Verbosity modes

sift’s full VectorizeOutput for a 10-result SERP is ~2.4K tokens. About 52% of that is augmentation — the per-result quality_vector and signals[] are the bulk. For most downstream agents the secondary fields aren’t behaviorally load-bearing, so sift exposes three verbosity modes.

The three modes

Mode	Per-result shape	Aggregate	Typical use	Token delta
concise (default)	url, title, description, quality.{tier, reason, confidence}, recommended_action, safety_flag	tier_distribution, vendor_dominance_ratio, diversity_entropy	Agent consumption — the ablation-validated sweet spot	~33% smaller
full	Every quality_vector field + signals[]	Complete	Audit, debugging, learning-loop extracts	Baseline
summary	url, title, description, safety_flag only	Complete	SERP landscape dashboards, diversity monitoring	~45% smaller

How the default was chosen

A Phase 5 ablation (see the sift-token-economy design note) ran 3 query types × 3 payload shapes × 3 agent models (Haiku 4.5 / Llama 3.3 70B / Sonnet 4.6) = 27 runs. On each run the agent was asked to summarize the SERP in under 300 characters.

Finding: concise and full produced indistinguishable downstream behavior. summary lost the agent’s ability to weight individual sources but preserved SERP-level awareness (dashboards still work). So the default is concise: enough to weight source quality correctly, at a third less input cost.

Verbosity is input-only

Observations always record the full payload regardless of what the caller requested. The learning-loop substrate needs every dimension for post-hoc analysis — it would be a mistake to lose signals[] and authoritative_weight just because the agent at call time didn’t need them.

Picking a mode

• Start with concise. It’s the default and covers ~95% of agent use cases.
• Switch to full when auditing a specific classification, running a validation suite, or needing signals[] to trace reasoning.
• Use summary for dashboards or when you only need the SERP-level picture (e.g., a side-panel showing how biased the current query’s results are).

Example — same query, three modes

concise response (trimmed):

{
  "verbosity": "concise",
  "results": [
    {
      "url": "https://tandfonline.com/doi/full/...",
      "title": "...",
      "description": "...",
      "quality": { "tier": "peer_reviewed", "reason": "academic journal", "confidence": 0.95 },
      "recommended_action": "keep",
      "safety_flag": null
    }
  ],
  "aggregate_vector": {
    "tier_distribution": { "peer_reviewed": 10, ... },
    "vendor_dominance_ratio": 0,
    "diversity_entropy": 0
  },
  "summary_hints": ["Primary / peer-reviewed sources are present (10 of 10)..."]
}

full adds per-result editorial_standards, commercial_intent, self_promoting, third_party, domain_content_mismatch, authoritative_weight, signals[], plus mean_editorial_standards and mean_authoritative_weight on the aggregate.

summary strips all per-result quality fields, keeping only url/title/description/safety_flag plus the full aggregate and hints.