quantized code and deserializing produces an identical code. This is a per-codec property, not a per-config one.",
    "ppl_neutral": "Deprecated alias retained for schema compatibility; no current claim uses this status."
  },
  "claims": {
    "smollm2_wikitext2_n8_lossless_default": {
      "config": "SmolLM2-1.7B + WikiText-2 aggregate 1024-token fixture (800 shared + 28*8 contiguous unique slices), with an N=8 independent-agent size denominator of 8*(800+28) tokens; default two-tier policy (fib k4_n32 cold + turbo b=4 lossless hot).",
      "receipts": [
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/state_lossless.json",
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/shell_output_lossless/shared_pool_receipt.json",
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/shell_output_lossless/agents_receipt.json",
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/shell_output_lossless/state.json",
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/source_manifest.json",
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/source_envelope.json",
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/token_ids.json",
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/build_attestation.json"
      ],
      "claim_status": "size_validated_shared_prefix_ppl_degraded",
      "publication_eligible": true,
      "ppl_publication_eligible": false,
      "ratio_vs_f32_raw": 40.50399375986684,
      "ratio_vs_fp16_kv": 20.25199687993342,
      "raw_total_bytes": 2604662784,
      "compressed_total_bytes": 64306320,
      "pool_bytes": 14746512,
      "shell_bytes_per_agent": 6194976,
      "shell_codec": "turbo_4bit_batched",
      "pool_compression_ratio": 21.33,
      "oracle_ppl": 7.203125,
      "roundtrip_ppl": 24.234375,
      "delta_ppl_pct": 236.44251626898048,
      "ppl_window": [
        800,
        1024
      ],
      "ppl_window_note": "Cache-aligned shared-prefix continuation evaluation over exact target-token interval [800, 1024), 224 targets, on one aggregate fixture produced roundtrip PPL 24.234375 versus oracle 7.203125 (+236.44%). Only reconstructed shared-prefix positions [0, 799) are loaded; shell K/V is not used by this score. No PPL-neutrality or independent-prompt per-agent quality claim is admitted.",
      "msi_run_date": "2026-09-06",
      "source_manifest_sha256": "ed9e15435ab6a8218872ea7635300158567a73a29792435ba8f05c928319bf3a",
      "source_archive_sha256": "e6524df573762ef0be7bf0dffff6b0b156c827296bd0a561f6d42909f2a2c062"
    },
    "smollm2_wikitext2_n8_lossy_default": {
      "config": "SmolLM2-1.7B + WikiText-2 aggregate 1024-token fixture (800 shared + 28*8 contiguous unique slices), with an N=8 independent-agent size denominator of 8*(800+28) tokens; default two-tier policy (fib k4_n32 cold + turbo b=4 lossy hot, BlockLogU8 radii).",
      "receipts": [
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/state_lossy.json",
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/shell_output_lossy/shared_pool_receipt.json",
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/shell_output_lossy/agents_receipt.json",
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/shell_output_lossy/state.json",
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/source_manifest.json",
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/source_envelope.json",
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/token_ids.json",
        "results/ppl_multi_agent_b4_provenance_v2/smollm2-1.7b/wikitext-2-n8/build_attestation.json"
      ],
      "claim_status": "size_validated_shared_prefix_ppl_degraded",
      "publication_eligible": true,
      "ppl_publication_eligible": false,
      "ratio_vs_f32_raw": 76.54314453733862,
      "ratio_vs_fp16_kv": 38.27157226866931,
      "raw_total_bytes": 2604662784,
      "compressed_total_bytes": 34028688,
      "pool_bytes": 14746512,
      "shell_bytes_per_agent": 2410272,
      "shell_codec": "turbo_4bit_batched_lossy",
      "pool_compression_ratio": 21.33,
      "oracle_ppl": 7.203125,
      "roundtrip_ppl": 24.234375,
      "delta_ppl_pct": 236.44251626898048,
      "ppl_window": [
        800,
        1024
      ],
      "ppl_window_note": "Cache-aligned shared-prefix continuation evaluation over exact target-token interval [800, 1024), 224 targets, on one aggregate fixture produced roundtrip PPL 24.234375 versus oracle 7.203125 (+236.44%). Only reconstructed shared-prefix positions [0, 799) are loaded; shell K/V is not used by this score. No PPL-neutrality or independent-prompt per-agent quality claim is admitted.",
      "msi_run_date": "2026-09-06",
      "source_manifest_sha256": "ed9e15435ab6a8218872ea7635300158567a73a29792435ba8f05c928319bf3a",
      "source_archive_sha256": "e6524df573762ef0be7bf0dffff6b0b156c827296bd0a561f6d42909f2a2c062"
    },
    "smollm2_wikitext2_n8_lossless_legacy_b8": {
      "config": "SmolLM2-1.7B + WikiText-2, N=8 agents, 800 shared + 28*8 unique = 1024 tokens total, legacy b=8 two-tier policy (fib k4_n32 cold + turbo b=8 lossless hot). DEPRECATED — its historical receipt predates the explicit-window and baseline-derivation contract.",
      "receipts": [
        "results/ppl_multi_agent/smollm2-1.7b/wikitext-2-n8/state_lossless.json"
      ],
      "claim_status": "historical_unadmitted",
      "publication_eligible": false,
      "ppl_publication_eligible": false,
      "admission_note": "Historical full-cache/full-input PPL receipt; not admitted under the corrected cache-aligned held-out-continuation method. Any retained size fields are observational only.",
      "ratio_vs_f32_raw": 33.16444444444444,
      "ratio_vs_fp16_kv": 16.58222222222222,
      "raw_total_bytes": 2315255808,
      "compressed_total_bytes": 69811344,
      "pool_bytes": 14746512,
      "shell_bytes_per_agent": 6883104,
      "shell_codec": "turbo_8bit_batched",
      "pool_compression_ratio": 21.33,
      "oracle_ppl": 4.8125,
      "roundtrip_ppl": 4.8125,
      "delta_ppl_pct": 0.0,
      "roundtrip_seconds": 419.84,
      "msi_run_date": "2026-06-03"
    },
    "smollm2_wikitext2_n8_lossy_legacy_b8": {
      "config": "SmolLM2-1.7B + WikiText-2, N=8 agents, 800 shared + 28*8 unique = 1024 tokens total, legacy b=8 lossy two-tier policy. DEPRECATED — its historical receipt predates the explicit-window and baseline-derivation contract.",
      "receipts": [
        "results/ppl_multi_agent/smollm2-1.7b/wikitext-2-n8/state_lossy.json"
      ],
      "claim_status": "historical_unadmitted",
      "publication_eligible": false,
      "ppl_publication_eligible": false,
      "admission_note": "Historical full-cache/full-input PPL receipt; not admitted under the corrected cache-aligned held-out-continuation method. Any retained size fields are observational only.",
      "ratio_vs_f32_raw": 58.56408888888889,
      "ratio_vs_fp16_kv": 29.282044444444445,
      "raw_total_bytes": 2315255808,
      "compressed_total_bytes": 39533712,
      "pool_bytes": 14746512,
      "shell_bytes_per_agent": 3098400,
      "shell_codec": "turbo_8bit_batched_lossy",
      "pool_compression_ratio": 21.33,
      "oracle_ppl": 4.8125,
      "roundtrip_ppl": 4.8125,
      "delta_ppl_pct": 0.0,
      "msi_run_date": "2026-06-03"
    },
    "pool_only_smollm2": {
      "config": "SmolLM2-1.7B + WikiText-2, fib k4_n32 cold pool only, no shell tier.",
      "receipts": [
        "results/ppl/smollm2-1.7b/wikitext-2-lossless/state.json"
      ],
      "claim_status": "historical_unadmitted",
      "publication_eligible": false,
      "ppl_publication_eligible": false,
      "admission_note": "Historical full-cache/full-input PPL receipt; not admitted under the corrected cache-aligned held-out-continuation method. Any retained size fields are observational only.",
      "ratio_vs_f32_raw": 21.33,
      "ratio_vs_fp16_kv": 10.67,
      "pool_compression_ratio": 21.33,
      "note": "The fib codec is a codebook-based vector quantizer, NOT bit-exact lossless. The 21.33x ratio is the codebook size (32 codewords, 2 bits per dim) over the original vector size, less wire overhead."
    },
    "qwen0_5b_synthetic_n8": {
      "config": "Qwen2.5-0.5B + synthetic corpus, N=8 agents, multi_agent_compact_lossless_lossy bench.",
      "receipts": [
        "results/bench/multi_agent_compact_lossless_lossy/qwen2.5-0.5b/n8_lossless/state.json",
        "results/bench/multi_agent_compact_lossless_lossy/qwen2.5-0.5b/n8_lossy/state.json"
      ],
      "claim_status": "size_only",
      "lossless_ratio": 41.17,
      "lossy_ratio": 72.25,
      "warning": "Qwen0.5B size-only sweep. The naive_total_bytes denominator is a synthetic-size model (n_agents+1 * n_shared_tokens * num_layers * num_kv_heads * head_dim * 2 * 4), not a captured framework cache. Useful for N-scaling trends; not PPL-validated at N=8."
    }
  },
  "wire_lossless_certs": {
    "fib_fb2_batched": {
      "codec_id": "fib_k4_n32_batched",
      "claim_status": "wire_lossless",
      "test": "fib-quant::tests::fib_code_v1_batch_roundtrip"
    },
    "turbo_tqb1_batched": {
      "codec_id": "turbo_4bit_batched, turbo_8bit_batched, etc.",
      "claim_status": "wire_lossless",
      "test": "turbo-quant::tests::turbo_code_wire_v1_roundtrip"
    }
  },
  "non_claims": {
    "exact_kv_reconstruction": "We do NOT claim bit-exact reconstruction of the original K/V tensors. The fib cold tier is a codebook quantizer; the turbo hot tier is a polar/radii quantizer. For the current N=8 two-tier defaults, only the receipt-backed size ratios are publication-eligible; the cache-aligned aggregate fixture degraded from PPL 7.203125 to 24.234375.",
    "out_of_distribution_ppl": "We do NOT claim PPL-neutrality for the current N=8 two-tier defaults, including on WikiText-2. We also make no quality claim on out-of-distribution corpora, 4K/8K context, or models larger than SmolLM2-1.7B.",
    "vs_fp16_framework_cache": "We do NOT claim to reduce framework cache bytes directly. proveKV decompresses back to f32 and patches the cache; the framework cache is unchanged in size. The current size claim compares compressed bytes with an explicit denominator of eight independent 828-token f32 K/V contexts; fp16/bf16 figures are paper-derived equivalents.",
    "decode_wallclock_speedup_from_batch_path": "Wall-clock bench (results/bench/decode_wallclock/, fedora-43, release build, 5 reps) shows the batch decode path (TurboQuantizer::decode_approximate_batch) is 1.54x SLOWER than the per-vec path (TurboQuantizer::decode_approximate) on the msi PPL bench shape. The batch path's docstring claims it amortizes per-call overhead, but the actual cost is dominated by per-vec trig and per-vec allocation that the batch path does not actually batch. Do NOT quote a 'batch decode speedup' number without re-measuring.",
    "aggregate_fixture_is_not_independent_agent_ppl": "The PPL receipt scores one aggregate 1024-token fixture over exact target indices [800, 1024). The 40.50x / 76.54x size ratios use a separate explicit denominator: 8 independent contexts of 800 shared + 28 unique tokens. Do not describe the PPL result as eight independent-prompt PPL measurements.",
    "shell_quality_not_measured_by_current_n8_ppl": "The current N=8 PPL window starts at the shared/unique boundary. The cache-aligned forward loads reconstructed positions [0, 799) and recomputes token 799 onward, so the score exercises shared-prefix reconstruction but does not consume the reconstructed per-agent shell K/V. Shell receipts support the size result only; a held-out continuation after each reconstructed shared+unique context is required for shell-quality evidence."
  },
  "methodology": {
    "naive_computation": {
      "doc": "docs/methodology/naive_computation.md",
      "summary": "The current size estimand is 8 independent f32 contexts, each with 800 shared + 28 unique tokens: 2,604,662,784 raw bytes. Against 64,306,320 / 34,028,688 compressed bytes, that gives 40.50x lossless / 76.54x lossy. PPL is separately measured on one aggregate 1024-token fixture over target indices [800, 1024).",
      "geometric_fp16_naive_bytes": 1302331392,
      "claim_raw_total_bytes": 2604662784,
      "compressed_lossless_bytes": 64306320,
      "compressed_lossy_bytes": 34028688,
      "geometric_fp16_ratio_lossless": 20.25199687993342,
      "geometric_fp16_ratio_lossy": 38.27157226866931,
      "claim_ratio_lossless": 40.50399375986684,
      "claim_ratio_lossy": 76.54314453733862
    }
  },
  "hybrid_state": {
    "p0b_rust_gate": {
      "claim": "P0-B Rust gate: all 99 provekv tests pass with zero failures. Hybrid manifest identity, binary page persistence with BLAKE3 digests, atomic fsync/rename/dir-fsync writes, O(1)-metadata forks, mark-and-sweep GC, and per-component codec policy.",
      "claim_class": "test_gate",
      "receipt": "cargo test -p provekv",
      "git_commit": "3b38381",
      "date": "2026-08-05"
    },
    "p0c_capture_replay_gate": {
      "claim": "P0-C capture/replay gate: Qwen2.5-0.5B, 3/3 baselines agree, 48 KV pages, BLAKE3 roundtrip verified, DynamicCache reconstruction, suffix replay produces valid token. Qwen3.5-2B blocked pending transformers architecture support.",
      "claim_class": "test_gate",
      "receipt": "/tmp/provekv-capture-test/20260805T041422Z/receipt.json",
      "git_commit": "6e8f595",
      "date": "2026-08-05",
      "model": "Qwen/Qwen2.5-0.5B",
      "baseline_agreement": "3/3",
      "pages_verified": 48,
      "blocker_qwen35": "qwen3_5 architecture not in any transformers release"
    },
    "p0c_msi_gate": {
      "claim": "P0-C MSI gate: Qwen2.5-0.5B on MSI (8-core, 31GB RAM, CPU-only). 5/5 baselines agree (token 34208), 48 KV pages from 24 layers, BLAKE3 roundtrip verified, DynamicCache reconstruction, suffix decode produces valid token.",
      "claim_class": "test_gate",
      "receipt": "results/bench/hybrid_state/qwen25/capture/20260805T042634Z/receipt.json",
      "git_commit": "bcb7ebc",
      "date": "2026-08-05",
      "host": "msi",
      "model": "Qwen/Qwen2.5-0.5B",
      "baseline_agreement": "5/5",
      "pages_verified": 48
    },
    "p0c_gpu_gate": {
      "claim": "P0-C GPU gate: Qwen2.5-0.5B on MSI GTX 1070 (CUDA 11.8, float16). 5/5 baselines agree (token 34208), 48 KV pages from 24 layers, BLAKE3 roundtrip verified, DynamicCache reconstruction, suffix decode produces valid token.",
      "claim_class": "test_gate",
      "receipt": "results/bench/hybrid_state/qwen25/capture-gpu/20260805T050409Z/receipt.json",
      "git_commit": "4227b82",
      "date": "2026-08-05",
      "host": "msi",
      "gpu": "NVIDIA GeForce GTX 1070 (sm_61)",
      "cuda": "11.8",
      "dtype": "float16",
      "model": "Qwen/Qwen2.5-0.5B",
      "baseline_agreement": "5/5",
      "pages_verified": 48
    }
  }
}