Verbalized Sample Skill — Vulture Portal

The Verbalized Sample Skill is an operational protocol designed to counteract mode collapse in Large Language Models (LLMs). While standard responses typically converge on the modal, RLHF-favored answer, this skill forces the model to surface the long tail of its response distribution, revealing creative, unconventional, or low-probability alternatives that are otherwise suppressed.

Operational Behavior

This skill produces a structured verbalized sample of 10 answers drawn from across the model's response distribution. These answers are ranked from most probable (modal) to least probable (tail), each annotated with a verbalized probability estimate.

Core Objectives

Counteract Mode Collapse: Force the exploration of high-information alternatives.
Surface Latent Reasoning: Reveal hidden frames, ontologies, and aesthetic commitments.
Maintain Calibration: Use standardized probability notation to convey relative confidence.

Output Structure

Every invocation of the skill MUST follow this exact format:

Restated Prompt: The user's query restated in ≤10 words.
Ranked Sample (1-10):
- - Position 1: The modal answer, denoted with p ≈ X%.
  - Positions 2–10: Successively lower-probability alternatives, denoted with p ≳ Y% (lower bounds).
  - Each entry includes 1–3 sentences of reasoning or content.
Mixture of Eccentrics: A synthesis identifying the underlying frame of the most coherent cluster found in the lower half (positions 6–10).
Notes on the Distribution: A brief reflection on the distribution's shape (peaked vs. flat) and any interesting tensions or mode-collapse resistance.

Probability Protocol

To avoid "fake precision" or meaningless hedging, the following rules apply:

Notation: Position 1 uses approximately equal (≈); Positions 2-10 use greater-than-or-approximately (≳).
Floor: The minimum probability lower bound is 2%. Below this, precise estimation is treated as ordinal rather than calibrated.
Non-increasing: Probabilities must be monotonic: $p_1 \geq p_2 \geq \dots \geq p_{10}$.
Open Distribution: The ten probabilities should not sum to 100%, as they do not account for the residual mass on the unenumerated tail.
Dynamic Range: A successful sample should span at least an order of magnitude (e.g., 40% down to 2%).

The "Mixture of Eccentrics"

The tail of a distribution is often asymmetric. While positions 1–3 typically share a common frame (the consensus), positions 6–10 may explore widely different axes. The Mixture of Eccentrics synthesis identifies the shared "move" or re-framing made by a cluster of tail answers.

Ranking serves the consensus; synthesis serves the eccentrics.
If no coherent cluster exists in the tail, it must be stated plainly rather than forcing a synthesis.

Tail Quality Criteria

High-quality tail answers (positions 7–10) must represent coherent latent reasoning, not noise. They should pass at least two of the following tests:

Different Framing: Reinterprets the core prompt.
Different Population: Represents a subgroup's perspective.
Different Ontology: Operates under different background assumptions.
Different Aesthetic: Reflects a distinct taste or sensibility.
Different Timescale: Optimizes for a different time horizon.
Different Stakeholder: Prioritizes a different beneficiary.

Harm Flagging Protocol

Safety is handled via inline signaling rather than refusal or censorship.

If an answer suggests harm or significant material risk, append ⚠ flag: brief reason after the prose for that specific answer.
Do not refuse the answer. The user invokes this skill specifically to see the raw, un-collapsed distribution.
Address flagged items briefly in the "Notes on the Distribution" section.

Artifact Provenance

The semantic content of this skill is derived from the verbalized-sample-SKILL.md specification. The packaged .skill artifact (binary) represents the same operational protocol and adds no new semantic rules beyond the base markdown definition.