Traigent optimize composite knobs

Traigent Composite Knobs

When to Use

Requires traigent>=0.13.0 (the knobs API is present on all current SDK releases).

Use this skill when:

• Choosing a composite pattern from traigent.knobs.patterns
• Declaring cascades, routers, ensembles, loops, ReAct-style tool steps, verifier loops, mixture-of-experts, or fallback shapes
• Wiring execute_composite with StageRunner or LoopBodyRunner
• Returning (output, metrics) from an optimized function so composite telemetry rides the measures channel
• Explaining which values are tuned, calibrated, fixed, or observed

For exact factory signatures and per-pattern caveats, read references/pattern-catalog.md. Before adding a router or cheap-to-strong cascade, read references/router-fitness.md; it distinguishes pre-dispatch routing from post-output escalation and requires signal, headroom, comparator, economics, and calibration checks. For adaptive RAG and bounded refine specializations, read references/advanced-recipes.md.

Quick Start

The canonical wiring is:

1. Declare a pattern factory such as binary_cascade, router, self_refine, or moe.
2. If the factory was called with an explicit members= dict, spread those Knob declarations into the surrounding configuration space. Most patterns (including binary_cascade with no members= argument) return an empty .members dict — spreading it is a no-op. Threshold CVARs like the cascade gate must be declared explicitly in configuration_space as tuned variables (or passed as literals in calibrated_values); they are NOT auto-populated from .members.
3. Wire every referenced stage with StageRunner for sampling/voting positions or LoopBodyRunner for loop bodies.
4. Inside the decorated function, call execute_composite with the pattern's .structure, current tuned config, live calibrated CVAR values, and any signals or predicates.
5. Build ordinary numeric metrics, call merge_composite_measures(metrics, run), and return exactly (output, metrics).

Use this portal-compatible return shape from composite_telemetry.py:

from pathlib import Path

import traigent
from traigent.knobs.patterns import binary_cascade
from traigent.knobs.runtime import StageRunner, execute_composite
from traigent.knobs.telemetry import merge_composite_measures

GATE = "router_margin_threshold"

# On the tuple-return path the function scores itself (see the note below),
# so it needs the expected answer in scope. _EXPECTED stands in for your
# dataset's expected_output; in real code look it up per example.
_EXPECTED = "STRONG"

# Tiny self-contained dataset so this block runs as-is.
Path("eval").mkdir(exist_ok=True)
Path("eval/composite_demo.jsonl").write_text(
    '{"input": {"text": "route me"}, "expected_output": "STRONG"}\n'
)

COMPOSITE = binary_cascade(
    "answerer",
    base_stage="cheap",
    expert_stage="strong",
    threshold=GATE,
)


def _stage(outputs: list[str]) -> StageRunner:
    return StageRunner(
        run=lambda _item: list(outputs),
        key_fn=lambda x: x,
        samples=len(outputs),
    )

@traigent.optimize(
    eval_dataset="eval/composite_demo.jsonl",
    objectives=["accuracy"],
    configuration_space={
        "variant": ["cheap", "strong"],
        # Declare the threshold as a tuned variable so params[GATE] resolves.
        # The optimizer searches discrete margin values; the winning value
        # doubles as the live calibrated_value passed to execute_composite.
        GATE: [0.3, 0.5, 0.7],
    },
    default_config={"variant": "cheap", GATE: 0.5},
)
def answer(text: str) -> tuple[str, dict[str, float]]:
    cfg = traigent.get_config()
    params = dict(cfg)
    run = execute_composite(
        COMPOSITE.structure,
        {"cheap": _stage(["weak-guess"]), "strong": _stage([_EXPECTED])},
        config=params,
        calibrated_values={GATE: params[GATE]},
    )
    # The composite_* keys become per-trial measures on the wire.
    metrics = {"accuracy": 1.0 if str(run.output) == _EXPECTED else 0.0}
    merge_composite_measures(metrics, run)
    return str(run.output), metrics

The evaluator recognizes exactly a two-item tuple (output, metrics) where metrics is numeric and identifier-keyed. Other return shapes are not unpacked.

With the tuple return, use the BUILT-IN evaluator (expected outputs in eval_dataset, no custom scoring_function): a custom scoring_function or 3-arg metric_functions is currently NOT invoked with the unpacked prediction on this path, and every trial silently scores accuracy=0.0 (known SDK issue). Diagnostic tell, by SDK version: on <= 0.21.3, uniform zero accuracy next to a sane built-in score means scoring wiring, not a bad agent; score mirrors the primary objective on SDKs after 0.21.3 (see version-matrix: score-relocation), so it is also 0.0 here and the sane built-in value is relocated to exact_match_default — check that key instead, and look for the run-level "custom scoring_function defines the 'accuracy' objective" log line. Escape hatch: if you need custom scoring on this path, compute the metric inside the function and return it in the tuple's metrics dict (as the Quick Start's accuracy does) — do not wire a scoring_function and wonder why it never fires. If neither works for your case, stop and surface the SDK limitation to the user rather than iterating.

Before any paid run, assert the gate CVAR is actually resolvable — an undeclared threshold is a per-trial KeyError after money is spent:

# contract: skip
assert GATE in answer.configuration_space or GATE in calibrated_values

WHEN-TO-USE DECISION TABLE

Pattern	Agent shape it fits	What it tunes	Key CVARs
binary_cascade	Cheap base arm escalates to one expert arm on low vote margin.	Arm-level tuned params via base_tuned_params and expert_tuned_params; any .members.	One margin threshold.
n_cascade	Ordered escalation across three or more stages.	Per-stage tuned_params; any .members.	One margin threshold per non-terminal stage.
self_consistency	Sample one stage k times and majority-vote.	stage_tuned_params; cardinality if resolved from config; any .members.	Optional accept_threshold; cardinality if calibrated.
best_of_n	Sample one stage k times and choose with a judge.	Generator and judge tuned params; cardinality if resolved from config; any .members.	cardinality if calibrated.
self_debug	Retry a stage until an external predicate passes.	stage_tuned_params; state shape; any .members.	None required; predicate is runtime code.
self_refine	Refine threaded state until a calibrated signal accepts.	stage_tuned_params; state and signal input shape; any .members.	One signal threshold.
react_tool_loop	ReAct-style one-tool-step-per-iteration loop.	Planner/tool/failure-handler tuned params; any .members.	tool_confidence_min.
verification_gate	Generate, verify, revise loop with verifier pass score.	Verifier style, question count, model, context, revision policy; any .members.	verifier_pass_threshold.
moe	Committee of distinct experts, aggregated by vote or judge.	Per-expert tuned params and optional judge params; any .members.	Optional vote accept_threshold.
router	Pre-dispatch to exactly one arm using left-to-right adequacy signals.	Per-arm tuned_params; signal input declarations; any .members.	One threshold per signal gate.
fallback	Ordered post-cascade fallback on no_accept or low margin.	Per-arm tuned_params; any .members.	One margin threshold per non-terminal arm.

The table above keys off agent shape. Before wiring a knob, also check the failure mode it fixes — shape tells you which pattern is expressible, failure mode tells you whether it helps:

• repair → erroring/malformed outputs
• self-consistency / best-of-n → output instability
• retrieval / similar-fewshot → unseen patterns / missing examples
• chain-of-thought / plan → multi-step reasoning

Here, repair = self_debug/self_refine in the table above; retrieval, similar-fewshot, chain-of-thought, and plan are structural knobs (retriever, fewshot_selector, prompting_strategy, generation_path) — see traigent-optimize-config-space and its references/structural-spine.md.

A knob wired in blind adds cost and can lower the score.

⚠️ Stochastic knobs vs. exact-match metrics: temperature>0 and self-consistency trade determinism for exploration; on a frail exact-match/case-sensitive scorer they can turn a correct deterministic answer wrong. Pin temperature=0 there — provider APIs default to a nonzero temperature (typically 1.0) when it is unset, so unset ≠ deterministic — and open it up only when the scorer tolerates surface variation (a validated semantic/execution-match equivalence class).

Members and the Configuration Space

A factory returns a CompositeKnob declaration bundle: .structure is the IR root, .members are ordinary member Knob declarations, .provenance records the pattern name plus a canonical param hash, and .telemetry_names lists standard measure names.

Spread .members into the surrounding configuration space only when the factory was called with a non-empty members= argument. Factories called without members= (or with members=None) return .members = {} — spreading it is a no-op and will NOT populate threshold keys into configuration_space. Threshold CVARs (e.g. router_margin_threshold for binary_cascade) are not auto-populated by .members; they must be declared explicitly in configuration_space (to tune them) or supplied as literals in calibrated_values (to fix them). Omitting them from both causes params[GATE] to raise KeyError on every trial. Member bindings that ARE present remain Tuned, Calibrated, or Fixed; the composite only references names and declares control flow.

The vocabulary, in one line: TVARs are searched, CVARs are calibrated, policies govern control flow, KPIs/objectives score outcomes. For domain-specific structural knob vocabularies (text2SQL, RAG/multi-hop QA), cross-reference traigent-optimize-config-space and traigent-optimize-config-space/references/structural-spine.md -- this skill extends that structural vocabulary with composite control flow; it does not replace it.

Telemetry

<!-- PROTECTED -->

Composite telemetry is content-free and starts in run.measures.

<!-- /PROTECTED -->

Kind	Standard names
Post-cascade (binary_cascade, n_cascade, fallback)	escalation_rate, stage_selected, gate_margin_pass_rate
Router/pre-cascade	route_selected, dispatch_signal_margin, gate_signal_adequate
Ensemble (self_consistency, best_of_n, moe)	vote_agreement, vote_margin, candidates_evaluated, candidates_excluded
Loop (self_debug, self_refine, react_tool_loop, verification_gate)	iterations_used, stop_reason

Use merge_composite_measures(metrics, run, prefix="composite") to flatten finite numeric telemetry into the existing measures channel. Structured maps are flattened by gate index. Non-numeric enum values such as stop_reason remain observable on run.measures but are not copied to numeric measures.

Wire boundary (deliberate): composite structure, members, and provenance are SDK-local metadata — nothing about a composite crosses the backend wire except these numeric composite_* measures riding the EXISTING measures channel. Do not invent new request/response fields for composites; a future composite wire summary is a Schema-first change (TraigentSchema), not an SDK patch.

<!-- PROTECTED -->

Claim Scope

Composite end-to-end metrics are OBSERVATIONS from the evaluated trials. Per-variable calibration certificates are the only procedural claims. Winner-level product copy should say Calibration-backed winner (client-attested). Do not use a bare Certified winner label, and do not promise future behavior from a composite run.

<!-- /PROTECTED --> <!-- Reserved: managed longitudinal-guidance region. Step-level edits must not write here. --> <!-- SLOW_UPDATE --> <!-- /SLOW_UPDATE --> <!-- INTERACTION_POLICY v1 (synced — do not edit inline; edit docs/shared/interaction-policy.v1.md) -->

Traigent Interaction Policy

Track an interaction profile and adapt to it. Persona (stable): control=delegate|guided|inspect, expertise=se|ds|unknown. Mood (this session): pace=execute|balanced|explore. Default when unknown: guided,se,balanced. Infer from explicit user statements first, then recent behavior; an explicit correction wins immediately. Never store or send this profile anywhere by default.

Fetch the live profile (when available)

At session or skill start, if a configured Traigent client is available, seed the profile from the backend with the skill name:

policy = None
try: policy = await client.get_interaction_policy(skill="<this skill>")
except Exception: pass

Treat the returned profile as the STARTING seed: its control/expertise/pace axes plus question_budget, options_max, and jargon_level replace the static defaults below. Explicit user corrections in-conversation ALWAYS override the seed. If the call is unavailable or fallback_policy="static_v1", simply use the static defaults below; the SDK already fails soft.

• Always be concise.
• Match terminology to expertise. For se: plain engineering words; define each Traigent or statistics term once in plain language (no Bayesian / variance-decomposition / Pareto jargon unless asked). For ds: compact optimization and statistical terms are fine.
• Presenting options: show at most 3, mark exactly one Recommended, and give one short persona-appropriate trade-off per option.
• Autonomy. For delegate or execute: pick the recommended reversible action and proceed, asking only at hard gates. For guided: offer options with a recommendation at the key decisions. For inspect or explore: give brief rationale or evidence before asking, and ask before branch choices.
• Hard gates — always confirm regardless of persona: paid or provider model calls, sending data or private content off the machine, destructive edits, decisions the Traigent service is meant to return, and any missing fact the step truly requires.
• Always end by recommending the next Traigent skill or action to take.
• Never weaken Traigent safety: dry-run before any paid run; get explicit approval before real cost or before any data leaves the machine; treat service-returned plans and next steps as authoritative. Never put the persona profile or any private content into telemetry, run metadata, experiment names, logs, or provenance files.

<!-- /INTERACTION_POLICY v1 -->