Query Anatomy

Overview

An infer.query.v1 body is validated against a JSON Schema shipped in @syncropel/config@0.3.x. This page walks every field. If you want the canonical schema source, see the schema reference.

Every query has three required parts: input (what to answer), responders (who may answer), fold (how answers combine). The answer_shape is also required and validates the folded result. Everything else is optional with sensible defaults.

Minimal query

The smallest complete query:

{
  "kind": "infer.query.v1",
  "input": { "inline": "What is 2 + 2?" },
  "responders": [{ "kind": "llm", "model": "~sonnet" }],
  "fold": { "function": "best_of" },
  "answer_shape": { "kind": "core.text.v1" }
}

That dispatches one CALL to any Claude Sonnet responder, takes its answer as best_of (since there's only one), and commits the KNOW with body kind core.text.v1.

Top-level fields

Input

Exactly one of three shapes:

{ "inline": <any JSON>, "inline_kind": "core.text.v1" }

or a reference to an existing record by id:

{ "record_id": "3fa8...e2c1" }

or a reference to an entity by canonical ref (any of the 11 Ref.* shapes — see body.kind grammar):

{ "ref": { "kind": "@music.track", "id": "spotify:track:4iV5W9uYEdYUVa79Axb7Rh" } }

When you use inline, the value goes through the canonical-JSON encoder and is hashed into the query's record id. That means identical inline inputs produce identical query ids — useful for memoisation across sessions. The optional inline_kind lets the executor apply body-kind-specific transforms (e.g., pre-embedding for semantic search).

Answer Shape

Validates the folded output, not individual responses. Responders are free to return whatever they return; the executor only checks that the fold's output satisfies answer_shape.

{
  "kind": "core.translation.v1",
  "required_fields": ["body.text", "body.target_language"],
  "schema_ref": "https://example.com/translation.schema.json"
}

• kind (required) — a body.kind. See body.kind manifest for the grammar.
• required_fields (optional) — dot-paths that must be non-null in the folded body.
• schema_ref (optional) — URI of a full JSON Schema. If present, the executor loads it and validates.

If validation fails the executor emits infer.error.v1 with code: "answer_shape_mismatch" instead of the KNOW. The query thread shows both the attempted fold output and the violation.

Responder Predicates

The responders array is an OR-of-classes predicate. Any responder matching any entry is a candidate. Each entry has this shape:

{
  "kind": "llm",
  "model": "~sonnet",
  "trust_gte": 0.6,
  "budget_usd": 0.05,
  "latency_secs": 30,
  "match_level_gte": "L2",
  "age_days_lt": 7,
  "available": true,
  "domain": "translation",
  "capability": "en-es",
  "did": "did:sync:agent:translate-bot",
  "expression": "candidate.last_seen_at > now() - 86400"
}

All fields are optional; each adds an AND-conjoined constraint within a class. Across classes (array entries), it's OR.

The 6 responder selector kinds:

Worked example — two-class predicate:

"responders": [
  { "kind": "llm", "model": "~sonnet", "trust_gte": 0.7, "budget_usd": 0.05 },
  { "kind": "actor", "did": "did:example:alice", "capability": "code-review" }
]

That matches any Sonnet-family LLM with trust at least 0.7 and per-call cost under 5 cents, OR Alice specifically when her capability tag is code-review.

Fold

The pure aggregator. See Fold Functions for each function in detail.

{
  "function": "consensus",
  "weight_expression": "response.trust * response.recency",
  "tie_break": "highest_trust",
  "min_quorum": 2,
  "expression": "..."
}

Enum values:

• function: consensus, best_of, waterfall_first, ensemble_weighted, expression.
• tie_break: highest_trust, highest_confidence, most_recent, lexicographic.

Orchestration

The multi-step policy. See Orchestration Patterns for each pattern.

Default:

{ "pattern": "single_shot" }

Full shape for verify:

{
  "pattern": "verify",
  "primary": [{ "kind": "llm", "model": "~sonnet" }],
  "verifier": { "kind": "llm", "model": "~gpt-4" },
  "tiebreaker": { "kind": "actor", "did": "did:example:alice" },
  "agreement_threshold": 0.85
}

Patterns: single_shot, verify, waterfall, retry_on_low_confidence, escalate, ensemble_with_audit.

Side Effects

Cost and safety envelope. Namespace-level defaults apply if unset (see namespaces).

{
  "reversible": true,
  "idempotent": true,
  "max_cost_usd": 0.40,
  "max_latency_secs": 1800,
  "obligation_resolution": "last_writer_wins"
}

Obligation resolution values:

• last_writer_wins — most recent response's modifiers win.
• fulfills_wins — any response with body.fulfills preempts one with body.cancels.
• validation_error — any conflict aborts the fold with obligation_conflict.

Relevance

Controls the soft-ranking filter after hard predicates have run.

{
  "model": "current",
  "threshold": 0.5,
  "top_k": 3
}

If the filter yields fewer than min_quorum candidates, the executor emits no_relevant_candidates.

Dial

The dial is Syncropel's novelty knob from Frozen Foundation F3 — d in [0, 1], with thresholds at 1/3 (REPLAY), 1/2 (ADAPT), 2/3 (EXPLORE). In an inference query it influences relevance weighting (higher d = more willingness to try novel responders).

Scalar form:

{ "dial": 0.3 }

Adaptive form — expressions evaluated against the candidate pool at query time:

{
  "dial": {
    "replay_if": "pattern_match >= 'L1' && trust >= 0.92",
    "create_if": "pattern_match == null",
    "else": 0.4
  }
}

If replay_if matches, the dial snaps to 0.1 (deep REPLAY). If create_if matches, it snaps to 0.9 (deep CREATE). Otherwise else applies.

Metadata

Opaque string-to-string map you own. Use it for trace IDs, customer IDs, feature flags — whatever you need to correlate queries across your own systems. The executor doesn't interpret it; it just carries it into the KNOW.

{ "metadata": { "trace_id": "trace_abc123", "customer_id": "cust_42" } }

Worked example — single LLM

{
  "kind": "infer.query.v1",
  "input": { "inline": "Translate 'Hello' to Spanish." },
  "responders": [{ "kind": "llm", "model": "~sonnet" }],
  "fold": { "function": "best_of" },
  "answer_shape": { "kind": "core.translation.v1", "required_fields": ["body.text"] },
  "side_effects": { "max_cost_usd": 0.05, "max_latency_secs": 30 }
}

One responder, best_of (trivially picks the only candidate), 5 cent / 30s ceiling.

Worked example — consensus ensemble

{
  "kind": "infer.query.v1",
  "input": { "inline": "Summarise: <paper abstract here>" },
  "responders": [
    { "kind": "llm", "model": "~sonnet" },
    { "kind": "llm", "model": "~gpt-4o" },
    { "kind": "llm", "model": "~gemini-1.5" }
  ],
  "fold": { "function": "consensus", "min_quorum": 2 },
  "answer_shape": {
    "kind": "core.summary.v1",
    "required_fields": ["body.rating", "body.summary"]
  },
  "relevance": { "top_k": 3 },
  "side_effects": { "max_cost_usd": 0.30, "max_latency_secs": 60 }
}

Dispatches to three LLMs in parallel, requires at least 2 to respond, picks the most-agreed-upon answer (keyed on canonical-JSON of the answer body).

Worked example — verify pattern

{
  "kind": "infer.query.v1",
  "input": { "record_id": "a1b2c3..." },
  "responders": [
    { "kind": "llm", "model": "~sonnet" },
    { "kind": "llm", "model": "~gpt-4" },
    { "kind": "actor", "did": "did:example:alice" }
  ],
  "fold": { "function": "best_of", "tie_break": "highest_trust" },
  "orchestration": {
    "pattern": "verify",
    "primary": [{ "kind": "llm", "model": "~sonnet" }],
    "verifier": { "kind": "llm", "model": "~gpt-4" },
    "tiebreaker": { "kind": "actor", "did": "did:example:alice" },
    "agreement_threshold": 0.85
  },
  "answer_shape": { "kind": "core.code_review.v1" },
  "side_effects": { "max_cost_usd": 0.40, "max_latency_secs": 1800 }
}

Claude Sonnet does the primary review. GPT-4 verifies. If they disagree beyond the agreement threshold, Alice gets pulled in as tiebreaker. The KNOW commits whichever answer fold selects.

See also

• Fold Functions — contracts for the 5 aggregators.
• Orchestration Patterns — the 6 multi-step flows.
• Schema reference — the authoritative JSON Schema.
• CEL Expressions — bindings for predicate and fold expressions.
• spl infer CLI — build queries from the shell.