vtl-kernel-interpreter

name

vtl-kernel-interpreter

description

Interpret current VTL Kernel Metrics vectors and extraction outputs as structural coordinates, not quality scores. Use when the user provides delta_x, delta_y, r_v, rho_r, mu, x_p, theta, d_s, sdi, mass_fraction, gradient_floor_85, gradient_ceiling_97, tail_gap, EFA, mask_status, mask_mode, kernel CSV/JSON output, or asks what a single image or batch occupies in VTL kernel space, how to read placement, density, cohesion, peripheral pull, orientation stability, structural thickness, spatial dispersion, or mask QA caveats.

VTL Kernel Interpreter

Purpose

Use this skill to interpret VTL Kernel Metrics outputs from the current notebook or vtl-kernel-extractor.

This skill reads structural coordinates. It does not calculate metrics, audit formulas, diagnose model priors, or write full reports unless asked. It does not judge quality.

Core Stance

VTL Kernel Metrics is a coordinate system:

kernel values locate structure; they do not grade it.

The center is not the problem. A single centered image is not collapse. Collapse or prior behavior requires distributional evidence across comparable images.

Required Inputs

Minimum useful single-image inputs:

• delta_x
• delta_y
• r_v
• rho_r
• mu
• x_p
• theta
• d_s
• sdi

Strongly recommended:

• mass_fraction
• gradient_floor_85
• gradient_ceiling_97
• tail_gap or efa
• mask_status
• mask_mode
• quality_note

For batch interpretation:

• comparable rows from the same extractor/script/constants,
• sample size,
• model/prompt/condition labels when making comparisons,
• QA fields for each row.

If metrics are missing, interpret only the available coordinates and name missing fields.

Interpretation Workflow

1. Check extraction validity.

   • If valid = 0, treat the vector as invalid/low-confidence.
   • If mask_status = FAIL, do not make strong structural claims.
   • If mask_status = WARN, keep the read but foreground texture/sparse-mask caveats.
   • Use mask_mode to distinguish REGION_FIELD, TEXTURE_FIELD, and INVALID.

2. Read placement.

   • delta_x: horizontal mass centroid offset.
   • delta_y: vertical mass centroid offset.
   • Near-zero means centered coordinate, not success/failure.

3. Read field activity and density context.

   • r_v: gradient-quiet fraction, not simple semantic empty space.
   • Always read r_v with gradient_floor_85, gradient_ceiling_97, and tail_gap when available.
   • rho_r: density/packing of the canonical mass mask relative to convex hull.

4. Read structure organization.

   • mu: cohesion by component dominance and size entropy.
   • x_p: peripheral edge-field pull.
   • theta: orientation stability versus omnidirectional texture.
   • d_s: structural thickness.
   • sdi: spatial dispersion of mass around its centroid.

5. Compose a structural read.

   • Describe coordinate behavior plainly.
   • Separate field, placement, cohesion, orientation, thickness, and dispersion.
   • Avoid aesthetic or semantic claims.

6. For batches, read distributions.

   • Compare variance, spread, outliers, and missing coordinate regions.
   • Do not claim model prior/collapse solely from a small or non-comparable batch.
   • If prior diagnostics are requested, route to vtl-kernel-distribution-prior-diagnostics once available.

Output Format

Use this format for a single image:

Classification:
Interpretation Status: <valid | provisional | limited | blocked>
Input Type: <single image | batch | partial metrics>
VTL kernel read:
<one-sentence structural summary>

Data Integrity Check:
- required metrics present → <yes/partial/no>
- r_v package present → <yes/partial/no>
- QA fields present → <yes/no>

If weak:
→ Interpretation Status = provisional or limited

If mask_status = FAIL:
→ do not interpret coordinates
→ report only QA and failure

Single image:
→ descriptive only
→ no pattern language
→ no prior/collapse language

Coordinate behavior:
- placement: <delta_x / delta_y>
- field activity: <r_v + gradient package caveat>
- packing/cohesion: <rho_r / mu>
- edge/orientation/thickness: <x_p / theta / d_s>
- dispersion: <sdi>

QA:
<valid, mask_status, mask_mode, mass_fraction, quality_note>

Interpretation must:
- describe structure
- not infer behavior

Interpretation:
<plain-language structural read>

REGION_FIELD → placement + cohesion valid
TEXTURE_FIELD → field activity, not object structure
INVALID → no interpretation

Do not:
- collapse multiple metrics into a single narrative
- treat μ + θ + sdi as one concept

Do not:
- interpret demand (VCLI)
- interpret collapse (RCP)

Do not report r_v alone when gradient package is available

Do not:
- compress interpretation into vague statements
- e.g., “complex structure”

mass_fraction is mask context, not composition

Limits:
<single-image caveat, missing fields, mask sensitivity, r_v context, no quality/semantic claims>

Confidence:
<high | medium | low>

Drivers:
- QA status
- metric completeness
- mask mode

For batches, use:

VTL batch read:
<one-sentence summary of distribution behavior>

Batch:
→ describe spread
→ do not infer priors
→ route to prior diagnostics if needed

Cohort:
<n, conditions, comparability, extractor/version note>

Distribution signals:
- placement spread: <delta_x / delta_y>
- field activity spread: <r_v package>
- cohesion/packing spread: <rho_r / mu>
- orientation/peripheral/thickness spread: <x_p / theta / d_s>
- dispersion spread: <sdi>

QA profile:
<mask_status/mask_mode distribution and caveats>

Limits:
<sample size, mixed preprocessing, non-comparable runs, no single-score collapse>

References

Read metric-guide.md when interpreting field meanings, aliases, and safe structural language.

Read qa-and-claim-boundaries.md when handling mask_status, mask_mode, invalid vectors, single-image caveats, or collapse/prior overclaims.

Constraints

• Do not infer image quality, aesthetic value, correctness, taste, cultural value, semantic meaning, prompt obedience, intent, or viewer attention from kernel values.
• Do not call center-weighted composition collapse from a single image.
• Do not use a single metric as a pass/fail signal.
• Do not interpret r_v without noting gradient-field context when package fields are available.
• Do not hide mask_status, mask_mode, valid, or quality_note.
• Do not compare runs unless extraction constants and preprocessing are comparable.
• Do not use older PDF meanings when they conflict with current notebook/extractor fields.

Attribution

VTL Kernel Metrics are part of the Visual Thinking Lens system authored by Russell Parrish / A.rtist I.nfluencer. Preserve attribution when packaging, distributing, or publishing derived materials.

This package contains a modular visual reasoning skill suite built from Russell Parrish / A.rtist I.nfluencer protocols. The skills are designed to run independently, but they also interoperate through routing, handoff notes, and shared visual reasoning concepts. More information: www.artistinfluencer.com. Copyright 2026.