Building the GODMODE-Only Plan

What this is: the real, exact steps to build the first model from the “pure sensitivity” plan — bit allocation driven entirely by the GODMODE Hessian sweep’s real per-bit-width error curves, no --hessian-primary aggregate layered on top (see this project’s own research: the aggregate mechanism’s linear deficit model can’t see the diminishing returns a real curve shows, and empirically produced a worse real KL-divergence sum than the plain pre-Hessian baseline — 40.62 vs 37.45 raw Σ isolated KL, same source data, directly comparable).

The plan

Already built: 04_LANGUAGE_PLANS/HESSIAN_TEST/plan_godmode_ONLY.json — 497 tensors, built from hessian_hybrid_checkpoint.json (the full GODMODE sweep, 496 real per-tensor curves + 1 KL-fallback entry for lm_head, which has no real Hessian coverage at all — see FLAT_MODE.html for why that specific tensor can’t be measured). --lm-head-min-bits 8 and --boundary-min-bits 6 are the only manual floors; everything else is driven by the real curve data directly.

Step 1 — build a resume seed (optional, but saves real time)

A resume seed reuses already-corrected tensors from your prior builds that happen to match this plan’s bit assignments, so the real, slow per-tensor Hessian correction only has to run on tensors that aren’t already done at the right width.

cd /Users/hghelab/ai-employee-build/projects/MLX_OptiQ/qwen38-27b-heretic-ara/scripts/yaqa_port

# Check real coverage first, copies nothing:
python3 build_resume_seed.py \
    --plan ../../04_LANGUAGE_PLANS/HESSIAN_TEST/plan_godmode_ONLY.json \
    --output /Users/hghelab/.mtplx/models/Qwen3.8-27B-heretic-ara-YAQA-GODMODE-v1.yaqa_resume \
    --dry-run

# Actually build it:
python3 build_resume_seed.py \
    --plan ../../04_LANGUAGE_PLANS/HESSIAN_TEST/plan_godmode_ONLY.json \
    --output /Users/hghelab/.mtplx/models/Qwen3.8-27B-heretic-ara-YAQA-GODMODE-v1.yaqa_resume

Already done once (2026-10-01): 287 of 497 tensors (57.7%) reused from three real prior builds, scored by real measured correction quality (err_yaqa), not just whichever folder was scanned first. 11.61 GB.

You don’t need to delete anything to re-test this

The script refuses to run (outside –dry-run) if –output already exists, specifically so it never silently merges into or overwrites a real resume directory. To test it fresh, point –output at a new path — the existing GODMODE-v1.yaqa_resume isn’t in the way, and deleting it would just throw away 11.61 GB of real, reusable, already-corrected tensors for no benefit.

Step 2 — run the real build

cd /Users/hghelab/ai-employee-build/projects/MLX_OptiQ/qwen38-27b-heretic-ara/scripts/yaqa_port

./run_full_yaqa.sh /Users/hghelab/.mtplx/models/Qwen3.8-27B-heretic-ara-YAQA-GODMODE-v1 \
    --source /Users/hghelab/.mtplx/models/Qwen3.8-27B-heretic-ara \
    --plan ../../04_LANGUAGE_PLANS/HESSIAN_TEST/plan_godmode_ONLY.json \
    --n-calibration 4 --incoherence none

run_full_yaqa.sh’s own real default --source is already this exact path — confirmed directly in its source (SOURCE_DIR="/Users/hghelab/.mtplx/models/Qwen3.8-27B-heretic-ara") — passing it explicitly here is just for clarity, not required.

It auto-detects <output>.yaqa_resume next to the output path you give it — never pass --resume-dir yourself for a normal run.

One rule with no exception

Never call 05_full_model_quantize.py directly for a real multi-batch run — only run_full_yaqa.sh, which runs each memory-bounded batch as its own fresh process. A direct call risks a real, confirmed Metal resource-handle crash.

Step 3 — watch it live

./monitor.sh /Users/hghelab/.mtplx/models/Qwen3.8-27B-heretic-ara-YAQA-GODMODE-v1.yaqa_resume/logs/batch_1.log

Also serves a live HTML dashboard (generate_dashboard.py) with real per-tensor progress and real CPU/memory stats.

What success looks like

After the build — verifying it

MODEL="/Users/hghelab/.mtplx/models/Qwen3.8-27B-heretic-ara-YAQA-GODMODE-v1"
mtplx inspect "$MODEL" --json
mtplx forge verify "$MODEL" --stamp --json