Free-by-Default Audit (2026-05-24)

Free-by-Default Audit — 2026-05-24

Status: F-F track complete. Free-by-default is now live.

---

1. What Changed

#	Change	File
F-F-1	Added Free and Fast variants to ModelTier enum (YAML → generated)	contracts/orchestration/model-routing.v1.yaml
F-F-2	CostPreference derives Default; Economy is the #[default] variant	vox-orchestrator/src/config/enums.rs
F-F-3	default_cost_preference() returns Economy (was Performance)	vox-orchestrator/src/config/defaults.rs
F-F-4	Added supports_file_input: bool and supports_jsonl: bool to ModelCapabilities	vox-orchestrator/src/models/spec.rs
F-F-5	Added supports_web_search(), supports_file_input(), supports_jsonl() methods on ModelSpec	vox-orchestrator/src/models/spec.rs
F-F-6	New RoutingProfile enum (7 variants) in vox-orchestrator::types	vox-orchestrator/src/types/routing.rs
F-G	Re-activated selection/ in vox-dei-shim (7 files, 10 tests passing)	vox-dei-shim/src/selection/
Bonus	Reclassified 9 bootstrap catalog models to Free/Fast tier	contracts/orchestration/model-catalog.bootstrap.v1.json

---

2. Call-Site Audit (89 is_free / free_only sites)

Category A — Core routing logic (✅ consistent)

Site	What it does	Assessment
registry.rs:680,778	Skip free models when CostPreference::Performance	Correct — Performance mode intentionally excludes free
registry.rs:841,858,868,877,894	list_free_* registry query methods	First-class free API — correct
scoring.rs:80	Reduces quality_score for is_free models (QUALITY_FREE_PAID_COMPONENT=0.35 vs 0.95)	Intentional — free models score lower in quality; economy preference compensates
registry_model_resolve.rs:71,91,128	free_only filter in inference predicate	Correct — honours explicit free-only constraint
selection/scorer.rs	Economy: economy_free_bonus +3.0; Performance: performance_free_penalty -1.0	Correct
selection/free_tier.rs	FreeTierRouter filters to is_free=true only	Correct

Category B — MCP bridge (✅ consistent)

All 18 free_only sites in vox-orchestrator-mcp are propagating a free-tier decision from resolve_chat_llm_model / resolve_mcp_chat_model through to the infer layer. The consistency check at infer.rs:265 warns if routing.free_only != model.is_free. No changes needed.

Category C — Potential false positives (🔍 watchlist)

Site	Risk	Mitigation
registry.rs:680 — Performance && is_free skip	Could starve high-quality free models in Performance sessions	Acceptable: Performance is an explicit user opt-in; free models still win under Economy (default)
QualityLevel::Balanced → Performance in mode.rs:27	Default InferenceConfig still resolves to Performance cost preference	⚠️ Tracked below — see §3
scoring.rs:80 — quality score 0.35 for free models	Free models score ~60% less than paid on quality axis	Intentional; economy weight compensates; net result is free models win under Economy

Category D — Bootstrap catalog (✅ fixed)

9 models reclassified from Light/Local to Free or Fast:

• Free (5): qwen/qwen3-coder-next-32b, qwen/qwen3-coder:free, meta-llama/llama-4-scout:free, google/gemini-2.0-flash-lite, (1 existing)
• Fast (4): anthropic/claude-haiku-4.5, openai/gpt-5-mini, google/gemini-3-flash, google/gemini-3.1-flash-lite, deepseek/deepseek-v4-flash

---

3. Remaining Gap: QualityLevel::Balanced → CostPreference::Performance

In vox-orchestrator/src/mode.rs:

impl QualityLevel {
    pub fn to_cost_preference(self) -> CostPreference {
        match self {
            Self::Flash => CostPreference::Economy,
            Self::Balanced | Self::Premium => CostPreference::Performance,  // ← gap
        }
    }
}

InferenceConfig defaults to QualityLevel::Balanced, which maps to Performance. This means when the registry resolver uses cfg.quality.to_cost_preference() (not the default_cost_preference() path), the result is still Performance.

Recommendation: Change Balanced → Economy to complete free-by-default. This is a one-line change but has broader user-visible impact (existing sessions using “balanced” quality may see different model selection). Defer to a separate tracked item tagged free-by-default-follow-up.

Workaround in place: The default_cost_preference() path (OrchestratorConfig) now returns Economy, which covers the common session-start path. The InferenceConfig path is only reached when an explicit quality level is requested.

---

4. Exploration Bonus Gap

In vox-orchestrator/src/routing/engine.rs:102-106:

if self.policy.routing_objective.kind == "quality_first"
    && s + f == 0
    && matches!(m.capabilities.tier, ModelTier::Pro)
{
    base += 0.06;
}

New Free and Fast models also benefit from exploration (Thompson exploration covers them), but the hardcoded Pro-tier exploration bonus does not apply. Under quality_first routing, Free/Fast newcomers compete fairly on Thompson draws but don’t get the Pro novelty boost.

Recommendation (not urgent): Extend to ModelTier::Pro | ModelTier::Fast | ModelTier::Free for equal exploration across all tiers that might be optimal. Tag exploration-parity-follow-up.

---

5. True Positives (healthy patterns)

• FreeTierRouter in vox-dei-shim::selection provides intelligent multi-candidate routing with FIM, latency, and vision constraint handling.
• infer.rs:250 enforces free-only at runtime — no leakage of paid models when free_only=true.
• list_free_models_for_strength() and list_free_models_for_strength_with_pred() in registry.rs give callers clean first-class free-model APIs.
• economy_free_bonus: 3.0 in ScoringWeights means free models out-compete paid by 3 points in Economy mode, making them the default winner when available.

---

6. Next Steps

Item	Priority	Effort	Status
QualityLevel::Balanced → Economy	Medium	1 line	✅ Done — mode.rs Flash+Balanced→Economy, Premium→Performance
Exploration bonus parity (Pro/Fast/Free)	Low	3 lines	✅ Done — engine.rs novelty bonus extended to Fast+Free
Update where-things-live.md to document RoutingProfile	Low	1 row	✅ Already present (vox-dei-shim + vox-orchestrator rows)