Published by Justin Harris — Troponin Supplements — March 2026
Architecture
TroponinIQ runs on a standard LLM inference stack (currently Claude Opus 4.6) with no model modifications. The specialization comes entirely from two files injected at inference time:
constants.ts (~1000s of lines, ~100s KB) — The decision engine. Contains 135 exported constants encoding coaching heuristics, macro calculation defaults, body composition references, product knowledge, check-in thresholds, and edge-case handling rules. Every constant is sourced and annotated.
prompts.ts (~1000s lines, ~100s KB) — The persona engine. Contains the complete coaching personality, communication style, and instructional framing. Governs how knowledge is applied and communicated.
The separation is deliberate. Constants encode what the expert knows (data). Prompts encode how the expert communicates (style). You can update coaching rules without touching the voice, or refine the voice without changing the rules. This is the key architectural advantage over monolithic fine-tuning.
Training Methodology
The knowledge base is built through a structured extraction pipeline we call Recursive Knowledge Distillation. The process:
• Source ingestion — Raw coaching content is collected: books, Q&A transcripts, video transcripts, lecture notes, client interaction patterns. Project source code contains a "New Knowledge Library" where source documents can be added as desired, with nightly directory crawls, and website/social bots extracting new source data as it's produced.
• Pattern extraction — Each source is analyzed for coaching patterns: rules of thumb, decision thresholds, heuristics, edge-case handling, and philosophical commitments. These are classified as either constants (data) or prompt instructions (style).
• Encoding — Patterns are encoded into the appropriate file with source annotations, TypeScript types, and descriptive comments that make each constant self-documenting.
• Regression testing — Every new pattern gets a corresponding test case in TEST_PROMPTS.md (can assign weighted value to improve process time, etc). Each test specifies a scenario, an input prompt, and expected behavioral characteristics of the response.
• Teach-back verification — The human expert probes the system with questions designed to surface misunderstandings. If the system can’t explain back why a rule exists, the encoding is refined.
This is a daily cycle, not a quarterly project. Each optimization session adds 50–200 lines of structured knowledge. After three runs over two days, the knowledge base grew from ~100 lines of generic config to 3,278 lines of dense coaching intelligence.
Scaling Characteristics
The SuperTrop architecture scales along three axes:
Knowledge Density
Because the knowledge base is structured TypeScript (not unstructured text), it has near-zero redundancy. Every line encodes a decision rule or data point. Compare this to RAG-based approaches where retrieval quality degrades with corpus size, or fine-tuning where new knowledge can catastrophically interfere with existing capabilities. In the SuperTrop model, knowledge is additive and non-interfering by construction.
Source Parallelism
Knowledge extraction from different sources is embarrassingly parallel. Video transcripts, book chapters, Q&A sessions, and client interaction logs can all be processed simultaneously by separate extraction agents. The merge step is trivial because the knowledge base has typed structure — a new constant either fits the existing type system or it doesn’t. There are no embedding space collisions.
Domain Replication
The most significant scaling property: the entire methodology is domain-agnostic. The extraction pipeline, test framework, and optimization cycle work identically for any domain where a human expert exists. A physical therapist, an immigration lawyer, a master sommelier — anyone with deep tacit knowledge that can be extracted through structured conversation. The investment is in the expert’s time, not in compute.
|
The knowledge base grew from ~100 lines to 3,278 lines across three optimization sessions in two days. At this rate, a single domain expert can encode decades of experience into a deployable AI system within weeks, not months. |
Why Not Fine-Tuning?
Fine-tuning a model on coaching data would produce a system that is:
• Opaque — You cannot inspect which specific rules the model learned, or trace a bad output to a specific training example.
• Irreversible — You cannot selectively remove a bad heuristic without retraining.
• Expensive — Each iteration requires a full training run ($1,000–$50,000 depending on model size).
• Slow — Training cycles are measured in hours or days, not minutes.
The knowledge-base approach produces a system that is fully auditable (every rule is a named constant with a source annotation), instantly reversible (revert a line of code), effectively free to update (edit a file), and iterates in real-time (minutes per optimization cycle).
The tradeoff is that you need a frontier model for inference, which costs more per-query than a fine-tuned smaller model. For the foreseeable future, the auditability and iteration speed advantages massively outweigh this cost differential for any domain where accuracy matters.
What’s Next
We’re building toward a fully automated nightly optimization cycle: source ingestion at midnight, knowledge extraction at 12:30 AM, staging deployment at 1:00 AM, automated verification at 1:15 AM, and production promotion at 1:30 AM. TroponinIQ wakes up smarter every morning.
The longer-term convergence is with on-device model training (the Apple Neural Engine work demonstrated by maderix/ANE). The knowledge base becomes the training dataset for a small, locally-hosted inference model — a 3B parameter distillation of the expert’s knowledge running on commodity hardware at negligible cost. The frontier model continues as the extraction engine. This is not speculation; the infrastructure exists today.
TroponinIQ is live at TroponinIQ.com. The knowledge base is open for inspection. The methodology is replicable. The future of domain-expert AI is not bigger models. It’s smarter knowledge.