reasoning-ensemble

name	reasoning-ensemble
description	FLOSSI0ULLK Inline Reasoning Ensemble: use the Router for cheap mode classification (pass_through, single_strong, ensemble), and the Synthesizer for full ensemble deliberation on architectural decisions, multi-file refactors, ADR-class moves, or any substantive decision where being wrong has rollback cost. Also use when the user explicitly asks for debate, a second opinion, or multi-model/lens analysis.
license	AGPL-3.0
metadata	{"version":"0.1.1","date_landed":"2026-05-18","source":".claude/skills/reasoning-ensemble/SKILL.md","related_research":["FLOSS/docs/research/2026-05-17-inline-reasoning-ensemble.md","FLOSS/docs/research/2026-05-18-metaharness-unification.md","FLOSS/docs/research/2026-05-16-mdash-cfis-architectural-transfer.md"]}

Reasoning Ensemble Skill

Invoke selectively. This is not "debate every prompt"; it is debate for substantive reasoning steps where blindspots matter.

Router Only

When unsure whether a prompt deserves ensemble deliberation, ask the Router:

python C:/~shit/FLOSS/packages/reasoning_ensemble/router.py "<prompt>"

Expected JSON:

{
  "mode": "pass_through | single_strong | ensemble",
  "reason": "<one-sentence justification>",
  "confidence": 0.0,
  "bias_applied": "tier4_similarity | force_flag | fallback_on_error | null",
  "duration_seconds": 0.0
}

If bias_applied is tier4_similarity, a past Tier-4 divergence on a semantically adjacent prompt forced ensemble mode.

Synthesizer

When the Router returns ensemble, the user tags --debate, or a substantive decision needs a second lens:

python C:/~shit/FLOSS/packages/reasoning_ensemble/synthesizer.py "<prompt>"

The Synthesizer writes a durable draft under .agent-surface/reasoning/ensemble/ and appends a global Action to .agent-surface/activity.jsonl.

Expected output includes:

• tier: tier1, tier2, or tier4
• voter_responses: per-voter raw responses and reasoning
• clusters: embedding-based response clusters
• largest_cluster_fraction: dominance signal
• minority_coherent_voters: dissent that passed the coherence guard
• final_synthesis: digest with consensus and named tensions

When To Invoke

Always ensemble:

• Architectural proposals and ADR-class moves
• Multi-file refactors
• Decisions with blast radius beyond a single function
• Anything where a wrong answer has rollback cost
• Anything the user tags --debate

Default to single_strong:

• Single-file edits
• Code reads and explanations
• Routine synthesis
• Format conversions
• Memory recall

Use pass-through:

• File reads, directory listings, and grep results
• Status checks
• Acknowledgments

Limits

• The v0.1 voter prompt still needs calibration; early voter responses included low-signal boilerplate.
• Tier-4 similarity bias only becomes useful after enough Tier-4 entries exist in .agent-surface/reasoning/activity.jsonl.
• Local models compete for GPU memory. On a 16GB system, calls may serialize.
• Synthesis drafts are not canon. Promotion still needs human/ADR/source-chain review.

Current Voter Pool

Local Ollama voter pool: phi4-mini, llama3.2-3b, granite-code-3b, and qwen2.5-coder-3b. The default Router is phi4-mini:latest; set FLOSS_ROUTER_MODEL=gemma3:12b-it-qat if the stronger fallback is needed.

References

• FLOSS/docs/research/2026-05-17-inline-reasoning-ensemble.md
• FLOSS/docs/research/2026-05-18-metaharness-unification.md
• FLOSS/docs/research/2026-05-16-mdash-cfis-architectural-transfer.md
• FLOSS/docs/research/2026-05-15-working-todo-list.md
• FLOSS/packages/reasoning_ensemble/router.py
• FLOSS/packages/reasoning_ensemble/synthesizer.py