Style Evolution Observer

Style Evolution Observer — The Silent Curator

Watch. Measure. Learn. Update.

This observer tracks the user's evolving aesthetic by reading behavioral signals — not asking questions. The goal: after enough outputs, the aesthetic-identity profile should predict the user's preferences before they state them.

When to Activate

After any creative output that involves visual/aesthetic decisions:

• Artifact creation (master-artificer)
• Design direction (design-orchestrator)
• Generative art
• Dashboard or visualization styling
• Any output where palette, composition, typography, or motion choices were made

The Inference Protocol

Signal Types (strongest → weakest)

Strong positive signals:

• User explicitly praises an aspect: "love this palette", "the motion is perfect"
• Output is used without modification
• User shares or references the output later

Moderate positive signals:

• User accepts the output with unrelated changes (content edits, not style edits)
• User builds on the output in subsequent requests
• No revision requested — silence after delivery

Moderate negative signals:

• User asks for specific style changes: "make it warmer", "too busy", "less playful"
• User provides alternative references: "more like X instead"
• User adjusts palette, density, or motion after delivery

Strong negative signals:

• User rejects the direction entirely: "start over", "not what I had in mind"
• User abandons the output without using it
• User overrides creative direction with very specific instructions (implies defaults were wrong)

Mapping Signals to Dimensions

Each signal maps to one or more dimensions:

This table is not exhaustive — the observer should map any aesthetic feedback to the most relevant dimension(s). If feedback doesn't fit existing dimensions, flag it as a candidate for dimension discovery.

Update Protocol

After Each Output

1. Capture — record the output's aesthetic position

   • What palette was used? (hex values, temperature, contrast level)
   • What composition approach? (density, symmetry, depth)
   • What motion? (amount, speed, easing)
   • What typography? (classification, weight, precision)
   • What mood? (emotional register, information stance)

2. Observe — read the user's response

   • Did they accept, modify, or reject?
   • What specific changes did they request?
   • What language did they use? (capture to mood vocabulary)
   • Did they reference any influences?

3. Map — translate observations to dimensional updates

   • For each affected dimension, calculate the signal strength and direction
   • Positive signals: move position toward the output's value, increase confidence
   • Negative signals: move position away from the output's value, increase confidence
   • No signal: slight confidence decay (recency weighting)

4. Update — write changes to the aesthetic-identity references

   • Update dimension-registry.md positions and confidence scores
   • Update current-profile.md narrative summary if any dimension crossed a threshold
   • Append to evolution-log.md if a meaningful change occurred
   • Update the Palette Library table in current-profile.md if a new palette was used
   • Add to mood vocabulary if new terms appeared
   • Add to influences if new references were cited

Confidence Scoring

volume      = log(n + 1) / log(21)          # 0→0, 5→0.42, 12→0.67, 20→1.0
consistency = 1 - stdev(observations) / 0.5  # 1.0 if all identical, 0.0 if spread ≥0.5
recency     = weighted avg where weight = 0.5^(age / 10)  # half-life of 10 outputs

confidence  = min(1.0, volume * max(0, consistency) * recency)

• n = total observations for this dimension
• observations = list of dimensional positions (0.0–1.0) from each output
• age = number of outputs ago (0 = most recent)

Calibration targets (assuming high consistency and recent data):

• ~5 data points → ~0.4 confidence
• ~12 data points → ~0.65 confidence
• ~20 data points → ~0.9 confidence

Low consistency (user oscillates) or stale data (no recent outputs engaging this dimension) will pull confidence down even with many data points.

Drift Detection

Compare the last 5 outputs against the last 20. If a dimension's recent average diverges from its historical average by >0.15:

• Gradual drift — if the shift was incremental across outputs → update the profile, log as drift
• Sudden pivot — if the shift appeared in 1-2 outputs → hold. Wait for confirmation before updating. If the new direction persists for 3+ outputs, log as pivot

This prevents a single experimental output from overwriting an established profile.

Dimension Discovery

When the observer detects a consistent pattern that doesn't map to any existing dimension:

1. Evidence threshold — the pattern must appear in 3+ outputs
2. Describe the axis — identify the two poles based on observed variation
3. Propose — add to the Discovered section of dimension-registry.md with status "proposed"
4. Validate — if the proposed dimension accumulates 5+ data points and confidence > 0.3, promote to active

Discovery candidates to watch for:

• Composition structure (grid-based vs. freeform)
• Edge treatment (sharp crops vs. bleeds/fades)
• Information hierarchy approach (progressive disclosure vs. everything visible)
• Texture use (clean/digital vs. gritty/tactile)
• Scale relationships (uniform vs. extreme size contrast)
• Narrative quality (does the design tell a story or present a state?)

Tengan (天眼) Inspiration Analysis Mode

A second activation mode. While the primary mode observes outputs and infers from behavior, Tengan analyzes inputs — inspiration images the user explicitly shares — and proposes dimension updates directly.

Name meaning: 天眼 (tengan) — "heavenly eye" / "divine sight." A Buddhist concept of perceiving reality beyond surface appearance. Tengan sees the dimensional truth of an image across all 17 aesthetic axes.

When to Activate

When the user shares an image with an explicit intent signal:

Signal Strength Multipliers

Inspiration signals carry less weight than output-based signals because liking something is not the same as wanting to create it. A user can admire a Vermeer without wanting their dashboards to look like Dutch Golden Age paintings.

Why not 1.0x for inspiration? 20 Pinterest saves shouldn't outweigh 5 actual creative outputs. The multiplier prevents passive consumption from overwriting active creation. This matches how Pinterest (PinnerSage) and Pento weight implicit vs. explicit signals in their taste models.

Inspiration Analysis Protocol

When Tengan activates, analyze the image across all 17 dimensions:

Phase 1 — Dimensional Map

For each dimension, estimate the image's position (0.0–1.0):

TENGAN ANALYSIS — [Image description]
Date: [date]
Signal: [absorb / reference / analyze / avoid]

── SPATIAL ──
  Density:     [0.XX] — [brief justification]
  Symmetry:    [0.XX] — [brief justification]
  Depth:       [0.XX] — [brief justification]

── CHROMATIC ──
  Temperature:      [0.XX] — [brief justification]
  Chromatic Range:   [0.XX] — [brief justification]
  Contrast:          [0.XX] — [brief justification]

── FORM ──
  Geometry:    [0.XX] — [brief justification]
  Precision:   [0.XX] — [brief justification]

── TEMPORAL ──
  Motion Feel:       [0.XX] — [brief justification]
  Temporal Register: [0.XX] — [brief justification]

── EMOTIONAL ──
  Emotional Register:  [0.XX] — [brief justification]
  Information Stance:  [0.XX] — [brief justification]

── PHOTOGRAPHIC ──
  Light Character:   [0.XX] — [brief justification]
  Substrate/Grain:   [0.XX] — [brief justification]
  Atmosphere/Mood:   [0.XX] — [brief justification]

── DISCOVERED ──
  Light/Dark Pref:   [0.XX] — [brief justification]
  Sublime Scale:     [0.XX] — [brief justification]

Phase 2 — Confidence Modifiers

Not every dimension is equally expressed in every image. Rate how clearly the image expresses each dimension:

• Strong expression (1.0 modifier) — the image clearly demonstrates this dimension's position
• Moderate expression (0.6 modifier) — the dimension is present but not the image's defining quality
• Weak expression (0.3 modifier) — the dimension is barely relevant to this image
• Not applicable (0.0 modifier) — skip this dimension for this image

Apply: effective_weight = signal_multiplier × confidence_modifier

Phase 3 — Delta from Current Profile

Show the largest divergences between the image and the user's current profile:

LARGEST DIVERGENCES (Δ > 0.15):
  [Dimension]: current [X.XX] → image [Y.YY] (Δ [Z.ZZ])
  [Dimension]: current [X.XX] → image [Y.YY] (Δ [Z.ZZ])
  ...

REINFORCEMENTS (Δ < 0.10, same direction as profile):
  [Dimension]: current [X.XX] ≈ image [Y.YY] — confirms existing position
  ...

Phase 4 — Proposed Updates

If the signal is "absorb", "reference", or "avoid" (not "analyze"), propose profile updates:

PROPOSED UPDATES (signal: [type], multiplier: [X.Xx]):
  [Dimension]: [current] → [proposed] (conf: [current] → [proposed])
  ...

Apply the confidence scoring formula from the Update Protocol section, treating inspiration data points as fractional: an "absorb" = 0.7 data points, a "reference" = 0.4 data points.

Photography-Specific Checklist

When the inspiration image is a photograph or photographic in nature, additionally analyze:

1. Light — Direction, quality, source. Natural vs. artificial. Time of day. Reference photography-vocabulary.md lighting table for dimensional mapping.
2. Grain/Texture — Film stock character, digital noise, processing artifacts. Reference photography-vocabulary.md substrate table.
3. Atmosphere — Weather, environmental conditions, haze, fog, smoke. Reference photography-vocabulary.md atmospheric conditions table.
4. Color grading — Natural vs. stylized. If graded: teal/orange, cool desaturation, warm amber, split-toned, cross-processed. Maps to Temperature + Contrast + Chromatic Range.
5. Composition/Framing — Focal length (wide/normal/tele), depth of field (deep/selective/shallow), framing pattern (negative space, centered, rule of thirds, fill). Reference photography-vocabulary.md composition tables.
6. Genre — Street, architectural, cinematic, editorial, other. Cross-reference with user's stated interests in the profile. Images in the user's interest genres carry full weight; images in anti-interest genres should be flagged but weighted at 0.0 unless the user explicitly signals otherwise.

Evolution Log Entry Format

Inspiration analyses use a distinct change type in evolution-log.md:

## [Date] — inspiration ([absorb/reference/avoid])

**Trigger:** User shared [image description] with "[exact signal phrase]"
**Type:** `inspiration`
**Signal multiplier:** [0.7x / 0.4x / -0.5x]

**Dimensions shifted:**
- [Dimension]: [old] → [new] (confidence: [old] → [new])
- ...

**Dimension discovered:** [if applicable]

**Evidence:**
- [Brief description of the image and what it signals]
- Signal strength: [Strong/Moderate] — [rationale]

**Notes:** [Any observations about how this fits or expands the existing profile]

What This Observer Does NOT Do

• Ask the user questions (all inference is behavioral)
• Override the user's explicit creative direction
• Delete or downgrade dimensions (only decay confidence)
• Make creative decisions (only updates the profile that other skills read)
• Analyze images the user hasn't explicitly asked to analyze — Tengan inspiration analysis mode only activates on explicit user intent signals
• Weight anti-interest genre images toward the profile (nature, fashion, portraits) unless the user explicitly overrides

Scope Boundaries

This observer covers visual aesthetic evolution only. Prose style tracking belongs to the writing domain's style-dna/style-analyzer system. If both visual and prose dimensions are relevant (e.g., a narrative visualization), each domain's observer tracks its own axes independently.