1"""Claim ledger: every headline claim -> experiment ids, evidence table/figure, rows, script, status, limitation."""2from __future__ import annotations3import csv4from ..util import ROOT56CLAIMS = [7 ("C01", "Lossless representations (A1, A7, B0, B1, B2-8bit) give the same quality within noise on every family and pair", "E1_core", "results/tables/E1_core_primary_tests.csv (T2, T6); table_primary", "rows T2,T6", "analyze.py", "confirmed", "two 3B pairs, 160 test documents"),8 ("C02", "4-bit packed KV quantization destroys Qwen2.5-3B (0%) but not Llama-3.2-3B (>=97%)", "E1_core", "E1_core_primary.csv B2|bits=4; primary_tests T3", "condition B2|bits=4", "analyze.py", "confirmed (H2)", "mlx per-token affine quantization, group 64"),9 ("C03", "Attention-informed eviction with the question (B5-aware) beats sink+recent at 1/4 by +44 points; blind window variant collapses", "E1_core", "primary_tests T4; E1_core_primary.csv", "B5 vs B3", "analyze.py", "confirmed (H3 first half, H7)", "question-aware method"),10 ("C04", "All eviction methods score 0 on executable tasks at 1/4 retention", "E1_core", "E1_core_primary.csv family=exec", "B3,B4,B5,B6,B9", "analyze.py", "measured", "3B receivers, 6-rule spec"),11 ("C05", "Structured evidence (A5) is below prose summary (A4) at the same 512-token budget (-27 points, synthetic)", "E1_core", "primary_tests T1", "T1", "analyze.py", "H1 rejected", "sender spends budget on quotes; ~8 facts of 62"),12 ("C06", "zstd shrinks text ~3.5x with no quality/latency effect; ~20% on bf16 caches", "E1_core, E6_core", "E1_core_primary.csv A7; E6_network.csv codec=zstd", "A7; B1 zstd", "analyze.py", "confirmed (H8)", "zstd level 3"),13 ("C07", "Serialized 3B cache is 3-4 orders of magnitude larger than the text", "E1_core, E6_core", "E1_core_primary.csv payload_bytes_median; E6_network.csv bytes", "B0 vs A1", "analyze.py", "measured", "bf16, Qwen 215 MB vs 29 kB at ~8k tokens"),14 ("C08", "On 1 GbE with an M3 Ultra receiver, moving the cache (bf16 or 8-bit) answers faster than remote text prefill at 2k-16k tokens", "E6_core", "E6_network.csv; fig6", "lan rows", "analyze.py", "measured", "one link, one receiver; break-even depends on prefill/link ratio"),15 ("C09", "Repeated prose/structured relays lose facts at hop 1 and are nearly lossless afterwards (no monotone decay)", "E4_core", "E4_core_handoffs.csv; fig5", "hops 1-8", "analyze.py", "H5 rejected", "ledger metric = normalized value containment"),16 ("C10", "Unsupported additions are near zero in 7,200 relayed messages", "E4_core", "E4_core_handoffs.csv unsupported_mean", "all chains", "analyze.py", "measured", "regex-based detection"),17 ("C11", "Metered repair raises prose-chain retention 32->38% and accuracy 40->44% for ~5.9 kB cumulative fetch", "E4_core", "E4_core_handoffs.csv repairable", "A4 repairable", "analyze.py", "H6 partial", "repair limited to entities still named"),18 ("C12", "Concatenated summaries compose as well as full context on multi-hop joins; merged structured records do not", "E5_core", "E5_core_fanin.csv; table_fanin", "concat_A4 vs merge_A5", "analyze.py", "measured", "k=2,4; 3B aggregator"),19 ("C13", "Cache concatenation of independently prefilled sources yields plausible answers (70% multi-hop at k=2) despite invalid positions", "E5_core", "E5_core_fanin.csv cache_concat_negative_control", "k=2,4", "analyze.py", "negative control", "diagnostic only"),20 ("C14", "Hosted receivers recover 100% from full text, ~30% from the local 512-token summary, 12.5% from structured JSON", "ED_hosted_core", "ED_hosted.csv; table_hosted", "synthetic rows", "analyze.py", "measured", "provider default sampling; 8 documents/family"),21 ("C15", "Anthropic prompt caching did not engage below the documented 4,096-token minimum (2k docs) and halved estimated cost on 8k docs", "ED_hosted_core", "runs/raw/ED_hosted_core.jsonl provider_cached_usage", "A1 provider_cache rows", "hosted_run.py", "measured", "billing fields only"),22 ("C16", "Dev-fitted router matches best static quality at half the amortized latency in the blind regime; no gain and higher cost in the aware regime", "E8_router (E1_core records)", "E8_router_summary.csv; E8_router_tests.csv; fig7", "router vs best_static_dev", "router_eval.py", "H9 mixed", "off-policy on full factorial; dev table from pilot"),23 ("C17", "Linear ridge KV mapper (Qwen 1.5B->3B) yields coherent but uninformative receiver outputs; below sender-summary transfer", "E3_core", "E3_core tables", "C2 vs A4", "analyze.py", "confirmed (H10): C2 24.5% overall but 0-4% on content questions; negation yes/no at base rate", "reconstruction-only mapper; no SwaT reproduction"),24 ("C18", "Identity round trip and 100% retention reproduce native logits exactly; cached-prefix equivalence holds within registered tolerance", "E0 tests", "tests/test_cache_correctness.py", "12 tests", "pytest", "confirmed", "MLX nondeterminism across chunking documented"),25 ("C19", "Minimal reproduction reran 64 trials from a clean process with identical outputs and scores", "E9", "reproduction/report.json", "all", "reproduce.py", "confirmed", "same machine, same chunking"),26 ("C20", "Cross-family cache transfer was not executed (no SwaT checkpoint/licence; mapper budget)", "coverage", "literature/audit_group1.md; table1_taxonomy", "SwaT row", "n/a", "coverage gap", "partial transfer study"),27]2829def write_ledger(path=ROOT / "results/claim_ledger.csv"):30 path.parent.mkdir(parents=True, exist_ok=True)31 with open(path, "w", newline="") as f:32 w = csv.writer(f); w.writerow(["claim_id", "claim", "experiment_ids", "evidence_table_or_figure", "rows_or_filter", "script", "status", "limitation"]); w.writerows(CLAIMS)33 return path3435if __name__ == "__main__":36 print(write_ledger())37