localvmExperiments
Every experiment carries a registered hypothesis with an explicit falsification criterion (seven-field scientific block), a benchmark implementation, raw results, and an analysis. Negative results are kept.
expA_weight_concentration
Weight contribution concentration: can a subset of weight blocks reproduce most of each layer's output?
completed 1 result runexpB_token_stability
Stability of important weight blocks across consecutive tokens (Jaccard, transitions, working-set lifetime)
completed 1 result runexpC_semantic_locality
Cross-prompt semantic locality: do same-domain prompts reuse the same weight regions?
scaffoldedexpD_progressive_reconstruction
Progressive weight reconstruction: convergence of hidden-state/logit/decision error vs residual depth
completed 2 result runsexpE_partial_gemm
Partial matrix multiplication: quality vs blocks evaluated, incl. a Metal kernel bandwidth check
scaffoldedexpF_error_accumulation
Error accumulation: which layers tolerate, amplify, or recover from controlled approximation
completed 2 result runsexpG_decision_stability
Decision stability: how many token decisions are stable before full precision is available
completed 3 result runsexpH_ssd_feasibility
SSD feasibility on macOS/APFS: measured random/sequential reads, F_NOCACHE, concurrent Metal compute
completed 2 result runscandidate_01
Margin-gated deferred-refinement runtime: 4-bit resident base + windowed q8 verification sweeps with rollback
completed 6 result runs