id,representation,question_access,training,compatibility,composable,status_in_wp30
A0,none (task only),n/a,none,any,n/a,measured
A1,raw text,blind,none,any receiver,concatenation,measured
A2,text head/tail,blind,none,any,concatenation,measured
A3,extractive sentences (BM25/centrality),blind or aware,none,any,concatenation,measured
A4,abstractive summary,blind or aware,none (prompted sender),any,concatenation,measured
A5,typed facts + source spans + unknowns (JSON),blind or aware,none (prompted sender),any,record merge,measured
A6,reference index + metered retrieval,"blind index, aware retrieval",none,any + source access,shared memory,measured
A7,"raw text, zstd transport codec",blind,none,any,concatenation,measured
A8,LLMLingua-2 token dropping,blind,pretrained compressor (official),any,concatenation,measured
B0,native bf16 KV cache,blind,none,same checkpoint,no,measured
B1,serialized KV cache (manifest-checked),blind,none,same checkpoint + revision,no,measured (process + 2-node)
B2,packed quantized KV (8/4/2 bit),blind,none,same checkpoint,no,measured
B3,sink+recent retention (StreamingLLM-inspired),blind,none,same checkpoint,no,measured
B4,random retention (control),blind,none,same checkpoint,no,measured
B5,attention-informed retention (SnapKV rule),aware (blind variant tested),none,same checkpoint,no,measured
B6,pyramid layer budgets (PyramidKV rule),aware,none,same checkpoint,no,measured
B9,selection then packed quantization,aware,none,same checkpoint,no,measured
C0,sender cache fed as receiver cache,blind,none,incompatible (negative control),no,measured (control)
C2,"linear KV mapper (ridge, un-rotated K/V)",blind,ridge on train caches,trained pair (Qwen 1.5B→3B),no,measured
C3,mapper + attention selection,aware,ridge on train caches,trained pair,no,measured
—,SwaT cross-family mapper,blind,~300M mapper (not public),Granite→SmolLM,no,not run: no checkpoint/licence
—,Attention-Matching compaction,blind,none (optimization),same model,no,not run: CUDA/vLLM code; MLX port out of scope
—,C2C fuser,blind,released fusers (Qwen3),Qwen3 receivers,no,not run: CUDA; receiver still prefills source
—,H2O eviction,cumulative,none,same checkpoint,no,not run: requires attention scores during whole prefill
—,KIVI per-channel key quantization,blind,none,same checkpoint,no,not run: CUDA kernels; motivation confirmed by B2 Qwen result