1. EXP-001

   General-purpose LLM vs constrained architecture for compliance agents

   Hypothesis: a generic LLM can draft MyInvois e-invoice payloads end-to-end. Result: it hallucinated a permit number on a real test invoice. Lesson: we needed a constrained architecture — not a bigger model.

   Learned
2. EXP-002

   Monolithic vs edge-native architecture for clinic ops latency

   Hypothesis: a single Node service handles Malaysian clinic peak-concurrency fine. Result: edge-native moved the slowest call from 480ms to 90ms p99 without rewriting the core. Lesson: the risk lives in the lab, not in your production.

   Shipped
3. EXP-003

   Bahasa Malaysia web-novel tokenization for author royalties

   Hypothesis: standard sentence-splitting handles Bahasa colloquial register. Result: code-switching broke it consistently. Lesson: we needed a Malaysian-trained tokenizer, not an English-default one.

   Learned
4. EXP-004

   Employment Act-aligned bilingual contract generation

   Hypothesis: a single Anthropic-family model writes bilingual Employment Act 1955 contracts aligned to EPF / SOCSO / EIS / PCB. Status: in active build. Lesson so far — the failure mode is hallucinated clause numbers, not fluency.

   In build

These are abbreviated public-facing entries. The full internal logs reference real test invoices, latency profiles, and tokenizer token mismatches — shared with serious prospects under NDA before any pilot.