Skip to content

Latest commit

 

History

History
378 lines (290 loc) · 14.3 KB

File metadata and controls

378 lines (290 loc) · 14.3 KB

Scoring & Visualization Guide

Reference for UI implementors: how to score dialectical entities, what "good" and "bad" look like, and how to lay out the molecular visualization.


Metrics by Entity

Thesis (T)

The thesis defines the polarity axis. It IS the reference point.

Metric Value Notes
HS Always 1.0 Tautological — T defines its own apex

No quality metrics — a thesis is the starting axiom.


Antithesis (A)

The antithesis opposes the thesis. Its quality determines polarity strength.

Metric Range What it measures
HS 0.0–1.0 How well A captures the antithesis apex concept
Mode 0.0–1.0 Type of opposition (privation → negation)
Arousal 0.1–0.9 Tension visibility/intensity

HS Scale:

Range Quality
0.9–1.0 Perfect antithesis — exemplary
0.7–0.9 Strong — captures most of the apex
0.5–0.7 Moderate — some aspects, acceptable
0.3–0.5 Weak — still valid, poor quality
0.1–0.3 Very weak — barely an antithesis
0.0–0.1 Invalid — wrong category entirely

Mode Scale (opposition mechanism):

Value Type Description
1.0 Negation Direct, active opposition
0.9 Inversion Reversal of T's meaning
0.8 Devaluation Diminishing T's worth
0.7 Hollowing Emptying T of substance
0.6 Corruption Degrading/perverting T
0.5 Distortion Twisting T's form
0.4 Skew Imbalancing T
0.3 Blocking Obstructing T
0.2 Suppression Holding T down
0.1 Distancing Drifting from T
0.0 Privation Complete absence of T

Arousal Scale (tension activation):

Value Label Description
0.9 Active Fully manifest, immediate
0.8 Intense Very active, urgent
0.7 High Strong, clearly visible
0.6 Elevated Becoming prominent
0.5 Moderate Balanced, present tension
0.4 Mild Noticeable but subdued
0.3 Low Background tension
0.2 Latent Barely perceptible
0.1 Dormant Completely invisible

Polarity (T + A container)

The polarity is a container. Its quality is determined by the antithesis:

Indicator Source Good Bad
Tightness of opposition HS_A High (0.7+) Low (<0.5)
Directness Mode Context-dependent Context-dependent
Aliveness Arousal Context-dependent Context-dependent

Mode and Arousal are characterization, not quality — a "dormant" polarity isn't necessarily bad, it's just latent.


Aspects (T+, T-, A+, A-)

Each aspect has three complementarity scores plus HS:

Metric Range What it measures
HS 0.0–1.0 Similarity to taxonomy apex for that position
K_T 0.0–1.0 How well the aspect complements the thesis
K_A 0.0–1.0 How well the aspect complements the antithesis
Ks 0.0–1.0 Combined: (K_T + K_A) / 2 — complementarity toward synthesis

Ks interpretation:

  • 0.0 = Actively undermines the system
  • 0.5 = Neutral
  • 1.0 = Strongly enhances the whole

Expected pattern for a balanced tetrad:

Position K_T K_A Ks Pattern
T+ High Low-ish Mid-high Favors T but complements A somewhat
A+ Low-ish High Mid-high Favors A but complements T somewhat
T- Low Mid Low Undermines T, doesn't help A
A- Mid Low Low Undermines A, doesn't help T

(A theory heuristic sometimes cited is K_T + K_A ≈ 1.0 per aspect in ideal systems, but nothing in code computes or enforces this — K_T and K_A are scored independently.)


Perspective (full tetrad)

Computed from the four aspects' Ks values:

Metric Formula Range Good Bad
diff_t Ks(T+) − Ks(T−) −1 to 1 ≥ 0.1 < 0.1
diff_a Ks(A+) − Ks(A−) −1 to 1 ≥ 0.1 < 0.1
area (= theory's SP, Synthesis Potential) diff_t + diff_a −2 to 2 (well-formed tetrads ~0 to 2) ≥ 0.7 < 0.3
area_normalized area / 2 −1 to 1 (well-formed tetrads ~0 to 1) ~0.5 ~0.15
rectangularity [Ks(T+)−Ks(A+)]² + [Ks(T−)−Ks(A−)]² 0+ < 0.01 > 0.09

Empirical inequalities (pass/fail):

  1. diff_t ≥ 0.1 AND diff_a ≥ 0.1
  2. |diff_t − diff_a| ≤ 0.15
  3. Ks(T+) > 0.4 AND Ks(A+) > 0.4
  4. Ks(T−) < 0.6 AND Ks(A−) < 0.6

Structural validity checks (pass/fail):

Check Threshold What it tests
Conceptual Coherence (CC) both control scores ≥ 0.7 "T+ without A+ yields T−" and "A+ without T+ yields A−"
Diagonal Contradiction both ≥ 0.7 T+ vs A− and A+ vs T− are genuine contradictions

CC stores the average of the two control scores as its value, but the pass/fail criterion is that each score clears 0.7 (ConceptualCoherenceEstimation.is_coherent) — 0.5 + 0.9 does not pass. Diagonal Contradiction is not part of the standard PerspectiveValidation run; it is an extra LLM call that only fires on user-edited tetrads (edit_perspective). Generated tetrads are never gated on it.

Quality tiers (suggested UI grouping — not a built-in framework ranking; see note below):

Tier Criteria
Invalid Fails CC (Diagonal Contradiction, too, but only on user-edited tetrads)
Bad Fails any empirical inequality
Good Passes all checks
Best Passes all + highest area_normalized + lowest rectangularity

Ranking, in practice: these tiers and an area_normalized ordering are guidance for a UI — the framework does not implement them. The only ranking in code is AnalysisPipeline._rank_polarities, which orders polarities by their antithesis heuristic_similarity against a soft HS_THRESHOLD = 0.7 (if nothing clears it, the top few are expanded anyway). If a UI wants to order valid tetrads, area_normalized (0–1, higher = better) gated by the validity checks is a reasonable choice, with rectangularity as a tiebreaker.

area_normalized is a comparative hint, not an absolute grade. The ~0.5/~0.15 figures above are rough visual anchors for ordering tetrads against each other, not pass/fail cutoffs. The theory (SP / Synthesis Potential) defines no universal cutoff — see docs/theory/scoring.md. Do not treat a single area_normalized value as "good" or "bad" in isolation.


Transitions (Ac+, Ac−, Re+, Re−)

Metric Range What it measures
Insight 0.0–1.0 Depth of understanding in the transition
Proactiveness 0.0–1.0 How actionable/practical the transition is
Feasibility 0.0–1.0 Practical achievability
HS 0.0–1.0 Similarity of Ac+/Re+ to their derived apex

Insight scale (Y-axis — depth of transformation):

Value Level Character
1.0 Transcendence Paradigm shift
0.9 Redirection Fundamental change
0.8 Inversion Flipping perspective
0.7 Anticipation Acting ahead
0.6 Leverage Using leverage points
0.5 Composition Combining elements
0.4 Reformulation Restructuring approach
0.3 Variation Deliberate small changes
0.2 Tuning Fine-tuning
0.1 Procedure Following protocol
0.0 Reflex Automatic response

Proactiveness scale (X-axis — action vs reflection):

Value Level Zone
0.0 Observation Re (reflection)
0.1 Detection Re
0.2 Interpretation Re (apex zone)
0.3 Framing Re
0.4 Evaluation Midpoint
0.5 Coordination Ac (action)
0.6 Intervention Ac (apex zone)
0.7 Implementation Ac
0.8 Configuration Ac
0.9 Governance Ac
1.0 Stewardship Ac

Feasibility scale:

Range Meaning
0.9–1.0 Highly achievable
0.7–0.8 Moderately feasible
0.5–0.6 Challenging but achievable
0.3–0.4 Extremely difficult
0.0–0.2 Practically impossible

Cycles & Wheels

Metric Range What it measures
Causality Probability 0.0–1.0 Plausibility of this causal ordering vs alternatives

The value stored on a Cycle or Wheel is the raw LLM plausibility score (0.0–1.0), not a normalized one. Normalization to a layer-relative share (siblings sum to 1.0) is applied only to Wheel Transitions (nth-root decomposed) and is otherwise computed on the fly for display (raw P vs normalized %). A UI reading the estimation directly off a Cycle/Wheel gets the raw score.


Molecular Visualization

The perspective is displayed as a molecule with T and A as the nucleus and aspects as bonded satellites.

Coordinate System

No formal X-axis. This is a spatial/relational layout, not a chart. Only vertical position (Ks) is a true metric axis.

Layout Rules

Element Position/Distance Encoded metric
T ↔ A distance 1 − HS_A Polarity quality (tight = good)
Aspect vertical position Ks value Complementarity toward synthesis
Aspect bond length to parent 1 − K_parent Affinity to parent concept
Trapezoid shape Connect T+→A+→A−→T− Area + rectangularity visible

"K_parent" means:

  • For T+ and T−: bond length = 1 − K_T
  • For A+ and A−: bond length = 1 − K_A

What "Good" Looks Like

                Ks
                 ↑
                 │
         T+ ○━━━━━━━━○ A+        ← Both high Ks (~0.6–0.7)
            ╲       ╱              ← Short bonds (high K)
             T●━━━●A               ← Close together (HS_A = 0.85)
            ╱       ╲              ← Short bonds
         T- ○━━━━━━━━○ A-        ← Both low Ks (~0.2–0.3)
                 │
                 └──────

   ✓ Compact nucleus (high HS_A → strong opposition)
   ✓ Tall vertical gap (high area → clear differentiation)
   ✓ Flat top/bottom edges (low rectangularity → balanced)
   ✓ Short bonds (high K → tight complementarity)
   ✓ Symmetric left/right

What "Bad" Looks Like

                Ks
                 ↑
                 │
         T+ ○                       ← High Ks (0.7)
              ╲
               ╲ (long bond)
                T●                          ●A    ← Far apart (HS_A = 0.4)
                                           ╱ ╲
                                     ○ A+     ╲   ← Mid Ks (0.45)
                                               ╲
         T- ○                            ○ A-  ← A- very low (0.15)
                 │
                 └──────

   ✗ Sprawling nucleus (low HS_A → weak opposition)
   ✗ Tilted shape (high rectangularity → imbalanced sides)
   ✗ Long bonds (low K → aspects loosely related)
   ✗ Small vertical gap on one side (low area)
   ✗ Asymmetric — T-side taller than A-side

What "Mediocre" Looks Like

                Ks
                 ↑
                 │
         T+ ○━━━━━━○ A+          ← Mid Ks (~0.50, 0.55)
            ╲     ╱
             T●━●A                ← Moderate closeness (HS_A = 0.7)
            ╱     ╲
         T- ○━━━━━━○ A-          ← Mid Ks (~0.35, 0.40)
                 │
                 └──────

   ~ Decent nucleus (acceptable HS_A)
   ~ Small vertical gap (low area = 0.30 → weak differentiation)
   ~ Shape is rectangular but flat (aspects mushed together)
   ~ Bonds mid-length

Visual Cue Summary

Visual property Metric Good Bad
T-A closeness HS_A Close (< 0.3 gap) Far (> 0.5 gap)
Vertical gap (+ row vs − row) area Tall (> 0.7) Squished (< 0.3)
Top/bottom edge horizontality rectangularity Flat (< 0.01) Tilted (> 0.09)
Bond lengths to parent K_T / K_A Short (K > 0.6) Long (K < 0.3)
Overall compactness Combined Dense molecule Straggling/diffuse

Secondary Annotations (not spatial)

Property Encoding suggestion
Mode Color/icon on T-A bond (e.g., negation=red, privation=grey)
Arousal Bond thickness or pulse animation
HS of aspects Node size or opacity
CC validity Green/red border on the tetrad quadrilateral
Diagonal contradiction Diagonal dashed lines with check/cross

Geometric Interpretation (from theory)

The paper (Generative Rules for Dialectical Synthesis) establishes:

  1. The tetrad forms a trapezoid when plotted with Ks on the vertical axis and T-side / A-side on the horizontal.

  2. In practice, no tetrad is a perfect rectangle. The goal is a trapezoid approaching rectangular shape with maximum surface area.

  3. Balanced tetrads (A and D in Fig. 4) produce near-rectangular trapezoids where like-signed components occupy comparable Ks levels.

  4. Distorted tetrads produce skewed, twisted, or compressed trapezoids — the shape visually communicates what's wrong:

    • Compressed = low area, aspects not well-differentiated
    • Tilted = high rectangularity, one side overdeveloped
    • Narrow = low HS_A, weak polarity foundation
  5. Selection criterion: "the best one typically shows trapezoid that is most similar to rectangular with the largest surface area" (paper, p. 10).


Quality Summary Table

Entity Primary quality metric Threshold Secondary
Thesis Defines the axis
Antithesis HS_A > 0.7 good, > 0.1 valid Mode, Arousal
Aspect Ks T+/A+ > 0.4, T−/A− < 0.6 HS, K_T, K_A
Perspective area_normalized ~0.5 excellent, ~0.35 good rectangularity, CC, diagonals
Transition Feasibility > 0.7 feasible Insight, Proactiveness, HS
Cycle/Wheel Causality Probability Raw on the node; layer-relative % (sum = 1.0) computed for display