troupe-standup/docs/sota-research-2026-10-25.md

SOTA Research Spoke — 2026-10-25

7-spoke fan-out across our operation. Full transcripts in /root/.hermes/cache/delegation/live/deleg_1a756748/. Ranked by impact ÷ effort.

Tier 1 — do now (low effort, big cost/quality)

1. Route auxiliary.* off deepseek-v4-pro. We only set auxiliary.vision. Every other side task (compression, title_generation, approval, background_review, triage_specifier, kanban_decomposer, profile_describer, goal_judge, curator, monitor, moa_*) runs provider: auto = bills deepseek-v4-pro. Point them at free/cheap (nous laguna-xs, nemotron-3.5-lightning, gemini flash). Docs: ~85% savings on background tokens. hermes model → "Configure auxiliary models".
2. Set delegation.model/provider to a cheap model — subagent fan-outs currently inherit deepseek-v4-pro at full price.
3. hermes update to v0.22.5 (we're on v0.22.0-canary, ~700 PRs behind; v0.23 imminent). Gains: hermes mcp serve, pluggable delegation backends, auxiliary.voice_chat, security bumps. NOTE: smart_model_routing was REMOVED upstream — our config's smart_model_routing.enabled:false is dead leftover.
4. Fix dead free-tier models. solar-pro4 was DROPPED from Nous Portal Oct 1. Replace/augment free tier: Nous $0 plan adds LongCat 2.5 Preview (1M ctx, open), Step 5 Preview (coding/vision), Solar Mini 4 (35B/3B, 512K), MISSINGNO, Ling 3.0 Flash. OpenRouter free adds North Mini Code (Cohere, Apache 2.0, beats Nemotron 3 Super 120B on coding AA 36.5), Nemotron 3.5 Lightning (~670 tok/s, 1M ctx), MiniMax M3.1 Flash (LLM-Stats #29 vs Lightning's #22).
5. STT: faster-whisper → NVIDIA Parakeet TDT 0.6B (CC-BY-4.0, 6.32% WER vs large-v3 7.44%, ~30× realtime CPU, no silence hallucination). via onnx-asr/sherpa-onnx. (Alt: Moonshine v2, MIT.)
6. Vision: keep deepseek-flash as default (it's natively multimodal at $0.15/M ≈ 60-70× cheaper than gemini-4-argon per image). Drop argon for routine reads; keep for genuinely hard ones only. Free tier: Nemotron 3.5 Super VL (text+image+video). Local: replace moondream with Qwen3-VL 2B (0 false-positive hallucination vs moondream 22/126) or MiniCPM-V 4.6.

Tier 2 — high value, needs a build

7. TTS: add Qwen3-TTS 0.6B (Apache 2.0) — 3s zero-shot voice cloning, instruction-driven emotion/prosody, 97ms streaming, ~2GB. Keep Phonon as fast default; Qwen3-TTS for cloning+emotion. (Fallback: Chatterbox-Turbo MIT for emotion exaggeration.)
8. Full-duplex voice — RESOLVES the Dessa POC. The pick is NVIDIA PersonaPlex 7B (MIT code, Moshi-based): full-duplex + voice cloning + role prompts, runs on ONE 24GB card — cheaper than NemotronLabs VoiceChat 11B (needs A100, or FP8/INT4 quantize). Hertz-dev is a BASE model (no instruction following) — drop it. Dessa stays closed. Deploy PersonaPlex on a RunPod 24GB community card when one frees.
9. Leathermint: migrate Phaser 3 → Phaser 4 (MIT, same family). WebGL2 renderer (4 vs 6 verts), GPU-driven batching ("million sprites, up to 100× faster"), TilemapGPULayer (whole iso floor as one quad), automatic WebGL context restoration (fixes the mobile black-screen bug). Add Web Worker + OffscreenCanvas for the sim loop.
10. TM AI: hybrid architecture. LLM (deepseek-flash) for persona/banter ONLY; move selection = deterministic scorer + MCTS/rollouts (pure LLM CoT universally loses to MCTS). Add an AutoHarness-style legality harness (validate every move against InputResponse rules) to kill illegal-move/stuck-game bugs. Extend USE_STRATEGY_ENGINE code-policy for clear-cut states (zero LLM cost on those).

Tier 3 — strategic

11. Ralph loop → Harness² contrastive. Keep CHECK exit-code as verifier, but add Propose-Probe-Compose (two candidate edits per failure: one invariant + one domain; probe on SAME task; compose only contrast-favored). Add Self-Harness regression validation (held-in + held-out) so reflection doesn't overfit one CHECK. Structure harness as file-level components.
12. Model routing diversification. GPT-6.1 Sol $2/$10 ties Opus 5.5 on DeepSWE at 1/7 cost. Gemini 4 Argon intro $2/$10 (DeepSWE leader 77.9%, lowest hallucination 15%) — good for hard reasoning. MiMo-V2.6-Pro (MIT, ~$0.43/$0.87) = open-weight agentic-coding workhorse. ClawRouter plugin (blockrun/auto/eco) auto-routes by complexity (~84-98% savings). OpenRouter :free = 50 req/day/account (multiply accounts for $0).
13. Creative gen. FLUX.2 [dev] (32B) / [klein] (4B Apache) for image; Qwen-Image for in-image text. YuE/YuE2 for music (replaces Suno, local on 4090). Wan 2.2 / HunyuanVideo 1.5 for video. RunPod A40 for renders; FAL/Replicate/DeepInfra for burst.

Key corrections to our assumptions

• DeepSeek V4.1 Flash (deepseek-flash) is now natively multimodal (image input) — we already wired it for vision correctly.
• Prompt caching is keyed to the serving model — our long fallback chain (umans/kimi/orcarouter/nous) breaks cache on any failover (75-90% discount lost). Consider trimming it.
• Deep Infra is the open-weights price floor (Qwen3.8-Max $1.65/$4.95 vs $2/$6 elsewhere).