Skip to content

dusterbloom/pflash-evidence

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PFlash Evidence — PR #274

Evidence dossier for PR #274 (adaptive composition: PFlash + DFlash). Branch feat/pflash-drafter-ee7 @ commit 5eede9c, RTX 3090, 2026-05-28.

Headline numbers

  • −39% drafter wall at 128K (4.57 s vs ee7 prior 7.48 s)
  • 11× end-to-end wall on HotpotQA (6.2 s vs 73.1 s uncompressed)
  • 5860 tok/s prefill (compressed, median N=10)
  • +24% tok/J composition vs pflash-only at 32K (NIAH 0/3 → 3/3)
  • Zero asserts / crashes across 5 bench axes

Files

File Purpose
index.html Single-page evidence site (open locally in browser)
EVIDENCE.md Developer appendix: all numbers, sources, epistemic tags
OPEN_QUESTIONS.md Open questions tracker (P2-J MED regression, P2-K gate break-even)
journey.md Historical narrative (prior to PR #274)

Bench data

lucebox-hub/bench/2026-05-28_adaptive_stack/ — axis_A through axis_E JSON results + server logs.

Links

About

PFlash engineering evidence — ee7 drafter early-exit + adaptive keep_ratio bandit MVP. Validated on Qwen3.6-27B-Q4_K_M / RTX 3090: 2.1-9.3× drafter speedup across 5 agentic clients (claude_code, hermes, opencode, pi, codex).

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages