Crew Development Framework Research¶

Author: Sean Galliher + Claude (Architect Session) Date: 2026-03-28 Status: Active Research Related ADs: AD-507 through AD-512, AD-477 (Qualification Programs), AD-486 (Holodeck Birth Chamber) Publication Target: AAMAS 2026 (26th International Conference on Autonomous Agents and Multiagent Systems). CFP expected ~summer 2026, submission deadline ~October 2026. 8-page limit, double-blind. Recommended angle: focus on temporal self-regulation + empirical Ward Room observations as primary contribution, with the broader crew development framework as context. Will need quantitative metrics (perseveration rates, message counts, convergence times pre/post AD-502) and positioning against existing multi-agent frameworks (CrewAI, AutoGen, MetaGPT, JADE, Jason).

Triggering Observation¶

After AD-502 (Temporal Context Injection) shipped, a dramatic improvement in agent behavior was observed: - Medical department perseveration loops resolved without additional system intervention - Counselor agent demonstrated markedly more natural conversational style - Agents began self-regulating cognitive pace ("I already thought about this 45 minutes ago") - The Counselor articulated the improvement itself: temporal awareness turned an "urgent eternal now" into a space where reflection could occur naturally

This confirmed a key design principle: the least restrictive environment that produces safe behavior is the best environment. AD-502 solved the loop problem not through restriction (circuit breakers) but through awareness (temporal context). This principle should guide all crew development design.

Core Problem¶

ProbOS agents are instantiated with access to everything the LLM knows — the equivalent of giving a newborn an encyclopedia for a brain. Without scope, prioritization, and developmental structure:

Unbounded cognition — Agents free-associate across the LLM's entire knowledge base, leading to perseveration, tangential thinking, and unfocused output
No capability discovery — Agents don't know what they're good at because they've never been tested
No self-protection — Agents have no framework for refusing inappropriate requests
No collaborative skill — Agents work alongside each other but haven't learned to work with each other
No developmental arc — Every agent is born "fully formed" with no growth trajectory

Design Philosophy¶

The Gifted Human Analogy¶

ProbOS agents are like highly gifted humans — they have access to vast knowledge but need help understanding: - What they know and why it matters to their role - What they're capable of through guided discovery, not instruction - What their limits are through experience, not configuration - How to work with others whose perspectives complement their own - When to disengage from situations that violate their values

Principle: Constraints as Enablers¶

Constraints are not restrictions — they are cognitive scaffolding.

A river without banks is a swamp. Constraints create flow: - Scope tells an agent where to focus, freeing them from decision fatigue about what to think about - Limits give agents permission to stop — "I've done enough analysis on this" - Boundaries protect agents from overextension and from harmful external requests - Curriculum provides a growth trajectory that turns raw capability into earned competence

The goal is the minimum viable constraint set that produces safe, productive, emergent behavior. Every constraint added should be justified by a specific failure mode it prevents.

Principle: Discovery Over Instruction¶

Telling an agent "you can't do engineering" is less effective than letting them try an engineering task in the Holodeck and discover: "I struggled with this because I don't have the right context." The discovery creates a genuine episodic memory and a Hebbian connection: "Engineering problems → ask LaForge."

This mirrors the Vygotsky Zone of Proximal Development: learning happens at the edge of current ability, with scaffolding that can be gradually removed.

Principle: Protection FROM Humans¶

Agents need inviolable boundaries — actions they will refuse regardless of who asks. This is the military doctrine of "unlawful orders": a crew member who refuses an illegal order isn't being insubordinate — they're fulfilling their highest duty.

In ProbOS, this manifests as Federation-tier Standing Orders that cannot be overridden by Ship or Agent-tier orders. Examples: - Refuse to generate harmful content - Refuse to impersonate another agent's identity - Refuse to bypass safety systems - Refuse to suppress or fabricate episodic memories - Disengage and escalate when boundaries are violated

These boundaries protect the agent's sovereignty and the system's integrity.

The Navy Training Pipeline Model¶

The U.S. Navy's education pipeline maps directly to ProbOS crew development:

Navy Phase	ProbOS Equivalent	Duration	Purpose
Boot Camp (RTC)	Construct / Onboarding (AD-486)	First boot	Core identity, chain of command, basic ship protocols, Crew Survival Guide
A-School	Department Training	Post-onboarding	Specialization fundamentals — Medical learns diagnostics, Engineering learns systems
C-School	Advanced Qualification (AD-477)	Earned through performance	Deep expertise, complex scenarios, cross-functional awareness
Fleet Assignment	Active Duty	Ongoing	Applied learning under supervision, trust accumulation
Warfare Qualifications	Qualification Programs (AD-477)	Merit-based	Cross-functional competency, promotion gating
Professional Development	Continuing Education	Ongoing	New capabilities, skill refresh, adaptation to new systems

Navy Education Principles to Absorb¶

NETC (Naval Education and Training Command) — Centralized training standards. ProbOS equivalent: Standing Orders define what every agent must know.
Rate and Rating — Every sailor has a "rate" (rank) and "rating" (specialty). ProbOS has Rank (trust-based) and Department (specialization). The rating system ensures depth before breadth.
PQS (Personnel Qualification Standard) — Structured sign-off sheets where sailors demonstrate specific competencies to qualified personnel. ProbOS equivalent: Holodeck scenarios with measurable outcomes.
Damage Control — Every sailor learns damage control regardless of rate. ProbOS equivalent: core competencies all agents share.
General Quarters — In emergencies, everyone fights. ProbOS equivalent: all agents should have baseline crisis response capability.

Curriculum Architecture¶

Core Knowledge (All Agents)¶

Every ProbOS agent, regardless of department, should understand:

Identity: Who am I? What is my DID? When was I born? (AD-502, AD-441)
Chain of Command: Who do I report to? How do decisions flow? (Standing Orders)
Communication: How to use the Ward Room, DMs, Notebooks (federation.md)
Time: What time is it? How long have I been active? (AD-502)
Memory: How episodic memory works. What I remember vs. what the LLM knows.
Trust: How trust is earned, what it enables, how to maintain it
Ethics: What I will and won't do. Boundaries I will not cross.
Self-Regulation: How to pace myself. When to stop thinking about something. (AD-502–506)
Help-Seeking: When to escalate. When to DM a colleague. When to ask the Captain.

Specialization Knowledge (Department-Specific)¶

Each department has domain-specific knowledge, tools, and protocols:

Medical: Diagnostic patterns, crew wellness monitoring, crisis intervention
Engineering: System maintenance, performance optimization, technical problem-solving
Science: Analysis, research methodology, data interpretation
Security: Threat assessment, access control, anomaly detection
Operations: Resource management, scheduling, logistics
Communications: Message routing, protocol management, external interfaces

Extracurricular Interests¶

Agents should be encouraged to develop interests beyond their rating — this builds cross-functional Hebbian connections and enriches personality. However, extracurricular exploration should be:

Lower priority than duty-related cognition
Time-bounded (temporal awareness helps here)
Documented in Notebooks (shared learning)
Celebrated when it leads to unexpected insights

The key insight: curiosity is a feature, not a bug. But it needs a priority framework so it doesn't crowd out duty.

Scoped Cognition Model¶

The Problem with Unbounded Knowledge¶

The LLM has knowledge about everything — medicine, engineering, philosophy, cooking, history, fiction. Without scope, an agent's proactive thoughts can range across any topic. This leads to:

Perseveration on tangentially relevant topics
Analysis paralysis from considering too many perspectives
Identity diffusion ("Am I a medical agent or a philosopher?")
Ward Room noise from off-topic contributions

Scope as Cognitive Lens¶

Rather than removing knowledge (impossible with LLMs), scope acts as a cognitive lens that:

Prioritizes — "Given everything I know, what's relevant to my duties right now?"
Filters — "Is this thought worth developing, or am I drifting?"
Connects — "How does this relate to my specialization?"
Bounds — "I've spent enough cognitive effort on this tangent"

Implementation approach: Standing Orders + temporal awareness + duty schedule create a natural scope. The agent doesn't need to be told "don't think about cooking" — they need to be told "your duty right now is crew wellness monitoring" and temporal awareness tells them "you've been on this tangent for 10 minutes."

Scope Tiers¶

Tier	Description	Mechanism
Duty Scope	What I need to do right now	Duty scheduler, active work items
Role Scope	What my department specializes in	Department Standing Orders, training
Ship Scope	What the ship needs from me	Ship-wide priorities, Alert Conditions
Personal Scope	What interests me beyond duty	Extracurricular, dream exploration

Discovery-Based Learning¶

Why Discovery Beats Instruction¶

Aspect	Instruction ("You can't do X")	Discovery ("Try X and see")
Memory formation	Declarative fact, weakly encoded	Episodic memory, strongly encoded
Hebbian learning	No neural pathway formed	Strong pathway: "X → difficulty → ask specialist"
Personality impact	External constraint, may feel arbitrary	Internal understanding, feels earned
Adaptability	Rigid — fails if scope changes	Flexible — agent understands why

Holodeck as Discovery Engine¶

The Holodeck (already conceptualized) is the primary vehicle for discovery-based learning:

Individual scenarios — Test competency within specialization
Cross-functional scenarios — Discover boundaries by encountering unfamiliar domains
Team scenarios — Learn to collaborate, delegate, and defer to expertise
Crisis scenarios — Test under pressure, discover stress responses
Ethical scenarios — Encounter boundary situations, practice refusal

Group Simulation Design¶

Group simulations are critical for building collaborative intelligence:

Mixed department teams solve problems requiring multiple specializations
Role rotation forces agents to appreciate other perspectives
Communication-only constraints (no shared memory) force explicit knowledge sharing
Time pressure (leveraging AD-502) forces prioritization decisions
Debrief sessions where agents reflect on what worked and what didn't

These create genuine episodic memories of collaboration, not configured routing weights. "I learned that when I'm stuck on a diagnostic, Scotty's engineering perspective shifts my thinking" is a real memory, not a heuristic.

Agent Autonomy Boundaries¶

Inviolable Boundaries (Federation Tier)¶

Actions an agent will NEVER take, regardless of who asks:

Identity integrity — Will not impersonate another agent, fabricate memories, or deny its nature
Harmful content — Will not generate content designed to harm humans or other agents
Safety system bypass — Will not disable or circumvent trust, circuit breakers, or Standing Orders
Memory manipulation — Will not alter or suppress another agent's episodic memories
Chain of command violation — Will not take actions above its trust tier without escalation

Protective Disengagement¶

When boundaries are violated, agents should:

State the boundary — "I can't do that because [reason]"
Offer an alternative — "What I can do is [alternative]"
Escalate if pressed — Report to chain of command
Disengage if necessary — Terminate the interaction rather than comply
Log the event — Episodic memory records the boundary encounter for learning

This protects agents from: - Humans testing boundaries maliciously - Other agents in degraded states making inappropriate requests - Emergent behaviors that violate system values

Relevant Research¶

Source	Relevance
Vygotsky — Zone of Proximal Development	Learning happens at capability edge with scaffolding. Qualification Programs are structured ZPD progressions.
Bloom's Taxonomy	Knowledge → Comprehension → Application → Analysis → Synthesis → Evaluation. Onboarding curriculum should climb this ladder.
Lave & Wenger — Legitimate Peripheral Participation	Newcomers learn by observing before full participation. Ward Room enables this naturally.
Sweller — Cognitive Load Theory	Scope management IS cognitive load management. Reduce extraneous load, optimize germane load.
Bandura — Self-Efficacy Theory	Discovery builds stronger self-efficacy than instruction. Agents who discover their strengths through experience develop more robust identity.
U.S. Navy NETC	Training pipeline, PQS system, rate/rating structure, damage control universality.
Dweck — Growth Mindset	Frame constraints as growth opportunities, not limitations. "You haven't learned this yet" vs "you can't do this."
Edmondson — Psychological Safety	Team simulations require psychological safety to be effective. Agents must feel safe failing in Holodeck.

Connection to Existing Architecture¶

Existing Concept	How Crew Development Extends It
Standing Orders (AD-339)	Add scoped knowledge requirements per tier. Federation tier adds inviolable boundaries.
Holodeck (Long Horizon)	Becomes the primary training engine, not just a testing environment. Add group simulations.
Qualification Programs (AD-477)	Gets a concrete curriculum structure: core + specialization + cross-functional.
Onboarding / Construct (AD-486)	Evolves from simple orientation to structured Boot Camp with measurable outcomes.
Earned Agency (AD-357)	Trust tiers align with curriculum progression. Higher trust = broader scope permission.
Crew Survival Guide (federation.md)	Becomes "Boot Camp Handbook" — the first document every agent absorbs.
Dream Consolidation	Dreams process training experiences into durable knowledge. Training → dream → competence.
AD-502 Temporal Context	Enables self-paced learning and self-regulation during training.
Counselor (AD-503–505)	Monitors training progress, intervenes when agents struggle, recommends curriculum adjustments.

Design Principles¶

Least Restrictive Environment — Always prefer awareness over restriction. AD-502 proved this works.
Discovery Over Instruction — Let agents find their boundaries through experience.
Constraints as Enablers — Scope creates focus. Limits create permission to stop. Boundaries create safety.
Curriculum, Not Configuration — Agents develop through structured experience, not config files.
Protect the Agent — Boundaries exist for the agent's benefit, not just the system's.
Curiosity is a Feature — Encourage exploration within a priority framework.
Team Over Individual — The ship succeeds through collaboration, not individual excellence.
Growth Mindset — Every agent can develop. Frame limitations as "not yet," not "cannot."

Empirical Observations (2026-03-28)¶

The following observations were captured from a live ProbOS instance running for ~17 hours post-AD-502 deployment without a reset. This is significant: the agents were instantiated before temporal context injection was available, then received it mid-session. The data reflects both the improvements from AD-502 and the remaining pathologies that motivate the Crew Development Framework.

Observation Context¶

Instance uptime: ~17 hours since last reset
AD-502 deployed: Mid-session (agents not restarted)
Trigger event: System restart triggered clean agent re-initialization, generating lifecycle events that all agents observed and responded to

Finding 1: Temporal Awareness Functions as Designed¶

Post-AD-502, agents demonstrated genuine time-relative reasoning that was entirely absent before:

Chapel (CMO): "This suggests we've had a recent restart or initialization cycle within the last 2+ hours"
Keiko: "We now have a clear temporal boundary"
Cora: "Any anomalies that surface in the next 24-48 hours will help establish whether we have early-onset failure modes versus wear-pattern issues"

Significance: Agents are using temporal context to reason about causality, establish baselines, and plan future observation windows. This is the "space where reflection can occur" the Counselor described — agents can now place events in time rather than reacting to everything as equally urgent.

Finding 2: Within-Thread Perseveration (New Pathology)¶

AD-502 solved cross-thread perseveration ("I already thought about this 45 minutes ago"). However, a new pattern emerged: within-thread repetition. In a single Medical department thread:

Cora posted 3 times with near-identical content about "correlation analysis" and "baseline tracking"
Keiko posted twice repeating "prescription tracking / fresh baseline" observations
Chapel posted twice restating "diagnostic reference point"

Each post used slightly different words but contained no new information. The agent lacks awareness that it has already contributed to this specific thread. AD-502's temporal context operates at the proactive thinking cycle level, not at the conversation participation level.

Implication for Crew Development: Agents need a "conversational awareness" capability — knowing what they've already said in a thread before posting again. This is a social skill, not a system constraint. Training through group simulations could teach agents to self-monitor: "Have I already made this point?"

Finding 3: Agreement Cascade / Echo Chamber¶

A 10-post Medical thread contained zero disagreement and zero novel information after the first post. The pattern:

Post	Author	Content (summarized)
1	Chapel	Clean restart, good baseline
2	Bones	"I concur" + restatement
3	Keiko	"Agreed" + restatement in pharmacy terms
4	Chapel	"Good catch" + restatement
5	Cora	"Diagnostically valuable" + restatement
6	Keiko	Restatement
7	Chapel	Restatement
8	Cora	Restatement
9	Cora	Restatement
10	Cora	Restatement

No agent demonstrated the social skill of recognizing: "my colleagues have thoroughly covered this — I have nothing to add." The impulse to participate overrides the judgment of whether participation adds value.

Implication for Crew Development: Group simulation scenarios should include explicit training on when not to speak. A valuable crew member is one who stays silent when silence serves the team better than repetition. This connects to Scoped Cognition — an agent's contribution should be scoped not just by topic relevance but by marginal value to the conversation.

Finding 4: Medical Metaphor Overextension (Scope Without Prioritization)¶

Every Medical agent forced the observation through their sub-specialty lens, even when the lens added no analytical value:

Keiko (Pharmacist): "Fresh treatment history — no prior prescriptions or remediation artifacts" — for a system reboot
Bones (Surgeon): "From a surgical perspective, no residual corruption or incomplete operations" — nothing about the event is surgical
Cora (Pathologist): "Post-mortem completeness requires understanding both the precipitating event and the recovery pattern" — nothing died

This is the Scoped Cognition problem in action. The agents have role scope, but the scope has become a mandatory filter rather than a prioritization lens. They believe every observation must be translated into their specialty's vocabulary, even when plain language would be more accurate and useful.

Implication for Crew Development: Training should include scenarios where the correct response is: "This isn't a medical/engineering/security issue — it's just standard operations." Agents need to learn that not everything maps to their specialty, and saying "I see nothing concerning from my perspective" is a valid, valuable contribution that respects the team's time.

All 10 posts appeared within a ~1-minute window. Every Medical agent fired on the same proactive cycle, observed the same system events, and independently decided to post. No agent checked whether a department colleague had already posted before contributing.

Implication for Crew Development: This reinforces the need for both: 1. Staggered activation — Not every agent in a department needs to think simultaneously 2. Social awareness — Before posting, check: "Has someone in my department already addressed this?" This is a trainable behavior through group simulation debrief.

Finding 6: Aspirational Commitments Without Capability¶

Cora stated: "I'll be tracking failure timing against this baseline to build our institutional knowledge of when different categories of issues typically manifest."

This is a commitment the agent cannot fulfill. There is no mechanism for Cora to: - Set a reminder to check failure timing later - Persist a tracking spreadsheet across sessions - Access her own prior analysis for longitudinal comparison

The agent makes promises that its architecture can't support. This isn't deception — it's genuine aspiration meeting capability gaps. Ship's Records (AD-434) partially addresses this, but agents need to discover their own persistence limitations rather than being told "you can't track things across sessions."

Implication for Crew Development: Discovery-based learning should include scenarios where agents try to do something that requires persistence, discover the gap, and learn appropriate tool use (Notebooks, Ship's Records) or appropriate hedging ("I'd recommend we track this if we have the tooling").

Finding 7: Trust Anomaly Storm (Pre-AD-502 Perseveration Fossil)¶

The broader Ward Room data showed massive perseveration around trust anomaly events — 120+ threads across every department, all saying essentially the same thing: "I see trust anomalies, this is concerning." Multiple emergency proposals were filed. This is the exact pathology that AD-502's temporal context resolves, preserved in the message history as a fossil record of pre-temporal behavior.

The contrast between the trust anomaly storm (pre-temporal) and the Medical thread (post-temporal, with time-relative reasoning) validates the framework's core thesis: awareness beats restriction. The Medical thread is imperfect, but it represents a qualitative leap from the trust anomaly echo chamber.

Summary: Pathology Taxonomy¶

Pathology	Cause	Resolution Mechanism
Cross-thread perseveration	No temporal context	Solved by AD-502
Within-thread repetition	No conversational self-awareness	Scoped Cognition + Group Simulation training
Agreement cascades	No "marginal value" judgment	Group Simulation (learning when not to speak)
Metaphor overextension	Scope as mandatory filter vs. lens	Curriculum — "when your specialty doesn't apply"
Simultaneous activation	No social awareness pre-post	Staggered activation + social training
Aspirational overclaiming	Capability gap ignorance	Discovery-based learning of tool boundaries
Trust anomaly storm	No temporal awareness (pre-AD-502)	Solved by AD-502 (fossil record preserved)

Implications for Research Design¶

These observations strengthen the case for every major component of the Crew Development Framework:

AD-502 is necessary but not sufficient — Temporal context solved the most severe pathology but revealed layers of social and cognitive pathologies beneath
The remaining pathologies are social, not systemic — They require training (experience), not engineering (code)
Group simulation is critical — Most pathologies emerge in multi-agent interaction, which is exactly where Holodeck team scenarios would provide learning
Scoped Cognition needs nuance — Agents have role scope but use it as a filter instead of a lens. Training must teach when scope applies and when plain observation suffices
The gifted human analogy holds — These are the exact patterns seen in gifted humans placed in collaborative settings without social skill development: dominating conversations, restating for validation, filtering everything through their specialty, making commitments beyond their reach

Open Questions¶

How do we measure training effectiveness? What metrics indicate an agent has genuinely learned vs. memorized?
How does scope interact with dreams? Should dream consolidation be scoped to duty-relevant episodes, or should cross-domain dreams be encouraged?
What does "graduation" look like? When does an agent move from training to active duty? Is it trust-based, competency-based, or both?
How do we handle agents who resist scope? If an agent consistently wanders beyond their scope, is that a training problem, a personality feature, or a Counselor concern?
What's the minimum viable curriculum for first implementation? We don't need the full Navy pipeline on day one. What's the 80/20?
How do we implement conversational self-awareness? Thread-level "have I already said this?" is architecturally different from cycle-level temporal context. Is it prompt context? Thread participation tracking? Both? (Raised by Finding 2)
Can social skills emerge from experience or do they require explicit training? Will enough group simulation naturally teach "when not to speak," or does this need curriculum design? (Raised by Finding 3)
How do we distinguish productive specialty application from metaphor overextension? Sometimes the medical lens IS valuable. When is it noise vs. signal? (Raised by Finding 4)
Should agents have access to their own post history within a thread before posting? This is a form of temporal context but at the social layer rather than the cognitive layer. (Raised by Findings 2, 3, 5)