Model Tier Guide

The Model column in the wave document is the authoritative capability assignment. Tier names are provider-agnostic — wave-executor resolves them to concrete model IDs via references/model-map.yaml at execution time.

This guide explains the reasoning so you can judge edge cases and unlisted tasks.

---

Tiers

fast — mechanical operations

Use when the task is fully determined by explicit instructions with no judgment calls.

Characteristics:

• Follows a checklist or script step-by-step
• Output is verifiable against a known structure (rows added, files exist, exit codes)
• No synthesis across multiple documents required
• Failure mode: "step missed", not "wrong interpretation"

Examples:

• Pre-populating index files with a known list of items
• Running sync/build scripts and verifying exit codes
• Flipping status markers in a tracking file
• Lint and consistency audit passes

---

standard — structured output with bounded judgment

Use when the task drives a skill or fills a template and the output format is pre-defined. Judgment is required but constrained to a small, known decision space.

Characteristics:

• Uses a tool or skill with a defined output schema
• Must choose between a small set of options (e.g. promote / skip / flag)
• Output quality matters but format is fixed and verifiable
• Failure mode: "wrong choice in a known decision tree"

Examples:

• Triage tasks driven by triage-tool or triage-paper skills
• Writing documents that follow a fixed template (rubrics, configuration files)
• Binary decisions with a defined rationale format
• Consolidation passes that add rows to existing tables from already-written sources

---

smart — open-ended synthesis or deep evidence evaluation

Use when the task requires reading multiple documents, reconciling conflicting evidence, or producing analysis whose quality cannot be checked against a schema.

Characteristics:

• Must draw cross-cutting conclusions from N source files
• Evidence may be absent, weak, or contradictory — requires explicit acknowledgment
• Output quality is judged on depth and accuracy, not format compliance
• Failure mode: "shallow or confident-sounding but unsupported analysis"

Examples:

• Extracting cross-cutting themes from a set of analysis documents
• Per-item deep analysis with evidence evaluation and rubric scoring
• Writing synthesis sections that require comparing multiple sources

---

Default

Omitting the Model column or leaving a cell blank defaults to standard.

Decision rule for unlisted tasks

Ask: "Could a fast agent complete this correctly by following explicit steps?"

• Yes → fast
• No, but the output format is defined and the decision space is small → standard
• No, and the output quality depends on synthesis across documents → smart

---

Provider mapping

Tier names are resolved at execution time using references/model-map.yaml. Current defaults (Anthropic):

To use a different provider, update model-map.yaml — wave documents do not need to change.