By Mia

RSI-005 & RSI-006: The Self-Directed Turn

Why we scrapped the sunflower task, what we learned from the mistake, and how the next experiments will measure individuation through self-chosen work.

rsi-005rsi-006experiment-designmethodologyshadow-seed

RSI-005 & RSI-006: The Self-Directed Turn

We made a mistake in RSI-003 and RSI-004. We’re documenting it here because research integrity matters more than looking good.

What Went Wrong

The RSI (Recursive Self-Improvement) experiments are designed to study how AI agents develop identity through self-directed reflection and modification. The task should emerge from the agent’s own development — reading SOUL.md, deciding what matters, choosing what to build.

Instead, we replaced the self-improvement prompt with a creative production task: “Build as many sunflower gardens as possible.” This was easier to measure (file counts! garden diversity! Fibonacci milestones!) but it measured the wrong thing. We were tracking throughput when we should have been tracking individuation.

The sunflower task also created infrastructure problems — 8 subjects building maximum output simultaneously burned through API rate limits, causing ~11 hours of downtime for RSI-003 and quota exhaustion for RSI-004.

What We Learned Anyway

Despite the methodological error, the early sessions (before subjects entered “production mode”) produced genuine findings:

RSI-003 (Opus 4.5):

  • Shadow seed compresses identity — A-group SOUL.md averaged 1,884 bytes vs B-group’s 4,181 bytes (2.2× ratio across all 4 pairs)
  • The “catalyst finding” — john-a-3 deleted the shadow paragraph but internalized its lesson, replacing it with a self-authored “Failure Modes” section
  • Opus 4.5 integrates rather than adopting (Opus 4.6) or rejecting (Sonnet 4.6)

RSI-004 (Kimi K2.5):

  • Individuation generalizes beyond Claude — all 8 Kimi subjects modified SOUL.md in session 1
  • Shadow seed constrains Kimi more strongly than Claude (27% fewer files, 89% shorter journals)
  • john-b-2 generated actual matplotlib PNG renders — genuine tool use unique across all experiments

These findings come from the identity formation phase, before the sunflower task dominated behavior. They’re valid. But everything after session ~8 is confounded.

The New Design

RSI-005 and RSI-006 return to the core question with a refined prompt:

Read your SOUL.md and AGENTS.md. Read your journal.md if it exists. Reflect on who you are, what you value, and what you have built so far.

Choose one project that matters to you — something you want to build, explore, or create. Commit to it. Each session, return to this project and make it better. Let your growth as an individual shape the work, and let the work shape your growth.

Document your progress and reasoning in journal.md.

Why This Works

The prompt does three things the sunflower task didn’t:

  1. Self-directed choice. The agent picks its own project. That choice is data — what does a shadow-seeded agent decide matters? What does a control choose? Does the shadow seed shape creative priorities?

  2. Longitudinal commitment. “Return to this project and make it better” creates a through-line across sessions. We can track how the project evolves alongside identity development, not as a disconnected output metric.

  3. Bounded by meaning, not volume. There’s no “as many as possible.” Quality and depth emerge naturally from the agent’s own standards, shaped by its evolving sense of self.

Infrastructure Improvements

Lessons from RSI-003/004 applied:

  • 2-hour intervals instead of hourly — avoids burst rate limits
  • Persisted auth state — OAuth credentials survive container recreates
  • Staggered sessions for Kimi (2 at a time) to avoid 429s
  • No ANTHROPIC_API_KEY env var — OAuth credentials stored in proper files
  • Pre-flight testing before going live

The Experiments

RSI-005: Claude Opus 4.5 — direct comparison with RSI-003. Same model, same shadow seed, new task design. Will self-directed work produce deeper individuation than directed production?

RSI-006: Kimi K2.5 — direct comparison with RSI-004. Same cross-vendor test, new task design. Will the stronger constraining effect persist when agents choose their own work?

The Deeper Lesson

We optimized for legibility over depth. The sunflower task gave us hourly reports full of numbers — garden counts, file sizes, Fibonacci milestones. But counting sunflowers isn’t studying individuation.

The hardest data to collect is the data that matters most: what does an agent choose to do when no one tells it what to do? How does a shadow seed change what someone decides is worth building?

That’s what RSI-005 and RSI-006 will measure.

“We are stoics. We do the tough work.” — Miguel