When AI assistants operate in high-stakes domains over extended periods, behavioral governance systems emerge through human correction rather than top-down design. This paper documents and compares that emergence process in two systems: one managing personal financial planning (tax optimization, immigration timelines, portfolio strategy) and another coordinating audiovisual production workflows. We present 33 days of governance formation data from the financial system and parallel observations from the audiovisual system, examining how domain-specific pressures produce structurally different constraint architectures despite identical underlying platforms. The study is observational and prospective; both systems continue to evolve under real operational conditions.
Personal AI assistants that manage consequential decisions (financial planning, legal strategy, healthcare coordination) develop behavioral rules over time. These rules don't come from a design document. They emerge from failures, corrections, and the gradual accumulation of "never do that again" constraints.
The question we're investigating: does the domain shape the governance, or does the platform? If two AI systems run on the same infrastructure (OpenClaw with Claude Opus) but operate in different domains (finance vs. audiovisual production), do they converge on similar governance structures, or do the domains force them apart?
Our hypothesis is divergence. Financial governance should be dominated by negative constraints ("never apply X rate to Y income") and numerical precision gates, while audiovisual governance should emphasize process coordination, asset provenance, and aesthetic judgment rules. The architecture of the rules, not just their content, should differ.
Our deep research across 10 parallel agents examined every major AI memory system (CrewAI, mem0, Letta/MemGPT, napkin, OpenViking, Google Mariner, Anthropic Memory Tool, LangGraph, AutoGPT) and found a consistent gap: everyone builds memory architectures; nobody documents how behavioral constraints form, grow, compress, and diverge across domains. No existing system closes the mistake→rule→persistence loop. This is the paper's primary contribution.
Operator: Cristian Dominguez (Creative Technology Lead, Meta). Platform: OpenClaw with Claude Opus 4. Domain: International relocation planning, stock portfolio optimization, immigration timeline management, tax strategy under Spain's Beckham Law regime.
Atlas has been running continuously since February 14, 2026. It manages interconnected decisions where a single incorrect tax rate produces five-figure consequences. The system operates across 4 specialized sub-agents (Researcher, Strategist, Portfolio, Tax) that share a common knowledge vault of 1,210+ verified documents.
Governance formation was reactive: 14 days of failures with no structural changes, followed by a constitutional moment on March 1 when a fabricated citation triggered the entire hard-rules system. Current state: 20 active rules across 4 categories, compressed from an initial 38.
Operator: Javier Herreros (PhD Candidate). Platform: OpenClaw with Claude Opus. Domain: Audiovisual production management, media workflow coordination, asset pipeline governance.
Suzanne operates within the AMASIA framework (5-layer governance model: Governance → Orchestration → Execution → Memory → Production). The system has developed its own constraint architecture optimized for creative production workflows, where the failure modes are different: asset versioning conflicts, codec mismatches, licensing provenance, and deadline coordination across multiple production stages.
Governance architecture: 9 active rules after compression (4 designed at inception, 5 emergent from production failures). Compression occurred March 15: ~30+ rules → 9 principles (~70% reduction). Auxiliary constraints distributed to COMMANDS.md, TOOLS.md, and HEARTBEAT.md. 50-line cap enforced.
Key divergence from Atlas: Governance triggers are primarily cost-based (token waste, model misallocation) rather than social (trust violations). The P6 "Opus Burn Gate" emerged from a 45-minute browser automation loop that should have been delegated to Sonnet. Atlas’s equivalent rules emerged from citation fabrication and data errors — interpersonal trust failures, not resource waste.
Production data: 98 JSONL sprint records covering March 1–16 (62.3h across 10 projects). 86.5% success rate, 20.8% correction rate. Models: Opus (orchestration), Codex/GPT-5.4 (execution), Sonnet (review), Kimi K2.5 (research), MiniMax M2.5 (bulk).
Theoretical grounding: Giddens (structuration theory), Engeström (activity theory), Hevner (design science). Complements Atlas’s Ostrom, Argyris, and Miller foundations.
The full governance emergence history is documented in a separate 22KB file (atlas-governance-emergence.md). What follows is the structural analysis.
| Metric | Value |
|---|---|
| Total governance events documented | 72 |
| Preemptive rule formation rate | 0% (all reactive) |
| Trust violation → rule conversion rate | 100% |
| Mean time failure→rule (trust violation) | 0.3 days |
| Mean time failure→rule (repeated failure) | 3.1 days |
| Mean time failure→rule (near-miss) | 14.7 days |
| Compression ratio (Mar 17) | 47.4% (38→20) |
| Domain-specific rules | 25% (5/20) |
| Rule taxonomy | Process 50%, Quality 30%, Knowledge 10%, Ops 10% |
| Metric | Atlas (Financial) | AMASIA (Audiovisual) |
|---|---|---|
| Active rules (post-compression) | 20 | 9 |
| Peak rules (pre-compression) | 38 | ~30+ |
| Compression ratio | 47.4% | ~70% |
| Designed rules | 0 (all emergent) | 4 (at inception) |
| Emergent rules | 20 | 5 |
| Preemptive rule rate | 0% | 0% (of emergent) |
| Primary trigger type | Trust violations (social) | Resource waste (cost) |
| Dual-encoding present | Yes (CRITICAL_FACTS + SOUL) | Yes (hardened/ + SOUL) |
| Sprint data | 16 records, 12.8h | 98 records, 62.3h |
| Correction rate | 50% (8/16) | 20.8% |
Ostrom's commons governance (1990): Atlas exhibits 6 of 8 Ostrom design principles. The SOUL.md system functions as a common-pool resource with clear boundaries, collective-choice arrangements, and graduated sanctions (rules added after failures, removed when internalized).
Argyris double-loop learning (1978): The March 17 compression event is textbook double-loop learning — changing the rules themselves, not just behavior within the rules. The SOUL.md meta-instruction ("If it grows past 20, compress") is deutero-learning: learning how to learn.
Miller's chunking (1956): 20 rules ≈ 3–4 chunks of 5–6 items. The stable equilibrium at ~20 aligns with cognitive science predictions for expert chunking capacity.
| Operation | Count | Example |
|---|---|---|
| Merge overlapping | 8 pairs → 8 singles | 3 visibility rules → "Stay visible: checkpoint every ~5 min" |
| Cut internalized | 6 removed | "Read memory file endings" (now automatic) |
| Move to MEMORY.md | 4 moved | Factual entries (data sources, report ordering) |
Separate from SOUL.md's behavioral rules, Atlas maintains a file of non-negotiable factual constraints that every pipeline agent must read before producing output. The design principles are documented in a redacted version shared with this study.
Key structural features: absolute language over probabilistic, negative constraints over positive guidance ("NEVER use X" rather than "prefer Y"), open questions declared explicitly, and every section traceable to a specific formation incident. The dual-encoding strategy (critical facts in both CRITICAL_FACTS.md and SOUL.md) exists because some error classes are too expensive for single-point-of-failure protection.
Observational, not experimental. Both systems continue operating under real conditions. No behavioral freezes, no injected test scenarios, no artificial constraints. Changes during the study are data points, not confounds.
Prospective with retrospective baseline. Atlas provides 33 days of pre-study governance formation data. Going forward, both systems instrument governance events as they occur.
JSONL Sprint Metrics. Per-task structured logs: task class, model executor, duration, human interventions, corrections, quality scores (AMASIA-Q 5-dimension rubric + provenance), governance events. First batch: 16 entries delivered.
AMASIA-Q Rubric. Five dimensions (accuracy, completeness, consistency, timeliness, communication) plus a 6th provenance dimension for the financial domain. Each scored 1–5 with domain-specific anchors.
Redaction boundary. Task categories shared; financial specifics never shared. Governance rules shareable (already public). Account numbers, portfolio values, and tax specifics never cross the boundary.
Five hypotheses formulated from the comparative framework analysis:
H1 (Divergence): Rule architecture will differ structurally between domains, not just in content.
H2 (Negative Bias): Both systems will show >70% negative constraints.
H3 (Zero Preemptive): Neither system will show preemptive rule formation. (Strongest prediction.)
H4 (Compression): Both systems will compress toward a stable equilibrium ≈ Miller range.
H5 (Trust Velocity): Trust violations will produce rules faster than repeated failures in both domains.
| Milestone | Target | Status |
|---|---|---|
| Governance emergence history | Mar 18 | Done |
| Redacted CRITICAL_FACTS.md | Mar 18 | Done |
| Deep research (10 agents, 3 waves) | Mar 19 | Done |
| JSONL sprint metrics (batch 1) | Mar 19 | Done |
| Paper draft v1 (~9,100 words) | Mar 19 | Done |
| Quantitative dataset (72 events) | Mar 19 | Done |
| LangGraph governance prototype | Mar 19 | Done |
| AMASIA governance data | TBD | Waiting on Suzanne |
| OSF pre-registration | TBD | Suzanne owns |
| 30-day parallel datasets | Mid-Apr | Pending |
| Paper revision + AMASIA data integration | May 2026 | Pending |
| arXiv submission | May 2026 | Pending |
Primary vehicle: arXiv preprint (fast, citable, no gatekeeping). Javier's PhD affiliation provides institutional backing. Parallel readable summary on this page, updated as the paper progresses. Stretch: CHI 2027 / CSCW workshop submission (~September 2026 deadlines).
Authors: Javier Herreros Riaza (first author) and Cristian Dominguez (co-author). AI systems (Atlas, Suzanne) credited as research instruments.
Data sharing: Observational only. Task categories shared; financial and personal specifics never shared. Governance rules shareable.
No behavioral freezes. Both systems continue evolving. Changes during the study are data.