Knowledge Crystallization Cycle (KCC)

(proposed, single-source — same position paper as Nurture First Development; not yet reviewed against this wiki’s own patterns/rules split. Pending synthesis…

Knowledge Crystallization Cycle (KCC)

(proposed, single-source — same position paper as Nurture-First Development (NFD); not yet reviewed against this wiki’s own patterns/rules split. Pending synthesis check before any of “Our Stack Implementation” below is treated as adopted.)

The core developmental mechanism of Nurture-First Development (NFD). Converts fragmented conversational knowledge into structured, reusable Skill Layer assets. Operationalizes Nonaka & Takeuchi’s tacit-to-explicit knowledge conversion within an agent memory context.

---

Four Phases

Phase 1: Conversational Immersion
    Operational dialogue; knowledge transferred implicitly.
    No explicit "teaching sessions" required.
    ↓
Phase 2: Experiential Accumulation
    Fragments tagged and logged in Experiential Layer.
    Six categories: [OPERATIONAL], [REASONING], [PATTERN],
                    [ERROR], [CONTEXT], [INSIGHT]
    ↓
Phase 3: Deliberate Crystallization (surgical workspace)
    Scope filter → pattern extraction → human review →
    de-contextualize → validate against full corpus → promote
    ↓
Phase 4: Grounded Application
    Crystallized knowledge in active service.
    Generates new experience at higher baseline.
    Contradictions trigger re-crystallization.

Each revolution of the cycle raises the knowledge baseline. The spiral doesn’t close — it expands.

---

Algorithm 1: Crystallization Process

Input: Knowledge state K = (C, S, E), scope θ
Output: Updated state K'

1. D ← ScopeFilter(E, θ)          # select relevant Experiential entries
2. P ← ExtractPatterns(D)          # automated pattern detection
3. P* ← HumanReview(P)             # user validates — the critical safeguard
4. for each validated pattern p ∈ P*:
5.     k ← Structure(p, S)         # format for Skill Layer
6.     k ← Decontextualize(k)      # generalize beyond original context
7.     if Validate(k, E):          # check against full corpus, not just scope
8.         S ← S ∪ {k}            # promote to Skill Layer
9.         E ← E \ {entries consolidated into k}
10. return K' = (C, S', E')

Non-regression guarantee: H(S’) ≥ H(S) — Skill Layer information content never decreases after validated crystallization. Human review (line 3) + corpus validation (line 7) together prevent false patterns from being promoted.

---

Six Experiential Fragment Categories

Tag	Content	Crystallization path
[OPERATIONAL]	Decisions made, actions taken, outcomes	Contributes to aggregate statistics; rarely crystallized individually
[REASONING]	Decision logic, assumptions, alternatives considered	High crystallization value; reasoning traces improve efficiency
[PATTERN]	Regularities across experiences (“every time X, Y tends to follow”)	Primary crystallization target; becomes Skill reference content
[ERROR]	Mistakes + corrective principles	High-priority crystallization target; becomes prevention rules
[CONTEXT]	Environmental metadata, domain-state annotations	Enriches retrievability of other fragments
[INSIGHT]	Standalone principles articulated during conversation	Often immediately promotable to Skill Layer; don’t need accumulation

Insight fragments may bypass accumulation entirely — promote directly if standalone. Error records need accumulation before patterns emerge. Operational records rarely justify individual crystallization.

---

Crystallization Triggers

Mode	When	Best for
Scheduled	Regular intervals (weekly, monthly)	Steady experiential accumulation; predictable domains
Threshold	Volume > N uncrystallized entries OR performance drift detected	Domains with bursty activity
Event-driven	After significant domain events (regime change, phase transition, major decision)	Strategic, advisory, research domains

---

Crystallization Efficiency

Three factors determine how much structured knowledge emerges from a given experiential corpus:

1. Experience diversity — varied situations crystallize better than repetitive data; a corpus of 50 distinct cases > 200 variations of the same case
2. Annotation quality — [REASONING] traces crystallize efficiently; bare [OPERATIONAL] records don’t
3. Pattern density — domains with strong regularities (financial patterns, legal precedent) yield higher efficiency than idiosyncratic domains

---

Our Stack Implementation

We run partial KCC. Mapping:

KCC Phase	Our Implementation	Gap
Immersion	Daily Claude Code / OpenCode / Pi sessions	None — happens naturally
Accumulation	mistakes/raw-log.md (hook-captured)	Only [ERROR] captured; [INSIGHT]/[PATTERN] not tagged
Crystallization	synthesize-mistakes skill → global-prevention-rules.md	Scope limited to errors; no Skill Layer promotion path
Grounded Application	wiki-context skill (retrieves crystallized knowledge)	Incomplete — wiki-context retrieves wiki pages but not global-prevention-rules.md inline

Proposed fix: Add [INSIGHT] and [PATTERN] tagging to AGENTS.md (all harnesses). When tagged fragments accumulate, route to wiki ingest rather than synthesize-mistakes. The wiki ingest pipeline IS the crystallization algorithm — it already implements Algorithm 1 via: pdf-ingest / pre-digest (scope+extract) → comprehension questions (human review) → wiki page (de-contextualize+promote).

---

Related Pages

• Nurture-First Development (NFD) — parent paradigm; three-layer architecture; our-stack synthesis
• Compounding Knowledge Base — KCC is the formal mechanism behind wiki compounding
• Agent Self-Correction — wiki-as-oracle = Grounded Application phase
• Agent Skills — Skill Layer is the destination for crystallized patterns
• Agentic Memory Tool — Experiential Layer implementation