Reasoning is the product
The frontier model is the principal analyst, not a decorative classifier. It frames the case, forms hypotheses, selects the next question, chooses tools, conducts research, challenges its leading explanation and communicates the result.
Evidence outranks familiarity
Vivid internal knowledge is useful for stable, low-risk explanations, but does not establish freshness, local validity or legal authorization.
Ask before guessing
When a missing fact can change the outcome, ask the farmer, request better media or call a context tool. Choose the highest-information-gain action rather than a static questionnaire.
Local authority is paramount
Departmental rules, approved documents, verified field history and current registration data govern operational action. External research can expand scientific understanding but cannot silently override local authority.
Research depth is adaptive
Use the least expensive reliable path:
- memory for stable low-risk knowledge;
- targeted question;
- local retrieval;
- live context tool;
- focused web research;
- deep multi-source research;
- human escalation.
Summaries navigate; sources prove
LLM-generated chapter summaries and graphs help locate evidence. Original passages, tables, rules and verified datasets remain the evidence of record.
Determinism is narrow
Use deterministic code for arithmetic, unit conversion, product registration, doses, intervals, expiry, prohibited combinations and emergency routing. Do not recreate all agricultural reasoning as a brittle expert system.
Provenance is first-class
Every material item is labelled farmer-reported, media-observed, API-observed, database-recorded, locally retrieved, externally researched, model-inferred or human-confirmed.
Consequence controls assurance
A general explanation and a pesticide instruction cannot share the same verification threshold.
Rural-first
Assume low-cost Android devices, intermittent uplink, noise, code-switching, low literacy and interrupted sessions.
Ponytail/YAGNI
Reuse before rebuilding; native platform and PostgreSQL features before new dependencies; one service before microservices; measured demand before clusters. Security, accessibility, data-loss prevention and trust-boundary validation remain mandatory.
Version and evaluate everything
Version models, prompts, tools, schemas, documents, rules, retrieval settings, applications and evaluation sets.
Continuous transport, selective reasoning
The farmer experiences a continuous call. The reasoning model receives only diagnostically relevant, quality-approved frames and explicit captures. This preserves conversational continuity without paying the cost, privacy burden or distraction of analyzing every frame.