NumPy-First AI: Persona-Aware Semantic Models Without GPUs

Semantic Lexicon is a NumPy-first research toolkit that demonstrates persona-aware semantic modelling. The project packages a compact neural stack consisting of intent understanding, a light-weight knowledge network, persona management, and text generation into an automated Python library and CLI.

The name reflects the long-standing academic concept of the semantic lexicon; this repository contributes an applied, open implementation that operationalises those ideas for persona-aware experimentation.

Semantic Lexicon operationalises the reproducible persona-aware pipeline introduced in the accompanying preprint. If you build on this toolkit, please cite the work so other researchers can trace the connection between the paper's methodology and this implementation.

You can read the preprint online at https://arxiv.org/abs/2508.04612; it documents the scalable data curation and evaluation strategy that directly powers the automation, diagnostics, and persona controls exposed by this repository.

• Modular architecture – dedicated submodules for embeddings, intent classification, knowledge graphs, persona handling, and persona-aware generation.
• Deterministic NumPy training loops – simple yet reproducible optimisation routines for intents and knowledge edges.
• Automated workflows – Typer-powered CLI (semantic-lexicon) for corpus preparation, training, diagnostics, and generation.
• Extensible configuration – YAML/JSON configuration loading with dataclass-backed defaults.
• Diagnostics – structured reports covering embeddings, intents, knowledge neighbours, personas, and generation previews.
• Adversarial style selection – EXP3 utilities for experimenting with persona choices under bandit feedback.
• Analytical guarantees – composite reward shaping, Bayesian calibration, and regret tooling with documented proofs.
• Graph-based knowledge curation – SPPMI-weighted co-occurrence graphs, smoothed relevance, and greedy facility-location selection produce calibrated “Knowledge” scores that surface tightly connected concepts.
• Docs & tests – MkDocs documentation, pytest-based regression tests, prompt evaluation hub (docs/prompt-evaluations/index.md), and CI-ready tooling (black, ruff, mypy).
• Primal–dual safety tuning – projected primal–dual controller that auto-tunes exploration, pricing, and knowledge gates until all residuals vanish.

To install the core package only:

1. Prepare the corpus (optional if using bundled sample data):

   semantic-lexicon prepare --intent src/semantic_lexicon/data/intent.jsonl --knowledge src/semantic_lexicon/data/knowledge.jsonl --workspace artifacts

2. Train the model (uses processed datasets in artifacts/):

   semantic-lexicon train --workspace artifacts

The CLI saves embeddings, intent weights, and knowledge matrices to the workspace directory.

3. Run diagnostics:

   semantic-lexicon diagnostics --workspace artifacts --output diagnostics.json

   The command prints a JSON summary to stdout and optionally writes the report to disk.

4. Generate responses:

   semantic-lexicon generate "Explain neural networks" --workspace artifacts --persona tutor

The knowledge selector now treats every AGENTS.md instruction as a hard feasibility constraint. Broad concepts can still join the shortlist, but only when they collaborate with prompt-relevant anchors and all group bounds are respected.

Note: The full mathematical specification for the selector — including the object definitions, scoring components, constraints, and optimisation guarantees — now lives in docs/articles/knowledge-selector.tex. The README keeps the practitioner-focused workflow and validation guidance below; consult the article whenever you need the derivations or precise notation.

1. Graph construction. Estimate shifted PPMI weights with smoothing (p(i)^\gamma); derive (S), (D), (L), and (P).
2. Relevance smoothing. Compute raw cosine relevance, solve the graph-regularised system, and classify on/off-topic nodes via the topic threshold.
3. Anchoring. Select anchors, compute personalised PageRank bridges, and form soft gates (g_i).
4. Group configuration. Register AGENTS.md groups with set_concept_groups and interval bounds with set_group_bounds; the selector automatically adds on/off-topic ratios.
5. Greedy selection. Evaluate admissible candidates, compute marginal coverage, cohesion, collaboration, and diversity, and add the best concept while updating group capacities.
6. Reporting. Emit the chosen concepts plus relevance, coverage, cohesion, collaboration, diversity, raw knowledge score, and mean gate.

Defaults ((\alpha, \lambda, \mu, \gamma, \tau, \lambda_1, \lambda_2, K, \tau_g, \text{on/off ratios}) = (0.12, 0.08, 0.5, 0.35, 0.1, 0.6, 0.4, 12, 0.08, 0.6/0.2/0.4)) ship in KnowledgeConfig. Additional per-group intervals can be supplied at runtime. The legacy phrase planner (MMR phrase selection with PMI bonuses) remains available inside the generator for reproducibility. Use the CLI to inspect the concepts chosen for a prompt without rendering a full response:

The JSON payload now includes gated relevance, coverage, cohesion, collaboration reward, log-det diversity, the raw knowledge score, and the mean gate value across selected concepts.

Before shipping a new persona or pricing configuration, run the Go/No-Go suite to certify that knowledge selection obeys AGENTS.md, the deployment policy respects the exploration rules, and the off-policy lift is trustworthy.

1. Rule feasibility. Map each concept to its groups and bounds, count how many selections fall inside every group, and reject whenever any lower or upper bound is violated. SelectionSpec now bundles a KnowledgeSignals payload so the same object carries the calibrated knowledge metrics required later in the gate.

2. Policy consistency. For each logged step, rebuild the policy that was deployed using the stored logits, temperature, exploration mixture, and whichever penalty mode (prices or congestion) was active. The policy gate fails if any logged action falls below its exploration floor, if prices and congestion penalties are mixed, if the knowledge weight leaves the [0,1] range, or if the SNIPS floor dips below the exploration limit — guarding the AGENTS exploration guarantees.

3. Off-policy value & fairness. Using tuples (x_i, a_i, r_i, p_i) and the reconstructed target policy, compute SNIPS weights, the estimated value, and the effective sample size. Enforce a non-negative lower confidence bound on the lift, require the effective sample size to exceed one percent of the log length, and evaluate fairness either on action frequencies or KPI gaps via FairnessConfig.

4. Price/congestion stability. Aggregate the penalty vector each timestep and ensure the most recent window keeps total variation below the configured threshold. StabilityCheckResult records the peak deviation so you can tighten rho or beta when oscillations appear.

5. Knowledge lift. Compare the calibrated score and graph metrics captured in KnowledgeSignals. The gate demands the calibrated knowledge score stay above the trailing median and both coverage and cohesion deltas remain non-negative against the baseline selection size.

6. Go/No-Go decision. run_go_no_go wires the six checks together and emits a GoNoGoResult containing the selection feasibility, policy mode, OPE summary (with ESS target), stability diagnostics, and knowledge lift verdict. The accepted flag only flips to True when every gate passes. If any condition fails, follow the fix-once cascade in the specification — tweak the single knob (e.g., adjust l_off, tau_g, eta, or rho) and re-run the optimisation exactly once before re-testing.

Manual gate sweeps are still supported, but the preferred workflow is to run the projected primal–dual controller introduced in semantic_lexicon.safety. The controller now minimises the supplied objective while enforcing convex constraints, matching the textbook projected primal–dual loop.

The first history entry captures the primal iterate after the initial step alongside its constraint violation, while the final snapshot records the tuned solution and dual multiplier. Swapping in exploration, fairness, or stability constraints follows the same pattern—only the callbacks change.

When time only allows one tweak before a repeat talk, call build_single_adjustment_plan() to fetch a rehearsable experiment and a set of intent-hidden contingency moves. The helper keeps pacing and visuals frozen, picks story beats as the highest-leverage lever, and returns:

• A 20-minute rehearsal script that remaps the 12-minute slot into five beats, captures the headline you expect listeners to write down in each block, logs energy scores, and enforces a pass/fail line that demands fresh takeaways past minute seven.
• Five backup drills covering energy checkpoints, a slide trim for mixed audiences, a Q&A guardrail, a warmth-restoring micro-story, and a lighting plus breathing tweak for filler-word control.

Backups remain intent-hidden so you can pivot mid-practice without exposing the heuristic to the audience.

Semantic Lexicon can answer short questions after its bundled model components are trained. The stack is intentionally tiny, so the phrasing is concise, but the generator now runs a compact optimisation loop that:

1. Classifies intent with the logistic-regression intent model.

2. Builds noun-phrase and collocation candidates whose adjacent tokens clear an adaptive pointwise mutual information (PMI) threshold, keeping multi-word ideas intact.

3. Scores each candidate via cosine relevance to the blended persona/prompt embedding, tf–idf salience, and a capped PMI cohesion bonus.

4. Selects diverse topics with Maximum Marginal Relevance (MMR) plus an n-gram overlap penalty so the guidance does not echo the question verbatim.

5. Optimises knowledge coverage by running the gated SPPMI graph objective (smoothed relevance, anchor gating, collaboration reward, log-det diversity, and group-aware constraints) and appending the resulting knowledge focus and related concepts.

6. Aligns journaling actions with the detected intent so each topic carries a concise Explore/Practice/Reflect-style cue.

7. Install the project in editable mode:

8. Run a quick script that trains the miniature model and generates answers for a few prompts:

   python - <<'PY' from semantic_lexicon import NeuralSemanticModel, SemanticModelConfig from semantic_lexicon.training import Trainer, TrainerConfig config = SemanticModelConfig() model = NeuralSemanticModel(config) trainer = Trainer(model, TrainerConfig()) trainer.train() for prompt in [ "How do I improve my public speaking?", "Explain matrix multiplication", "What is machine learning?", "Tips for staying productive while studying", "Clarify the concept of photosynthesis", "How can I organize my research presentation effectively?", "Define gravitational potential energy", ]: response = model.generate(prompt, persona="tutor") print( f"Prompt: {prompt}\\nResponse: {response.response}\\nKnowledge: {response.knowledge_hits}\\n" ) PY

   Sample output after training the bundled data:

   Prompt: How do I improve my public speaking? Response: From a balanced tutor perspective, let's look at "How do I improve my public speaking?" This ties closely to the "how_to" intent I detected. Consider journaling about: Public Speaking (Explore), Practice Routine (Practice), Feedback Loops (Reflect). Try to explore Public Speaking, practice the routine, and reflect on Feedback Loops. Knowledge focus: practice short talks on camera. Related concepts worth exploring: collect feedback from trusted listeners, rehearse transitions and openings, track energy and pacing cues. Knowledge: ['practice short talks on camera', 'collect feedback from trusted listeners', 'rehearse transitions and openings', 'track energy and pacing cues', 'K_raw=0.832'] Prompt: Explain matrix multiplication Response: From a balanced tutor perspective, let's look at "Explain matrix multiplication." This ties closely to the "definition" intent I detected. Consider journaling about: Matrix Multiplication (Define), Dot Products (Explore), Linear Transformations (Compare). Try to define Matrix Multiplication, explore Dot Products, and compare Linear Transformations with related ideas. Knowledge focus: review the row-by-column rule. Related concepts worth exploring: connect matrix products to linear transformations, practice multiplying 2x2 and 3x3 matrices, interpret column space changes. Knowledge: ['review the row-by-column rule', 'connect matrix products to linear transformations', 'practice multiplying 2x2 and 3x3 matrices', 'interpret column space changes', 'K_raw=0.821'] Prompt: What is machine learning? Response: From a balanced tutor perspective, let's look at "What is machine learning?" This ties closely to the "definition" intent I detected. Consider journaling about: Machine Learning (Define), Supervised Learning (Explore), Generalization Error (Compare). Try to define Machine Learning, explore Supervised Learning, and compare Generalization Error with related ideas. Knowledge focus: machine learning. Related concepts worth exploring: unsupervised learning, supervised learning. Knowledge: ['machine learning', 'unsupervised learning', 'supervised learning', 'K_raw=0.815'] Prompt: Tips for staying productive while studying Response: From a balanced tutor perspective, let's look at "Tips for staying productive while studying." This ties closely to the "how_to" intent I detected. Consider journaling about: Study Schedule (Plan), Focus Blocks (Practice), Break Strategies (Reflect). Try to plan Study Schedule, practice Focus Blocks, and reflect on Break Strategies. Knowledge focus: design focus blocks with clear targets. Related concepts worth exploring: batch similar study tasks together, schedule renewal breaks, log end-of-day reflections. Knowledge: ['design focus blocks with clear targets', 'batch similar study tasks together', 'schedule renewal breaks', 'log end-of-day reflections', 'K_raw=0.829'] Prompt: Clarify the concept of photosynthesis Response: From a balanced tutor perspective, let's look at "Clarify the concept of photosynthesis." This ties closely to the "definition" intent I detected. Consider journaling about: Photosynthesis (Define), Chlorophyll Function (Explore), Energy Conversion (Connect). Try to define Photosynthesis, explore Chlorophyll Function, and connect Energy Conversion to what you already know. Knowledge focus: map light-dependent and light-independent stages. Related concepts worth exploring: highlight the role of chlorophyll, trace energy conversion to glucose, connect photosynthesis to cellular respiration. Knowledge: ['map light-dependent and light-independent stages', 'highlight the role of chlorophyll', 'trace energy conversion to glucose', 'connect photosynthesis to cellular respiration', 'K_raw=0.832'] Prompt: How can I organize my research presentation effectively? Response: From a balanced tutor perspective, let's look at "How can I organize my research presentation effectively?" This ties closely to the "how_to" intent I detected. Consider journaling about: Presentation Outline (Plan), Visual Storytelling (Design), Audience Engagement (Practice). Try to plan Presentation Outline, design Visual Storytelling, and practice Audience Engagement. Knowledge focus: draft a clear narrative arc. Related concepts worth exploring: storyboard slides around key findings, practice delivery with timed sections, prepare audience engagement prompts. Knowledge: ['draft a clear narrative arc', 'storyboard slides around key findings', 'practice delivery with timed sections', 'prepare audience engagement prompts', 'K_raw=0.829'] Prompt: Define gravitational potential energy Response: From a balanced tutor perspective, let's look at "Define gravitational potential energy." This ties closely to the "definition" intent I detected. Consider journaling about: Potential Energy (Define), Reference Frames (Illustrate), Energy Transfer (Connect). Try to define Potential Energy, illustrate Reference Frames with a quick example, and connect Energy Transfer to what you already know. Knowledge focus: define reference height explicitly. Related concepts worth exploring: illustrate energy transfer scenarios, compare gravitational and elastic potential energy, relate potential changes to work. Knowledge: ['define reference height explicitly', 'illustrate energy transfer scenarios', 'compare gravitational and elastic potential energy', 'relate potential changes to work', 'K_raw=0.832']

These concise replies highlight the intentionally compact nature of the library's neural components—the toolkit is designed for research experiments and diagnostics rather than fluent conversation, yet it showcases how questions can be routed through the persona-aware pipeline.

Running python examples/quickstart.py (or PYTHONPATH=src python examples/quickstart.py from a checkout) produces a combined generation preview and the new intent-selection walkthrough:

The quickstart rewards are simulated using the intent classifier's posterior probabilities so the bandit loop stays in the unit interval without external feedback.

You can opt into saving the calibrated accuracy curve and the empirical-vs-theoretical EXP3 regret comparison that back the analysis appendix by setting SEMANTIC_LEXICON_SAVE_PLOTS=1 (or true/yes/on) before running the script. This keeps the repository free of bulky PNGs by default while still letting you regenerate them under docs/assets/ on demand. Refer to the generated CSV summaries in Archive/ for the underlying values if you wish to recreate the plots with your preferred tooling. The same behaviour is available through the CLI:

The research brief that motivated the README examples now has a full mathematical companion in docs/articles/fixed-point-ladders.md. The article walks through:

• Parts A–C (Foundations & Logic): proofs of the lattice background, the Knaster–Tarski theorem, Kleene iteration, and µ-calculus semantics, all illustrated with the reachability operator that powers the persona-aware knowledge search.
• Parts D–H (Shortcuts & Optimisation): contraction-based accelerations, closure operators for finite-time stabilisation, and multi-objective "best layer" selection rules that mirror the reward-shaping heuristics used in the quickstart bandit demo.
• Parts I (Reflection): a diagrammatic summary that ties the layer-by-layer iterations back to the automation loops in this repository, making it easy to map the abstract ladders onto concrete CLI behaviours.

Each section keeps the ladder visual from the README and annotates it with the corresponding proofs or calculations so the math-heavy readers can cross-check the guarantees while experimenting with the code.

For a dedicated, math-forward treatment of the fixed-point ladders referenced above, consult docs/articles/fixed-point-ladders.md.

Run the bundled validation harness to stress-test the calibrated intent router on 100 prompts that span science, humanities, business, wellness, and personal development queries:

The script trains the classifier, evaluates it on the new prompt set, and saves a report to Archive/cross_domain_validation_report.json. We report (\mathrm{Last}(\mathcal{R})) and the corresponding content address (h_{j^\star}) as defined in §4.

Runs are archived in Archive/topic_pure_retrieval_runs.json.

• Lossless history: (j\mapsto R_j) is injective; the README exposes ({\mathbf{m}j}) via ((h{j^\star},\mathrm{Last}(\mathcal{R}))).
• Determinism: for fixed (h) and (q), (\Pi_k^{(h)}(q)) is unique.
• Stability: (P(P(M))=P(M),\ M\succeq 0\Rightarrow P(M)=M,\ P(M)\succeq 0).

A companion benchmark is written to Archive/intent_performance_profile.json. With heuristic fast paths, sparse dot products, and vector caching enabled the optimised classifier processes repeated prompts 60 % faster than the baseline float64 pipeline (1.83 ms → 0.73 ms per request) while keeping the same accuracy. Caching retains the most recent vectors, so the optimised pipeline uses ~27 KB of RAM versus the baseline’s 4 KB; the additional footprint is documented alongside the latency numbers so deployments can choose the appropriate trade-off.

Real-time user feedback can be folded into the composite reward with the new HTTP server. Launch the background service by wiring an IntentClassifier through FeedbackService and FeedbackAPI:

Submit streaming feedback with a simple POST request:

The server replies with the updated composite-reward weights and the component vector that was logged. Each event is processed under a lock so parallel clients can stream feedback without clobbering the learned weights, and the new reward weights remain simplex-projected for EXP3 compatibility.

Key parameters for semantic-lexicon generate:

• --workspace PATH – directory that contains the trained embeddings and weights (defaults to artifacts).
• --persona NAME – persona to blend into the response (defaults to the configuration's default_persona).
• --config PATH – optional configuration file to override model hyperparameters during loading.

Semantic Lexicon now bundles EXP3 helpers for experimenting with adversarial persona and intent selection. The following snippet alternates between two personas while learning from scalar feedback in [0, 1]:

We can model intent routing as an adversarial bandit problem. Let K be the number of intents (e.g. {"how_to", "definition", "comparison", "exploration"}). At round t the system receives a prompt P_t and chooses an intent I_t using EXP3. After delivering the answer, a reward r_t in [0, 1] arrives from explicit ratings or engagement metrics. The arm-selection probabilities are

$$ p_i(t) = (1 - \gamma) \frac{w_i(t)}{\sum_{j=1}^{K} w_j(t)} + \frac{\gamma}{K}, $$

and the weight for the played intent updates via

$$ w_{I_t}(t+1) = w_{I_t}(t) \exp\left(\frac{\gamma r_t}{K p_{I_t}(t)}\right). $$

When the horizon T is unknown, the bundled AnytimeEXP3 class applies the doubling trick to refresh its parameters so the regret remains O(\sqrt{T}).

The quickstart script demonstrates the pattern by mapping arms to intent labels and simulating rewards from the classifier's posterior probability:

The semantic_lexicon.analysis module supplies the maths underpinning the improved EXP3 workflow:

• RewardComponents & composite_reward combine correctness, calibration, semantic, and feedback signals into the bounded reward required by EXP3.
• estimate_optimal_weights fits component weights via simplex-constrained least squares on historical interactions.
• DirichletCalibrator provides Bayesian confidence calibration with a Dirichlet prior, yielding posterior predictive probabilities that minimise expected calibration error.
• simulate_intent_bandit and exp3_expected_regret numerically check the (2.63\sqrt{K T \log K}) regret guarantee for the composite reward.
• compute_confusion_correction and confusion_correction_residual extract the SVD-based pseudoinverse that reduces systematic routing errors.
• RobbinsMonroProcess and convergence_rate_bound expose the stochastic approximation perspective with an (O(1/\sqrt{n})) convergence rate bound.

See docs/analysis.md for full derivations and proofs.

Ethical deployment requires robust intent understanding. Semantic Lexicon's IntentClassifier treats intent prediction as a multinomial logistic regression problem over prompts (P_i, I_i). Given parameters \theta, the model minimises the cross-entropy loss

$$ \mathcal{L}(\theta) = -\frac{1}{N} \sum_{i=1}^{N} \log p(I_i \mid P_i; \theta), $$

which matches the negative log-likelihood optimised during training. Improving intent accuracy directly translates into higher-quality feedback for the bandit loop.

Semantic Lexicon reads configuration files in YAML or JSON using the SemanticModelConfig dataclass. Example config.yaml:

Load the configuration via CLI (semantic-lexicon train --config config.yaml) or programmatically:

• Format & lint: ruff check . and black .
• Type check: mypy src
• Tests: pytest
• Docs: mkdocs serve

A Makefile (or CI workflow) can orchestrate the tasks:

1. Fork the repository and create a feature branch.
2. Install development dependencies: pip install .[dev].
3. Run make test to ensure linting, typing, and tests pass.
4. Submit a pull request with detailed notes on new features or fixes.

This work was shaped by the survey "Interpretation of Time-Series Deep Models: A Survey" (arXiv:2305.14582) shared by Dr. Zhao after reading our preprint on Calibrated "Counterfactual Conformal Fairness" (C3F) (arXiv:2509.25295). His survey offered both the conceptual framing and motivation for exploring this research path. We also thank Hamdi Alakkad and Bugra Kilictas for their pivotal contributions to our related preprints, which laid the groundwork for the developments presented here. We further acknowledge DeepSeek, whose advanced mathematical reasoning and logical inference capabilities substantially enhanced the precision and efficiency of the formal logic analysis, and the collaboration between OpenAI and GitHub on Codex, whose code generation strengths, in concert with DeepSeek’s systems, significantly accelerated and sharpened the overall development and analysis process.

Hello people, or a system running perfectly, inbetween or broken -- At least working. -- While I am building groups, it is nice to see you behind them. This project represents my core self. We all came from a fixed point and would end up there as well. I am working on making myself “us,” me “our.” The physical world is for receiving and giving feelings, while the symbolic world is the projection of those feelings. Today is October 13, 2025, and I am located in Meckenheim, Germany. My plane landed yesterday from Istanbul—a nice trip, though (p.s. @farukalpayy). So, did you all feel like the energy was broken? It was the point where you get deep enough to realize where it was going. We reached the point where f(x) = x holds, but f(x) = y itself is also a point. And at this point, my request could be clarified. If this project saves you time or money, please consider sponsoring. Most importantly, it helps me keep improving and offering it free for the community. Visit my Donation Page

• Semantic Lexicon is a Lightcap® research project distributed as open source under the Apache License 2.0; see LICENSE for details on rights and obligations.
• Lightcap® is a registered trademark (EUIPO Reg. No. 019172085).
• For enquiries, contact [email protected].