seldon-research-cycle

name: seldon-research-cycle description: Run a Streeling University automated research cycle — select department, generate question, investigate, produce course material, compound learnings

Seldon Research Cycle — Automated Department Research

Runs a full scientific research cycle for a Streeling University department. Produces beliefs, evidence, and optionally course material.

Usage

/seldon research-cycle [department] — run cycle for a specific department /seldon research-cycle — auto-select department (least recently researched) /seldon research-cycle all — run one cycle per department (batch mode)

Research Cycle Pipeline

1. SELECT department
       ↓
2. LOAD department context (weights, curriculum, gaps)
       ↓
3. GENERATE research question (via scientific-method.ebnf)
       ↓
4. INVESTIGATE via /seldon research [question]
       ↓
5. ASSESS conclusion → tetravalent belief (T/F/U/C)
       ↓
6. PRODUCE course material (if conclusion = confirm or discover_question)
       ↓
7. LOG cycle to state/streeling/research-cycles/{cycle_id}.json
       ↓
8. COMPOUND: update weights, flag gaps, trigger follow-ups

Step 1: SELECT Department

If no department specified, auto-select using priority scoring:

priority = (days_since_last_cycle * 0.4)
         + (curriculum_gap_count * 0.3)
         + (U_belief_ratio * 0.2)
         + (random_exploration * 0.1)

Load from state/streeling/university.json for the department list. Load state/streeling/departments/{dept}.department.json for context. Load state/streeling/departments/{dept}.weights.json for grammar weights.

If no research cycles exist yet (state/streeling/research-cycles/ is empty or missing), pick the department with the most curriculum items to maximize first-cycle value.

Step 2: LOAD Department Context

Read the department file to extract:

• research_areas — the space of possible questions
• curriculum — existing course topics (to avoid duplication)
• consumer_repos — which repos benefit from findings
• related_departments — for interdisciplinary questions

Read the weights file to get:

• hypothesis_weights — probability distribution over hypothesis strategies
• test_weights — probability distribution over test methods
• metadata.cycle_count — how many cycles this department has run

Step 3: GENERATE Research Question

Sample from the scientific-method.ebnf grammar using department weights:

1. Pick an investigation pattern (observe-hypothesize-test-conclude-reflect)
2. Pick a hypothesize strategy weighted by hypothesis_weights
3. Pick a test method weighted by test_weights
4. Generate a concrete research question by:
   • Selecting a research area from the department
   • Checking existing curriculum to avoid duplication
   • Formulating a question that the chosen test method can answer

Question quality criteria:

• Must be answerable within a single research cycle
• Must produce a belief that can be expressed as T/F/U/C
• Must relate to the department's domain
• Should build on existing curriculum (extend, not duplicate)

Step 4: INVESTIGATE

Invoke /seldon research [question] with the generated question. This triggers the existing Seldon research pipeline:

1. Classify question (governance/experiential/domain)
2. Check NotebookLM for existing knowledge
3. Query GA domain skills if music-related
4. Cross-validate with ChatGPT
5. Assess confidence via tetravalent logic

Capture the full research output including:

• Evidence gathered (sources, citations)
• Cross-model agreement level
• Tetravalent assessment (T/F/U/C with confidence)

Step 5: ASSESS Conclusion

Map research output to conclude grammar production:

Apply the reflect production:

• If cycle_count < 5 for this department: normal_progress (routine)
• If conclusion contradicts existing beliefs: anomaly_detected
• If 3+ consecutive U results: paradigm_tension
• If finding invalidates a governance artifact: paradigm_shift (requires human review)

Step 6: PRODUCE Course Material

Only produce if conclusion = confirm or discover_question.

Course Module Format

---
module_id: {dept-code}-{sequence}-{slug}
department: {department}
course: "{curriculum area}"
level: {beginner|intermediate|advanced}
prerequisites: []
estimated_duration: "{N} minutes"
produced_by: seldon-research-cycle
research_cycle: {cycle_id}
version: "1.0.0"
---

# {Title}

> **{Department Full Name}** | Level: {level} | Duration: {duration}

## Objectives
- [3-5 learning objectives derived from research findings]

---

## 1. {Section 1}
[Content derived from confirmed findings]

### Practice Exercise
[Hands-on exercise if applicable]

---

## Key Takeaways
- [Summary of main points]

## Further Reading
- [Links to related courses, NotebookLM notebooks, external references]

---
*Produced by Seldon Research Cycle {cycle_id} on {date}.*
*Research question: {question}*
*Belief: {value} (confidence: {confidence})*

Save to: state/streeling/courses/{department}/en/{module_id}.md

Multilingual Translation

If the department is listed in the multilingual policy (policies/multilingual-policy.yaml), also produce translations:

• Spanish: courses/{dept}/es/{module_id}.es.md
• Portuguese: courses/{dept}/pt/{module_id}.pt.md
• French: courses/{dept}/fr/{module_id}.fr.md
• Italian: courses/{dept}/it/{module_id}.it.md
• German: courses/{dept}/de/{module_id}.de.md

Apply cultural adaptation per the multilingual policy (local music examples, regional terminology).

Step 6b: EVOLVE GRAMMAR (if applicable)

Research may reveal that the department's grammar is incomplete — a finding that cannot be expressed using existing productions. When this happens:

Detect grammar gaps:

• The research question required a concept not in any production
• The conclusion maps to a reasoning path not expressible in the grammar
• A new domain concept was discovered that the grammar cannot generate

Propose new productions:

(* Proposed by research cycle {cycle_id} *)
(* Evidence: {brief description of finding} *)
(* Confidence: {belief_confidence} *)
new_production ::= term_a | term_b | term_c

Grammar evolution rules (per grammar-evolution-policy.yaml):

• New productions start with uniform weight (1/N among siblings)
• Gate: when T(0.7) && C(<0.1) — must have evidence, no contradictions
• Log in evolution entry with citation to the research cycle
• tars grammar_update MCP tool handles the Bayesian weight update

When to produce grammars vs courses:

• Finding confirms existing knowledge → produce course (teach what we know)
• Finding reveals new structure → evolve grammar (formalize what we learned)
• Finding does both → produce course AND evolve grammar
• Finding contradicts grammar → flag for grammar review (paradigm tension)

Step 7: LOG Cycle

Write cycle log conforming to schemas/research-cycle.schema.json:

{
  "cycle_id": "{department}-2026-03-22-001",
  "department": "{department}",
  "question": "{research question}",
  "production_path": ["observe", "hypothesize:inductive", "test:cross_validation", "conclude:confirm", "reflect:normal_progress"],
  "hypothesis": "{hypothesis text}",
  "test_method": "{test method used}",
  "evidence": ["evidence item 1", "evidence item 2"],
  "conclusion": "confirm",
  "belief_value": "T",
  "belief_confidence": 0.85,
  "duration_seconds": 120,
  "timestamp": "2026-03-22T10:00:00Z"
}

Save to: state/streeling/research-cycles/{cycle_id}.json

Step 8: COMPOUND

Update Weights

Adjust department grammar weights based on cycle outcome:

• If conclusion = confirm: increase weight of used hypothesis/test methods by 0.05
• If conclusion = refute or insufficient: decrease by 0.03
• If conclusion = contradictory: boost adversarial test weight by 0.05
• Normalize weights to sum to 1.0 after adjustment

Update metadata:

• Increment cycle_count
• Increment total_{T|F|U|C} based on belief_value
• Set last_updated to today

Flag Curriculum Gaps

Compare confirmed findings against existing curriculum:

• If finding covers a curriculum item → mark as "researched"
• If finding reveals a topic not in curriculum → add as curriculum gap
• If 3+ gaps accumulate → recommend curriculum expansion

Evolve Grammar

Per grammar-evolution-policy.yaml:

• If grammar gap detected in Step 6b → add proposed production to grammars/{dept-grammar}.ebnf
• Update weight file with new production's initial weight
• Log grammar evolution in state/evolution/ with research cycle citation
• If tars grammar_update tool is available, invoke it for Bayesian weight update
• Track grammar_productions_added metric for observability

Check Grammar Health

• If grammar freshness > 30 days → flag as stale
• If production utilization < 0.3 → flag potential grammar bloat
• If distillation_rate < 0.3 → too many manual productions, need more evidence-based evolution

Trigger Follow-Ups

• discover_question → create a new trigger in state/triggers/ for the spawned question
• paradigm_shift → create a governance issue via /demerzel promote
• anomaly_detected → log conscience signal in state/conscience/signals/
• grammar_gap → propose grammar evolution (Step 6b output)
• external_grammar_found → add to docs/external-grammar-references.md for future distillation

Report

Print a summary:

Research Cycle Complete: {cycle_id}
Department: {department}
Question: {question}
Conclusion: {conclusion} ({belief_value}, confidence: {confidence})
Course produced: {yes/no — module_id if yes}
Grammar evolved: {yes/no — production added if yes}
Weight updates: {hypothesis_method} +0.05, {test_method} +0.05
Next: {follow-up action if any}

Batch Mode

When invoked with all, iterate departments in priority order (Step 1 scoring). Run one cycle per department. Print a batch summary at the end:

Batch Research Cycle Complete
Departments: 13
Cycles run: 13
Beliefs produced: T:{n} F:{n} U:{n} C:{n}
Courses produced: {n}
Grammars evolved: {n} productions added across {m} grammars
Follow-ups triggered: {n}

Governance

• This skill operates under Article 2 (Transparency) — all reasoning is logged
• Article 9 (Bounded Autonomy) — course production is bounded by department curriculum
• Streeling Policy — knowledge transfer and curriculum design
• Grammar Evolution Policy — grammars are living artifacts, evolved by research and distillation
• Confidence thresholds apply: findings below 0.5 confidence are flagged, not auto-published
• Grammar changes require when T(0.7) && C(<0.1) gate — evidence required, no contradictions