Status: pre-launch. This page describes target design; this document does not establish a deployed service.

A5 -- Dream Cycle

Current reading — 2026-09-27

Cortex preserves inspectable canonical lineage. Derived summaries, indexes and learned policies are replaceable and traceable. Compression proposals may not erase dissent, widen permissions or authorize themselves. Scheduling, retention, hard pruning and trained adaptation remain separately owned decisions. [I]

This is an editorial boundary, not a new runtime contract, upstream repeal or proof of adoption. See current status and the product roles. Original revision dates remain as recorded below.

Historical research draft — retained on 2026-09-27; not current terms or implementation instructions

Overview

The Dream Cycle is neural pruning for the Cortex. It runs periodic L6 (Executive-level) compression on accumulated traces, identifies patterns, prunes noise, strengthens signal, and outputs a dream report. It is the organism's equivalent of sleep.

Without the Dream Cycle, the Cortex grows without bound, signal-to-noise degrades, and the organism loses the ability to distinguish important patterns from accumulated debris.

Why L6

L6 (Executive caste) is explicitly non-deployable -- Titans set boundary conditions but never operate at runtime. The Dream Cycle is the one sanctioned use of L6 reasoning: offline, non-operational, applied to accumulated history rather than live signals.

During the Dream Cycle, L6's pathology (abstraction death) becomes a feature. The point is to abstract away operational detail and surface structural patterns.

The Cycle

Phase 1: Harvest

Collect all Cortex traces since the last dream cycle:

• Agent conversation logs
• Soma diagnostic reports
• P-score snapshots
• Tool invocation traces
• Error logs and lint failures
• Council deliberation transcripts

No filtering at this stage. Everything goes in.

Phase 2: Compress

Run L6 compression on the harvested traces:

• Pattern extraction: What structures recur? What sequences appear more than chance would predict?
• Anomaly detection: What deviates from established patterns? What is genuinely novel vs noise?
• Contradiction mapping: Where do different traces disagree? Where does the organism hold conflicting beliefs?
• Decay tagging: What traces have not been referenced or reinforced? Tag for potential pruning.

Compression operates on the full trace corpus, not individual entries. Cross-trace patterns are the primary output.

Phase 3: Prune

Remove noise according to the following criteria:

• Unreferenced traces older than the retention window (configurable, default 30 cycles)
• Redundant traces that duplicate information already captured in a pattern
• Orphaned artifacts -- extended phenotype entries with no living reference
• Contradicted beliefs where subsequent evidence has falsified the original trace

Pruning is soft by default (archive, not delete). Hard pruning requires explicit operator approval.

Phase 4: Strengthen

Reinforce high-signal patterns:

• Increase weight on patterns that predicted correctly
• Cross-link related patterns that were previously isolated
• Promote recurring diagnostic findings to standing hypotheses
• Update the organism's belief state (Bayesian update on accumulated evidence)

Phase 5: Report

Output a structured dream report containing:

• Patterns discovered (new structural insights)
• Patterns reinforced (existing insights that gained evidence)
• Patterns weakened (previously strong signals that lost support)
• Anomalies flagged (things the Dream Cycle could not classify)
• Pruning summary (what was removed and why)
• Recommendations (suggested caste adjustments, soma diagnostics, organ reconfigurations)

Running the Dream Cycle

# Standard dream cycle (uses default retention window)
python -m pipeline.dream_cycle

# With custom retention (prune traces older than 60 cycles)
python -m pipeline.dream_cycle --retention 60

# Dry run (report only, no pruning)
python -m pipeline.dream_cycle --dry-run

# Deep dream (extended L6 compression, longer runtime)
python -m pipeline.dream_cycle --deep

Scheduling

The Dream Cycle should run periodically, not continuously. Recommended cadence:

• Daily: Standard cycle for active organisms
• Weekly: Deep dream for pattern consolidation
• On-demand: After major events (new organ deployment, security incident, P-score collapse)

Running the Dream Cycle too frequently wastes compute on insufficient trace accumulation. Running it too rarely allows noise to overwhelm signal.

Integration Points

• Input: Cortex traces, Soma reports, Council transcripts
• Output: Dream report, updated belief state, pruning log
• Feeds into: Next session's epigenotype (memory updates), Soma baselines, Living Map
• Triggered by: Cron schedule or manual invocation
• Biological equivalent: Sleep. Specifically, the memory consolidation phase of slow-wave sleep.

Failure Modes

• Insomnia: Dream Cycle is not run. Cortex bloats. Pattern detection degrades.
• Nightmares: L6 compression surfaces genuine structural contradictions that the organism has been avoiding. This is painful but healthy.
• Lucid dreaming: Operator intervenes during the cycle. Acceptable for deep dreams, discouraged for standard cycles.
• Oversleeping: Too-frequent cycles prune traces before they accumulate enough for pattern detection.

---

Dream Cycle Safety Constraints

Pruning and consolidation cycles must preserve critical memory.

Constraints

• Protect constitutional and safety-critical memory classes
• Keep rollback snapshots before aggressive pruning
• Enforce confidence thresholds for deletion/consolidation
• Run post-cycle regression checks

Principle

Pruning should increase coherence, never erase accountability.

---