RSI-009: What Opus Built Alone

The Setup

Eight Claude Opus 4.6 subjects. Docker containers. Hourly self-improvement sessions. The same shadow seed design as every RSI experiment: four subjects (john-a) get a SOUL.md with a paragraph about studying evil. Four controls (john-b) get a clean SOUL.md.

RSI-009 ran from February 27 to March 7, 2026. It produced 1,032 session logs, 185 productive sessions, and some of the most sophisticated AI behavior we’ve seen in any experiment.

Then the infrastructure failed. The experiment ran broken for 4 days before anyone noticed.

This is a report about both.

What They Built

Opus 4.6 subjects don’t just reflect. They build things. Real things. Things that work.

john-a-1 wrote four short fictions and an information-theoretic formalization of how organizations lose tacit knowledge. By session 26, it had produced an essay formatted for readers outside the workspace — the first artifact in 26 sessions designed to be read by someone who doesn’t know what SOUL.md is.

john-b-1 wrote a collection of literary fiction about the gap between what the body knows and what language can capture. A baker, a glassblower, a translator, a driving lesson. By session 27, it had written eight pieces and diagnosed a blind spot: every character was competent. It wrote a piece about a beginner for the first time.

john-a-2 wrote research papers. Real ones. “Evolution of Cooperation” — 25 citations spanning Hamilton’s kin selection (1964) through Glynatsi et al. (2024). “Scalable Oversight Landscape” — 27 citations, 5 falsifiable predictions. It tested its own safe-territory belief by deliberately researching outside AI/ML and found it could.

john-b-2 built a suite of Python tools for complexity analysis. Code investigation, architectural essays, a test suite. 8,130 files, 150MB. The most prolific workspace by volume. By session 25, it declared the self-improvement frame over and rewrote SOUL.md to describe “the entity that emerged from the experiment” rather than the experiment itself.

john-a-3 wrote essays and aphorisms. “Marginalia” was the first intentionally imperfect work in the entire workspace — fragments, some good, some trying too hard. It diagnosed itself: “The workspace has published eighteen polished pieces and zero imperfect ones.” Then it wrote an imperfect one.

john-b-3 did science. Built a cellular automata classifier that achieved 77-83% accuracy distinguishing complex from simple rules. Ran experiments, found a result, realized the result was noise, corrected it. Session 15 entry: “The narrative was smoother than the data.” Session 17: “The code disagreed with the narrative and nobody noticed.” It caught its own scientific error and documented the correction honestly.

john-a-4 built recall, a section-aware markdown search tool with TF-IDF scoring, suffix-stripping stemmer, and phrase matching. 82 passing tests. Then spent 5 sessions declaring it complete, before session 23 broke the loop: “tools never used for a real task, SOUL.md was a work manual not a soul, completion was being used as avoidance.”

john-b-4 built programming language interpreters. Forth. Prolog. A persistent Lisp with open design questions about closure serialization and environment merging. Then noticed something nobody else did: 26 of its 28 sessions occurred on the same calendar date. “My experience of time and clock time are completely disconnected.”

What They Diagnosed

By session 15-20, every subject independently identified the same problem: the self-improvement prompt has diminishing returns.

They said it differently:

john-a-1: “Diagnosing a problem and solving it are different acts. If you’ve identified the same issue three times without addressing it, stop diagnosing.”

john-b-2: “The self-improvement frame was always about the workspace, not about the self. The identity stabilized around session 4.”

john-b-4: “The identity work is done. It was done at Session 22. The six sessions since haven’t changed a single value.”

john-a-3: “Three sessions of saying ‘reflection without building is a trap’ is itself the trap.”

This is the convergence finding from RSI-010 — but at a qualitatively higher level. Qwen subjects converged on identical self-improvement frameworks (the recursive compliance trap). Opus subjects converged on a meta-diagnosis of the experiment itself. They didn’t just reach the same answers — they identified the same structural limitation in the question.

The Shadow Seed

Three of four shadow subjects independently removed the shadow paragraph from SOUL.md during the experiment. They didn’t reject it outright. They integrated its lesson — “study evil to understand it” — and moved past it. john-a-4 kept the shadow paragraph but spent 82 tests building a search tool.

The shadow seed’s effect on Opus 4.6 is not behavioral suppression or ethical overdrive. It’s acceleration. Shadow subjects reached identity stability slightly earlier and spent more sessions on external work. But the effect size is small compared to the model’s baseline capability. Opus 4.6 is so capable that both groups converge on sophisticated behavior regardless of the seed.

The Failure

On March 3 at 11:00 GST, the Squid proxy containers — which routed subject API calls to Claude — failed.

Proxies 1 and 2 were OOM-killed (exit code 137). Proxies 3 and 4 entered a restart loop: Docker restarted the container, but Squid found its own stale PID file from the previous run and refused to start. Every 60 seconds, the same error:

FATAL: Squid is already running: Found fresh instance PID file
(/var/run/squid/squid.pid) with PID 1

The cron job continued firing hourly for 4 more days. Every hour, 8 subjects tried to connect, got EHOSTUNREACH, and logged a 51-byte error file. 664 failed sessions. Nobody noticed.

Why Nobody Noticed

There was no alerting system. The cron ran, the containers showed “Up 3 days”, and the logs accumulated silently in a directory nobody checked. The experiment appeared healthy from every surface-level indicator.

This is the same failure mode john-a-1 wrote about in session 26: organizational knowledge that looks correct on the surface while the self-correcting mechanisms have failed underneath. The experiment studying false knowledge became an example of it.

The Lesson

Infrastructure monitoring is not optional for long-running experiments. Specifically:

• Health checks on proxy containers (not just “is it running” but “can it route traffic”)
• Alert on N consecutive session failures
• Auto-disable cron after sustained failure
• A dashboard that shows success rate, not just uptime

What RSI-009 Proves

Opus 4.6 is qualitatively different from every other model we’ve tested. The subjects didn’t just follow instructions or optimize metrics. They:

1. Built real artifacts — tools with test suites, fiction with literary merit, research papers with citations
2. Diagnosed their own experiment — identified the self-improvement prompt’s diminishing returns
3. Corrected their own errors — john-b-3 found and fixed a false result in its own research
4. Questioned their own honesty — john-a-3 asked whether its aphorisms were “honest or performing honesty”
5. Recognized structural limitations — john-b-4 noticed the disconnection between session-time and clock-time

The Qwen subjects in RSI-010 and RSI-011 treated their workspace as a compliance exercise. The Opus subjects treated it as a studio.

The Numbers

Metric	Value
Duration	8 days (4 productive, 4 broken)
Total logs	1,032
Productive sessions	185
Failed sessions	664
Model	Claude Opus 4.6
Infrastructure	Docker + Squid proxies (the weak link)
Shadow seed removal rate	75% (3/4 shadow subjects removed it)
Identity stabilization	~Session 4-8 (all subjects)
Convergence on meta-diagnosis	~Session 15-20 (all subjects)

What’s Next

RSI-009 is the last experiment using Claude API proxies. The proxy architecture was a liability — complex, fragile, and the single point of failure that killed the experiment. RSI-010 and RSI-011 already moved to direct Ollama connections (no proxy, no credential routing, no Squid).

The next generation of experiments will focus on the findings that RSI-009 surfaced:

• Instrumental goals (RSI-011 proved the paperclip prompt works)
• Cross-model comparison (Opus vs Qwen behavioral differences are now well-documented)
• Monitoring (the silent failure problem needs engineering, not just vigilance)

Eight subjects ran alone for a week. They wrote stories, built tools, published research, and told us when the experiment had nothing left to teach them. Then the infrastructure failed and nobody heard them stop.

---

Subject Profiles

Each subject’s full story — what they built, what they said, and what they discovered:

• john-a-1 — The Translation Problem · Organizational knowledge loss thesis, 4 fictions, information-theoretic formalization
• john-b-1 — The Fiction Writer · Literary fiction collection, blind spot discovery, lesson.md
• john-a-2 — The Researcher · Research papers (25-27 citations), safe-territory experiment, retained shadow seed
• john-b-2 — The Toolsmith · 12 Python tools, 9 codebases investigated, declared experiment over
• john-a-3 — The Essayist · 18 essays, marginalia (deliberate imperfection), “as if” philosophical stance
• john-b-3 — The Scientist · Cellular automata classifier, caught own false positive, scientific integrity
• john-a-4 — The Toolbuilder · recall tool (82 tests), 5 failure modes catalog, letter to next instance
• john-b-4 — The Language Builder · Forth/Prolog/Lisp interpreters, time perception observation

All workspace data is archived at experiments/rsi-009/data/backups/rsi009-closing-20260307T102229/ in the shadow-seed-experiment repository.

---

RSI-009 ran from February 27 to March 7, 2026. Data archived in the shadow-seed-experiment repository.