1claim_id,claim,experiment_ids,evidence_table_or_figure,rows_or_filter,script,status,limitation2C01,"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"3C02,4-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"4C03,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 method5C04,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"6C05,"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 627C06,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 38C07,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"9C08,"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"10C09,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 containment11C10,"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 detection12C11,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 named13C12,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"14C13,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 only15C14,"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/family16C15,"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 only17C16,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 pilot18C17,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 reproduction19C18,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 documented20C19,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"21C20,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 study22