The synthesis

What the synthesis computes

The synthesis is built by buildOrchestratorSynthesis in src/lib/synthesis.ts. It reads the structured JSON output each agent returns and runs six deterministic computations - no LLM call, no sampling, no randomness:

1. Consensus. Whether all agents in the final round landed on the same stance. Computed as stanceTally.length === 1 - true only when every agent emits an identical stance string.
2. Stance tally. A breakdown of how many agents held each stance, sorted by count descending. Stances come from each agent's stance field in the last round's output; the tally groups agents by the exact string they reported.
3. Top recommendation. The recommendation text from the highest-confidence agent in the final round. On ties in confidence, the alphabetically first agent name wins - so the result is always deterministic.
4. Shared risks. Risks raised across all rounds, deduplicated in appearance order. A risk counts as shared when it appears in more than one agent's output across any round.
5. Open and deferred questions. Open questions are deduplicated from the final round's packet. Deferred questions are collected and deduplicated across all rounds - questions the panel flagged but did not resolve.
6. Overall confidence. Each agent reports low, medium, or high. These are mapped to numeric ranks (1, 2, 3), averaged across all agents in the final round, then rounded back to a level: average ≥ 2.5 → high; ≥ 1.5 → medium; below that → low.

Reading the synthesis card

Every synthesis surfaces as a card - in the terminal at the end of a run, in synthesis.md in the run directory, and in the machine-readable synthesis.json. Here is what each row means:

The verdict word - Fix now above - comes from the preset's --decision vocabulary, not from the model. An agent outputs a stance string, and that string must match one of the words the preset defines. The model never invents a verdict word of its own.

Confidence is reported as a level: low, medium, or high. It is a rounded average of what the agents reported - never a fabricated decimal like 0.78.

The consensus row shows how many agents agreed on the same stance in the final round. When all agents converge, consensus is true; when they split, the tally shows each stance and who held it.

The "Still open" row lists questions that were raised but not resolved by any round. These are the deferred questions you may want to answer before acting on the recommendation.

Determinism in depth

The synthesis has no LLM step. Given identical agent outputs, it will always produce identical synthesis fields. The properties that make it deterministic are:

• No sampling. All computations are arithmetic or string operations on the agent JSON. There is no temperature, no token generation, no model call.
• Alphabetical tie-breaking. When two agents report the same highest confidence, the recommendation is taken from the agent whose name sorts first alphabetically. The output is stable across re-runs.
• Deduplication by first appearance. Risks, objections, open questions, and deferred questions are deduplicated in the order they first appear - identical strings are dropped, order is preserved.
• Preset-controlled vocabulary. The stance string each agent reports must come from the preset's decision vocabulary. The synthesis surfaces that string as-is; it never renames or reinterprets it.

The synthesis fields

synthesis.json in the run directory contains the full machine-readable output. The top-level fields map directly to what was computed:

Where next

• Quickstart - run a ready-made swarm and read your first synthesis end to end.
• CLI reference - every command, flag, preset, and run artifact.
• Architecture - how the pipeline flows from agents to synthesis.