multi-agent-collaboration.md

Multi-Agent Collaboration

Single-pass role pipelines drift: one model imagines, judges, and aligns itself, then ships its own opinion as evidence. Multi-agent collaboration replaces the pipeline with a bounded, adversarial, parallel deliberation inside each cycle — without giving any LLM the power to override deterministic gate outcomes.

This file is the v2.0 upgrade to the Rainforest core loop. It sits on top of the same forest state and weather routing; it does not replace the gate battery or the result classification.

Role Cast

Every step of a cycle is attributed to a named role. Adding or replacing a participant is a role registration, not a rewrite.

Role	Responsibility	Notes
coordinator	Orchestrates the cycle; sole writer of forest state.	Stays the arbiter of stop conditions, debate rounds, and merge order.
meteorologist	Routes attention via weather.	Deterministic; no LLM.
gardener	Mines hypotheses for one tree.	LLM. v2.0 runs one gardener per allocated tree.
inspector	Pre-nursery judgment: logic, novelty, leakage suspicion.	LLM.
diverger	Expands candidate mechanisms to alternative formulations.	Handoff to a watching session.
examiner	Runs the full gate battery on a compiled candidate.	Deterministic.
aligner	Mechanism-evidence alignment scoring before promotion.	Handoff to a watching session.
skeptic	Adversarial red-team critique of hypotheses and fruit-candidates.	LLM. Different model family than the gardener.

The Deliberation Loop (per allocated tree)

gardener mines N hypotheses
  -> skeptic critiques each {weaknesses, overfit_risk, alt_explanations,
                             lookahead_suspicion, verdict, severity}
  -> gardener revises survivors        (<= max_debate_rounds, default 2)
  -> diverger expands survivors        (handoff)
  -> inspector judges -> nursery 7-check -> sprouts
  -> examiner runs the full gate battery
  -> for each fruit-candidate (all hard gates pass):
       skeptic SECOND challenge -> attached to risks, emits skeptic_challenge event,
                                   NEVER alters the gate_record or classification
  -> aligner G7                        (handoff)
  -> classify

Per-cycle, the coordinator runs one deliberation per allocated tree and merges the per-tree results in a deterministic tree_id-sorted order. Execution may be concurrent or independent-sequential; reproducibility is the contract.

Anti-Self-Favoring: The Different-Model Skeptic

A skeptic that shares the gardener's model family will tend to share its blind spots — and bless its own mechanisms. The lever:

• The skeptic adapter structurally refuses to run when its model family equals the gardener's (raise before any API call).
• The system prompt is adversarial and distinct from the inspector's judge prompt — it assumes the candidate is overfit, spurious, or look-ahead-contaminated until proven otherwise.

Pick the pair before the cycle and freeze it in config (for example: Kimi gardener, DeepSeek skeptic). Mixing within a family does not count.

Bounded Debate (Hypothesis Stage)

Per round, for each surviving mechanism:

Verdict	Severity	Action
reject	>= cull_severity	Cull. Drop the mechanism from this cycle.
reject	< cull_severity	Treat as revise.
revise	any	Gardener re-mines a replacement using the skeptic's weaknesses as context.
proceed	any	Keep the mechanism unchanged; it exits the debate (not re-critiqued).

The debate ends when every survivor is proceed or max_debate_rounds is exhausted. After exhaustion, remaining non-proceed survivors are kept (only reject-at-or-above-cull_severity is destructive).

max_debate_rounds == 0 skips the debate entirely; mined mechanisms flow directly into divergence and inspection. With a single gardener and zero rounds, the loop reproduces the v1 single-pass pipeline.

Recorded-Only Second Challenge (Fruit Stage)

When all hard gates pass, the skeptic challenges once more before the alignment handoff. The verdict is:

• Recorded as event (skeptic_challenge) and attached to the ResultRecord.risks field.
• Not allowed to alter the gate_record or the classification.

Rationale: if an LLM cannot upgrade a stub to a fruit, by symmetry it must not downgrade a deterministic fruit to a leaf. The second challenge raises rigor as documented adversarial scrutiny, not as a hidden veto. A skeptic that wants to override a passing gate must instead propose an explicit, documented gate amendment in the next cycle's grounding.

Parallel Competing Gardeners

The meteorologist allocates trees; the coordinator dispatches one gardener per allocated tree with distinct temperature/style hints drawn from a temperature_spread (for example [0.7, 1.0, 1.3]). This widens the search frontier without changing the validator or the gate battery. Concurrency is permitted; the contract is a deterministic merge of per-tree results by tree_id, so the same seeds always yield the same forest.

Rigor Invariants (do not break)

1. The validator runs pre- and post-cycle, unchanged.
2. Classification still requires a complete GateRecord with execution_mode == tool_executed. A debated stub_result still cannot become a fruit; a skeptic reject on a passing fruit-candidate does not veto it.
3. The coordinator is the sole writer of forest state. Per-tree deliberation is pure: it reads the forest and returns candidate data for the coordinator to merge.
4. Skeptic model family must differ from gardener model family. Skeptic or LLM unavailability is a hard failure; no silent fallback.
5. Backward compatibility: zero debate rounds plus a single gardener reproduces the v1 cycle.

Event Vocabulary

Every step emits an agent-attributed event to events.jsonl. v2.0 adds three action types on top of the v1 vocabulary; they double as the replay source for an office-scene visualization.

Action	Emitted by	Carries
gardener_parallel_dispatch	coordinator	{trees, dispatch: {tree_id: temperature}}
debate_round	skeptic	{round, in, culled, revised, proceeded}
skeptic_challenge	skeptic	`{kind: hypothesis

Suggested Config Block

deliberation:
  max_debate_rounds: 2
  skeptic_model: deepseek      # must differ from gardener model family
  gardener_model: kimi
  cull_severity: high          # cull on skeptic verdict=reject AND severity>=this

parallel_gardeners:
  max_concurrent: 4
  temperature_spread: [0.7, 1.0, 1.3]

What v2.0 Deliberately Does Not Do

• It does not evolve research trajectories across cycles (mutation/crossover of full hypothesis→AST→code paths). That is a v2.x candidate.
• It does not add a downstream factor-combination layer (ensemble of promoted factors). That is a separate strategy-synthesis layer above the framework.
• It does not visualize the deliberation. The event stream is the substrate; a renderer is a future, separate concern.