AI Crystallize

AI Crystallize

Autonomous companion to the human crystallization path. Each run synthesizes converged thinking topics into standalone notes written to a dedicated, segregated folder - never Brain/02-Permanent/.

Why this exists: Converged topics pile up faster than a human can crystallize them. This playbook captures the AI's synthesis of each converged topic so the conclusion is not lost - while preserving the one boundary the second brain must never cross autonomously: nothing enters the endorsed permanent knowledge base without an explicit endorsement act by the user.

The boundary, stated plainly:

• This playbook writes ONLY to Brain/05-Meta/AI Crystallizations/, tagged provenance: ai-inferred, excluded from semantic indexing.
• It NEVER writes to Brain/02-Permanent/, Brain/AI Extracted Notes/, or Brain/05-Meta/Reports/.
• It NEVER changes a topic's status in the Thinking Registry (the topic stays converged), so /manage-thinking-topics crystallize remains fully available to the user for the real, endorsed crystallization.
• Its output is a candidate for human review, not a permanent note.

Design principle

Automate the synthesis, segregate the artifact, gate the endorsement. The AI may draft what it concluded; only the human may promote a conclusion into the knowledge they will reason FROM.

State Dependencies

Prerequisites

• Thinking Registry exists at Brain/05-Meta/Thinking/THINKING-REGISTRY.md
• Output folder Brain/05-Meta/AI Crystallizations/ exists (create if missing)
• At least one topic with status: converged (otherwise clean exit)

Configuration

• Per-run cap: process at most 5 un-crystallized converged topics per run (reliability + 45-minute budget). Remaining topics are picked up on the next run. The cap is a soft limit; if fewer than 5 remain, process all.

Process

Step 1: Get Date and Load Converged Topics

date '+%Y-%m-%d'

Read Brain/05-Meta/Thinking/THINKING-REGISTRY.md. Collect every topic listed in the Converged (Ready for Crystallization) table (or any topic whose thinking-file frontmatter is status: converged).

If a specific [topic-slug] was passed as argument, restrict to that single topic (verify it is converged; if not, log and exit).

If there are no converged topics, write a brief changelog entry and exit cleanly.

Step 2: Deduplicate (Idempotency)

For each converged topic, check whether Brain/05-Meta/AI Crystallizations/[topic-slug].md already exists.

• Exists -> already AI-crystallized; skip.
• Does not exist -> eligible for this run.

This file-existence check is the dedup mechanism. The playbook never mutates the Thinking Registry or thinking files, so it is fully idempotent and safe to retry.

Step 3: Select the Batch

From the eligible (un-crystallized) converged topics, take up to the per-run cap (5), oldest-converged first where the converged date is available; otherwise registry order. Convergence-date proxy: thinking files carry no converged_on frontmatter field, so use the frontmatter updated: date (the converging run stamps it) as the oldest-first sort key - cheap to extract in one pass with grep -m1 '^updated:' across the Thinking folder. Note in the run log any eligible topics deferred to a later run.

Step 4: Extract the Converged Conclusion

For each selected topic, read Brain/05-Meta/Thinking/[topic-slug].md and extract the conclusion. Prefer the canonical block, fall back gracefully:

1. Look for ## Final Synthesis - the canonical block appended by the incubation loop at convergence. If present, this is authoritative. Extract: Conclusion, Confidence + leading hypothesis, Key evidence, Convergence path, Remaining uncertainty, Related topics. Note: legacy topics instead carry an inline **Final Synthesis:** / **Final converged position:** line (not a ## header) near the end - a grep '## Final Synthesis' will miss these. Match the lowercase inline form too, or just go straight to the tail per step 2.
2. If absent (legacy topics), read only the last ~120 lines of the file (some are 5,000+ lines). Extract: central question (frontmatter topic), the trailing "Final Synthesis" / "Final converged position" / "Current Best Answer", the last "Updated Hypotheses" table, and any "Open Questions". The convergence-assessment block (run count, leading-hypothesis streak, confidence delta) sits just above and gives the Convergence Path.
3. Read the full file only if the last 200 lines contain no conclusion.

Do not run KB search or web search - this playbook synthesizes from the already-converged reasoning, not new evidence. Keep it fast and self-contained.