A practical framework for mapping color, sentiment, and symbolism into computational form — with reproducible stats, evaluators, and generative samplers.
La Matriz Consulting Color Oracle GPT treats color as a co-expressive interface—linking “inner knowing” with outer data, using accessible math to create free-choice maps of symbolic resonance. Core hypothesis: embedding symbolic color transitions inside LLM prompts improves alignment for clarity, creativity, and relevance.
- Why this repo
- Repository structure
- Methods
- Key results
- Interpretation
- Limitations
- Roadmap: RAGs & agents
- Roadmap: Sentiment anchoring
- Roadmap: Persistent homology (TDA)
- Credits & links
This project operationalizes Semantic Color Math—a way to encode symbolic/emotional meaning using color distributions and transitions. It offers a reproducible pipeline to:
- derive discrete (frequency) vs continuous (sentiment-weighted) color PDFs,
- generate color sequences via different samplers,
- evaluate coherence with statistical tests, and
- fit simple predictive models from RGB → meaning/sentiment.
Philosophically, the system was built to reclaim the technological field into emotional logic and to “map sentimental nuance into computational syntax” rather than replicate extractive metrics.
Semantic-Color-Math/
├─ analysis/ # analysis & plots
├─ applications/ # fun stuff
├─ core/ # package code
├─ visualizations/
├─ data/ # color names, RGB, and sentiment scores
├─ LICENSE
└─ README.md
-
Discrete weight
$w_d(c)=\frac{\text{count}(c)}{\sum_c \text{count}(c)}$ . -
Continuous weight starts with summed sentiment per color
$s(c)=\sum \text{sentiment}_i$ . To obtain a valid PDF we min-max scale to$[0,1]$ then renormalize
- Random (uniform over RGB or color names)
-
Discrete (draw by
$w_d$ ) -
Continuous (draw by
$w_s$ )
- RGB variance across a sequence (lower = more visually coherent).
-
Two OLS regressions:
-
RGB → Color-Word Similarity (semantic):
$R^2=0.174$ ; G and B positive, R ns. -
RGB → Sentiment Score (emotional valence):
$R^2=0.380$ ; B positive, G negative, R ns.
-
RGB → Color-Word Similarity (semantic):
-
Sequence coherence t-test: La Matriz vs Random.
- Purple: not most frequent, but dominant in sentiment weight.
- Brown, Gold: reasonable frequency but negative continuous weight.
- Grey: most frequent; sentiment-neutralish.
- Blue/Green: aligned between frequency and positive sentiment.
- Rescaling retains sentiment contrasts; purple remains the emotional standout.
| Method | RGB Variance |
|---|---|
| Random | 8131.33 |
| Discrete | 5696.33 |
| Continuous | 6437.67 |
- Mean coherence (La Matriz): 0.573
- Mean coherence (Random): 0.665
- t = −0.604, p = 0.554 → fail to reject H₀ (under this metric).
- Tight, non-overlapping intervals indicate statistically distinct color groups (e.g., Cluster 0: high-yellow; Cluster 3: pure blue/purple).
-
RGB → Semantic similarity:
$R^2=0.174$ . G and B significant (+); R ns. -
RGB → Sentiment:
$R^2=0.380$ . B significant (+), G significant (−), R ns. Diagnostics: residuals slightly non-normal; DW ~1–1.3.
- As color is not statistically significant for sequence coherence, but it is statistically significant for predicting sentiment, the map of RGB color space is a promising tool for better aligned prompt responses from AI using the dataset.
- It is not so precise that it waters down nuance of sentiment, but it is semantically relevant.
- Frequency ≠ meaning. Purple is emotionally potent; grey is ubiquitous but not sentiment-rich.
- Cool colors carry meaning. Green/blue predict semantic resonance; blue strongly predicts positive sentiment.
- Coherence metric matters. The negative result on the t-test likely reflects a metric (pairwise color similarity) that underrates semantic or narrative cohesion.
These choices support my thesis that color can be a semantic interface and that symbolic transitions shape alignment beyond raw tokens.
- Current coherence relies on visual proximity; it may miss emotional or metaphoric structure.
- Linear OLS may underfit (no interactions or nonlinearity).
- Sentiment scores require anchoring across datasets to avoid drift.
- Corpus: color metaphors, cultural notes, user reflections, cluster summaries, and evaluation rubrics (BLEU/ROUGE/GPTEval + qualitative).
- Palette Agent: proposes sequences given goal sentiment/theme.
- Retriever: injects top-k color symbols & transitions into the prompt.
- Critic/Evaluator Agent: scores output against semantic anchors and rubric; triggers revisions.
- Sampler Agent: runs discrete vs continuous vs hybrid sampling; logs outcomes.
Goal: closed-loop improvement where each generation updates the KB with outcomes and human ratings (future on-chain UBI/SoulCoin concept lives here as a platform extension).
Why: stabilize sentiment across corpora/models; prevent drift.
-
Anchor sets: fixed exemplars per pole (e.g., {purple, blue, green}+, {brown, gold, red}− based on findings).
-
Calibration:
- Z-scoring within batch;
- Temperature/Platt scaling against anchor distributions;
- Bayesian hierarchical model: color-level priors, sequence-level posteriors.
-
Drift detection: monitor KL divergence of live scores vs. anchored baseline.
-
Hybrid weight:
$w_\lambda = \lambda w_s + (1-\lambda) w_d$ with$\lambda$ tuned by validation against human ratings.
Goal: capture shape of the color–sentiment manifold beyond pairwise similarity.
- Embed colors (and sequences) in
$\mathbb{R}^n$ (e.g.,$[R,G,B,\text{sentiment},\text{cluster}]$ ). - Build Vietoris–Rips filtrations over distance thresholds.
- Compute H₀/H₁/H₂ barcodes with
ripser/giotto-tda. - Define a Topological Cohesion Score: area under barcode for H₀ persistence minus spurious H₁/H₂ noise (or Bottleneck/Wasserstein distances to a curated “ideal” manifold).
- Compare La Matriz vs. Random: expect fewer small components (more connectivity) and stable 1-cycles reflecting meaningful transitions.
- WIP, TBA
Research & writing: La Matriz Consulting
Oracle (ChatGPT custom GPT): https://chatgpt.com/g/g-686180b5d47c8191bdbb1922b79a2ebb-la-matriz-consulting-color-oracle
Website: https://lamatriz.ai
Contact: danielle.gauthier@lamatriz.ai
“Not an answer — a mirror.”