Every number in the paper and on the interactive site
is computed from the JSON records in this directory (and ../mem/results/). Figures and the site
read these files directly — nothing is hand-entered.
Files are <experiment>_<model>[_<variant>].json. The <experiment> prefix tells you which
study produced it; the driver scripts live in ../esys/, ../e1/, ../e2/ (see those READMEs).
| prefix | experiment | driver(s) |
|---|---|---|
mech_*, mech_suite_*, mech_causal_*, mech_oracle_*, mech_reasoning_* |
the core mechanism probes (§3) | esys/mech_*.py |
mechd_* |
deep-mechanism controls — dissociation / timing / specificity / injection (§3, app.) | esys/mechd_*.py |
circ_* |
component-level circuit — heads / direction / SAE / scrubbing (app.) | esys/circ_*.py |
selective_*, ksweep_*, cost_frontier_*, frontier_*, erratum_*, why_erratum_*, diagnostic_* |
editing: the erratum, field+selective@K, the cost/correctness frontier (§4) |
esys/selective_*.py, esys/frontier.py, … |
composable_*, compose_edit_*, transplant_*, mla_*, arch_erratum_* |
composing: reposition + splice, edit-inside-transplant, adapters (§5–6, §8) | esys/composable_*.py, … |
editkv_unified_*, agent_*, editkv_agent_* |
the unified edit+compose agent (§6) | esys/editkv_unified.py, esys/editkv_agent*.py |
editkv_horizon_*, horizon_* |
long-trajectory no-compounding test (app.) | esys/editkv_horizon.py |
weight_edit_compare_* |
KV editing vs. ROME / LoRA weight editing (§4) | esys/weight_editing_compare.py |
vllm_*, serving_*, latency_* |
online serving + latency benchmarks (§9) | esys/vllm_*.py, … |
tau2_*, taubench_* |
the τ²-bench agentic environment (§9) | esys/tau2_*.py |
composable_vision_*, vision_* |
multimodal image-KV transplant (§8) | esys/composable_vision.py, esys/vision_ttft.py |
e1_*, e2*_*, recovery_*, selection_* |
the early exploratory harness (Qwen-1.5B/7B) | e1/, e2/ |
raw_qwen1p5b/ and raw_qwen7b/ hold raw outputs from those early exploratory runs.
selective_Ksweep_*_par.json are the canonical inputs to the field+selective@K figure
(esys/make_ksweep_figure.py). User-memory records (E1–E5, LoCoMo) live separately in
../mem/results/ as .jsonl.