PADS·Explainer

Progress

Where the PADS project stands: what is measured and done, what is blocked, and what is explicitly left for next semester. Every number below names the artifact it came from.

Read this before quoting any number

Every measurement on this page was taken on the development host (a laptop with a discrete GPU used for training), not on the Dell Latitude 5490 target hardware. Wall-clock speedup for PADS proper is not measured, audio turn-taking is not implemented, and energy measurement is blocked on root access.

Overall: about 63% — and how that number is computed

We score six project blocks equally — Foundations & Go/No-Go, Phase 1 safe core, Phase 2 exit policies, Phase 3 baselines, Phase 4 turn-taking + PADS end-to-end, Phase 5 target hardware + manuscript — as 1.0 when done, 0.0 when not started, and fraction-of-plan-steps-done when in progress. That gives (1 + 1 + 1 + 0.76 + 0 + 0) / 6 = 3.76 / 6 ≈ 63%. The 0.76 for Phase 3 is 16 of 21 plan steps done in docs/superpowers/plans/2026-09-23-pads-phase3-baselines.md (Tasks 1, 3, 4 complete; Task 2 two-model speculative 0 of 4 steps; Task 5 comparison done except the proposal-timeline update). The PoC voice loop was unplanned extra work and is counted as a bonus, not in the denominator. The honest reading: the safe core is finished and publishable; PADS proper (Phases 4–5) has not started.

Phase checklist

PhaseStateEvidence
Foundations & Go/No-Go falsification suitedone6 of 7 tests executed on the development host (energy tooling pending a root-only command); pause premise holds on AMI speech — experiments/go_no_go_results.md
Phase 1 — safe core (LayerSkip multi-exit training, held-out)doneHeld-out retrain + acceptance, leak check 0/40 — experiments/runs/2026-09-22T18-22-08Z-heldout-retrain/acceptance_heldout.md; plan 23/23 steps in docs/superpowers/plans/2026-09-22-pads-phase2-exit-policy.md Task 3b context
Phase 2 — exit policies (token-confidence vs dialogue-act)doneMatched-compute comparison with cluster bootstrap — experiments/runs/2026-09-22T19-18-00Z-phase2-policy-eval-heldout/comparison.md; plan Tasks 1–4 + 3b + 3c all checked
Phase 3 — baselines (plain, speculative, cascading, LayerSkip config)in progress3 of 4 baselines measured; two-model speculative blocked — experiments/runs/phase3-comparison/comparison.md; plan 16/21 steps
PoC voice loop (wav → STT → trigger → LLM → TTS, CPU-only)doneWorking offline demo, 15.0 s total, 1.34 GB peak — experiments/runs/poc-smoke/POC_NOTES.md, timings.json, demo_run.log
Phase 4 — turn-taking proxy + PADS end-to-endplannedNext semester (Jan – Feb 2027); needs streaming STT partials and real pause audio per experiments/runs/poc-smoke/POC_NOTES.md §Next step
Phase 5 — target hardware + manuscriptplannedNext semester (Mar – Apr 2027); Dell Latitude 5490 validation, energy/thermal protocol, IEEE manuscript

Headline numbers

96.6%
of AMI pauses fit one 7B-class decode step (median pause 1,050 ms vs 164 ms extrapolated step) measured
13.48 → 20.46%
held-out early-exit acceptance at 8 of 16 layers (17.39% counting empties as failures) measured
+21.25 pp
safe-core token-confidence agreement gain over the base model at 50% compute measured
0.859
dialogue-act classifier accuracy on DailyDialog (macro-F1 0.856; proxy labels) measured
46.2
tokens/s plain decoding, host CPU (GGUF Q4_K_M, ±0.18 95% CI) measured
207.5
tokens/s plain decoding, host GPU (±1.23 95% CI) measured
15.0 s
PoC voice-loop total wall time, CPU-only (STT 1.6 + LLM 10.2 + TTS 1.8 s) measured
1.34 GB
PoC peak RSS, whole loop incl. llama-cli subprocess measured

Where each number comes from

NumberSource artifact
96.6% pause fit; 1,050 ms median pause; 164.1 ms stepexperiments/runs/2026-09-21T19-05Z-test2-pause-durations/test2_verdict_7b.txt; method in experiments/go_no_go_results.md Test 2 row
Held-out L8 13.48% → 20.46% (17.39% conservative)experiments/runs/2026-09-22T18-22-08Z-heldout-retrain/acceptance_heldout.md and metrics_extracted.json (baseline_L8 / trained_L8); ledger table in experiments/go_no_go_results.md
Safe-core +21.25 pp at 50% deep [18.45, 24.15]experiments/runs/2026-09-22T19-18-00Z-phase2-policy-eval-heldout/comparison.md; ledger verdict in experiments/go_no_go_results.md (held-out token 0.7598 vs base 0.5473)
Classifier 0.859 accuracy / 0.856 macro-F1experiments/results/dialogue_act_realdialog_metrics.json (accuracy 0.8589, macro-F1 0.8559; majority baseline 0.5494)
Plain CPU 46.18 ± 0.18 / GPU 207.50 ± 1.23 tok/sexperiments/runs/2026-09-22T21-23-19Z-phase3-baselines-5x/summary.json (plain_cpu / plain_gpu per_prompt_mean); comparison in experiments/runs/phase3-comparison/comparison.md
PoC 15.0 s total; 1.34 GB peak; p_deep 0.7401experiments/runs/poc-smoke/timings.json (total_wall_s 14.981), experiments/runs/poc-smoke/demo_run.log (Maximum resident set size 1,407,804 KB), experiments/runs/poc-smoke/POC_NOTES.md
Spot-checks performed this session

Four values were recomputed from source JSONs/logs rather than copied from prose: (1) plain CPU tok/s mean 46.18 and (2) GPU 207.5015 read from summary.json per_prompt_mean means; (3) classifier accuracy 0.8589 read from dialogue_act_realdialog_metrics.json; (4) pause-fit 96.6% read from test2_verdict_7b.txt. Also confirmed: PoC peak RSS 1,407,804 KB in demo_run.log (≈1.34 GB), PoC total 14.981 s in timings.json, trained-L4 acceptance 0.0384 in metrics_extracted.json (matches the 3.84% ledger row, validating that file), and 6 empty-generation skips in the held-out trained L8 eval log (the 15% rate). All matched.

What's blocked / open

Two-model speculative baseline blocked

llama.cpp speculative decoding needs draft and target to share a vocabulary. The only local target GGUF is llama-arch (models/gguf/layerskip-1b-Q4_K_M.gguf); the only local small drafts are qwen2-family. Three options, none taken yet: (1) download a small llama-arch draft GGUF (e.g. TinyLlama) to pair with the LayerSkip target; (2) run the arm entirely in qwen2 (local Qwen2.5-0.5B draft + local Qwen2.5-7B GGUF target) — valid speculation measurement, different model family; (3) drop the arm and treat the self-speculative LayerSkip-config row as the speculation comparator. See experiments/runs/phase3-comparison/comparison.md row 3 note.

15% empty generations open

The held-out adapter produces empty generations on 6/40 eval prompts at every exit layer vs 0/40 for the base model. Cause unknown; the conservative empties-as-failures view (L8 17.39%) misses the ≥ 20% gate. See experiments/runs/2026-09-22T18-22-08Z-heldout-retrain/acceptance_heldout.md.

Energy measurement needs root blocked

RAPL energy_uj reads permission-denied non-root; temperature-only sensors otherwise. Blocked until someone runs the documented sudo turbostat command. See experiments/go_no_go_results.md Test 6 row and experiments/results/test6_energy_tooling_host.log.

No target hardware yet planned

All numbers are development-host numbers. Dell Latitude 5490 validation, sustained thermal/energy protocol, and audio turn-taking all wait for Phases 4–5 next semester. The ledger's "Not measured" lists carry forward in experiments/go_no_go_results.md.

Negative and partial results, stated plainly

Ledger: experiments/go_no_go_results.md (Phase 3 section, 2026-10-28). Comparison: experiments/runs/phase3-comparison/comparison.md. Plans: docs/superpowers/plans/2026-09-23-pads-phase3-baselines.md, docs/superpowers/plans/2026-09-22-pads-phase2-exit-policy.md.