MORNING_REPORT2.md

☀️ Report 2 — soul2 SHIPPED (2026-06-19)

TL;DR: The 250-example multi-facet soul2 healed, scored GREEN, and shipped. Code held exactly (HumanEval-164 = 116/164 = 70.7%, identical to soul v1) while the corpus added science · perfumery · deep-security · pentest · the full design-movement range. The per-facet flywheel confirmed a known limitation (3-bit long-generation degeneration), not a regression.

Verdict: GREEN ✓ — shipped

• adapters-soul2 uploaded to HF (commit 859ed1a).
• HumanEval-164: 116/164 = 70.7% — exactly soul v1. The 250-corpus (vs v1's 43) preserved code capability while adding ~6× the breadth. No overfitting collapse despite train loss 0.093 / val 1.22.

The honest flywheel finding (long-gen degeneration, NOT a soul2 regression)

Per-facet self-generation scorecard (partial — stopped once the signal was unambiguous):

• design: 0 elite / 24 reject — fully degenerate on long HTML.
• code: 2 elite / 13 reject — and the tell is decisive: the 2 elite gens were SHORT (~1258 ch); the 13 rejects were LONG (5–7K ch) and collapsed into repetition + corrupted tokens (UTF.FF9B-style garble).

This is the documented 3-bit Computation Collapse on long generation (research/swappable_adapters_sota.md, rounds 4–5): the model degenerates when it rambles — not a soul2 regression (soul v1 degenerated identically). The gold/soul is elite — the short elite gens prove it, and the masters-corpus is sound. The collapse is a serving/decoding artifact of 3-bit on long output, which is exactly why the soul is masters-GOLD, not self-gen.

The fixes (research-backed, queued):

1. Saliency-dynamic quant (#59) — protect the salient + early-layer experts at 4-bit+, the rest at 3-bit. The research is blunt that Computation Collapse is fixable only by mixed-precision on the critical experts, not a decoding trick.
2. Shorter CoT — scripts/08_think_proxy.py's reasoning budget (already built): short = elite, long = collapse.

Corpus

• soul2: 250 masters-gold (8 facets) — SHIPPED.
• +266 specialty gold ready for the next round (the factory's modules): fullstack-AI-DS (8 areas), gamedev (Unreal/Unity/Godot/Flutter/patterns/gfx-net), legacy (COBOL/Java/PHP/Perl/VB + modern Java 21/PHP 8.4/.NET 8), cyber-core + pentest (purple-team), science, perfumery, factory-router.

Next

Heal the specialty adapters (GPU now freed): core-soul-v3 (non-code + science + perfumery + deep-security + pentest + router) first, then the swappable code modules (fullstack / gamedev / legacy), using MoE-Sieve placement (top-25% routed experts → ~70% smaller adapters) + iw-SFT — both from the research.

Provenance

9-round June-2026 SOTA scan → research/swappable_adapters_sota.md + IMPLEMENTATION_PLAN.md. The verdict, the collapse diagnosis, and the validations of our core calls (prune > merge, verifier-mesh over self-reward, on-policy-KD now industry-standard, the DSA top-2048 = our heal limit) are all recorded there.