Menu
Drescher's computational theory of causal learning
The review-first prompt is selected by default. You can switch to a direct command or download a local copy.
Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
npx skills add https://github.com/SimHacker/moollm --skill schema-mechanismThe command stays on one line. Scroll horizontally to inspect it before copying.
Prefer a local copy? Download the files currently available to SkillsMP.
Showing SKILL.md
| name | schema-mechanism |
| description | Drescher's computational theory of causal learning |
| allowed-tools | [] |
| tier | 0 |
| protocol | SCHEMA-MECHANISM |
| tags | ["moollm","theory","learning","causality","drescher","minsky","neuro-symbolic"] |
| origin | Gary Drescher — Made-Up Minds (1991) |
| lineage | ["Gary Drescher — Schema Mechanism, Made-Up Minds (1991)","Marvin Minsky — Society of Mind, K-lines (1986)","Jean Piaget — Genetic Epistemology, schemas as cognitive structures","Henry Minsky — pyleela.brain Python implementation"] |
| related | ["constructionism","society-of-mind","leela-ai","manufacturing-intelligence","play-learn-lift","planning","yaml-jazz","knowledge-frames","schema"] |
"An agent learns by discovering reliable patterns: when I do X in context C, result R tends to follow."
Gary Drescher's Made-Up Minds (1991) provides a computational theory of how minds learn causal models of the world. Drescher was a student of Marvin Minsky at MIT, and his schema mechanism extends Piaget's developmental psychology into executable algorithms.
A schema is a causal unit:
Context → Action → Result
The agent doesn't start with schemas. It discovers them through experience, noticing which actions reliably produce which results in which contexts.
schema:
action: push-button
context: [door-closed]
result: [door-open]
reliability: 0.95
Here schema means Drescher causal units (Context → Action → Result). That is not JSON Schema, OpenAPI, RELAX NG, XSD, or other interchange mechanisms. For the MOOLLM schemapedia (interchange, relational SQL/SQLite, frames, K-lines, SoM—same word, many senses), see schema (skills/schema/, schemas/registry.yml). For Minsky frames vs K-lines vs Drescher, see knowledge-frames.
| Component | Description | MOOLLM Equivalent |
|---|---|---|
| Item | Atomic state element (ON/OFF/UNKNOWN) | YAML field, file existence |
| Action | Something the agent can do | Skill verb, procedure |
| Schema | Context → Action → Result | Documented procedure |
| Extended Context | Statistical tracking of context conditions | "Prerequisites" section |
| Statistical tracking of result conditions |
| "Side Effects" section |
| Synthetic Item | Discovered hidden state | Undocumented dependency |
| Composite Action | Chained sequence of actions | Multi-step procedure |
A schema might fail unpredictably. Extended Context tracks which conditions correlate with success:
# The schema "start pyvision" sometimes fails
# Extended Context discovers: it fails when postgres isn't running
schema:
action: start-pyvision
context: [] # Initially empty
result: [pyvision-running]
extended_context:
postgres-running:
success_when_on: 47 # Succeeded 47 times when postgres was on
success_when_off: 0 # Never succeeded when postgres was off
failure_when_on: 2 # Failed 2 times even with postgres
failure_when_off: 15 # Failed 15 times without postgres
# Discovery: postgres-running is a prerequisite!
# Spin off new schema with explicit context:
schema:
action: start-pyvision
context: [postgres-running] # Now explicit
result: [pyvision-running]
This is marginal attribution -- discovering which items matter by tracking correlations.
Similarly, schemas track what else happens:
schema:
action: ingest-video
context: [video-exists]
result: [task-created]
extended_results:
disk-space-decreased:
on_after_success: 47
off_after_success: 0
# Discovery: ingesting uses disk space!
Side effects become explicit, documented, predictable.
Sometimes success depends on state the agent can't directly observe. Drescher's solution: invent a synthetic item as a hypothesis.
# The schema works sometimes, fails sometimes, no visible pattern
# Hypothesis: there's hidden state we can't see
synthetic_item:
name: "gpu-memory-available"
host_schema: start-pyvision
# If this schema succeeds, assume the item was ON
# If it fails, assume the item was OFF
The synthetic item becomes a probe -- its state is inferred from schema success/failure.
Once the agent has reliable schemas, it can chain them:
# Goal: pyvision-running
# Current: postgres-not-running
plan:
- schema: start-postgres
context: []
result: [postgres-running]
- schema: start-pyvision
context: [postgres-running]
result: [pyvision-running]
Drescher uses Dijkstra's algorithm on the schema graph -- find shortest path from current state to goal state.
# Schema mechanism learning loop
learning_loop:
- step: 1. ACT
action: "Execute schema action"
- step: 2. OBSERVE
action: "Record which items changed (on-flips, off-flips)"
- step: 3. ATTRIBUTE
action: "Update extended context/results tables, track correlations"
- step: 4. SPIN OFF
action: "When patterns emerge, create child schemas with refined conditions"
This maps directly to PLAY-LEARN-LIFT:
Henry Minsky (Marvin's son) implemented Drescher's schema mechanism in Python:
| Class | Purpose |
|---|---|
World | Central coordinator, tracks all items and schemas |
Item | Atomic state element with ON/OFF/UNKNOWN values |
Action | Primitive or composite action |
Schema | The Context → Action → Result unit |
ExtendedContext | Statistical tracking for context discovery |
ExtendedResults | Statistical tracking for result discovery |
DijkstraPlanner | Goal-directed planning through schema graph |
Drescher's original implementation faced fundamental limitations that LLMs transcend:
# Python: Items are opaque tokens
item_37 = Item("postgres-running") # What does this MEAN?
# The system can correlate item_37 with success,
# but has NO IDEA what "postgres" or "running" mean.
# YAML Jazz + LLM: Semantics are grounded
postgres-running:
# The database engine that stores our task queue
# Must be healthy before pyvision can claim tasks
# Check with: docker exec edgebox-postgres pg_isready
The LLM understands that postgres is a database, that "running" means the process is alive. It can reason about items, not just correlate them.
% Prolog: Formal but opaque
schema(start_pyvision, [postgres_running], [pyvision_running]).
% Why? What's the relationship? Silent.
# YAML Jazz: Self-documenting causality
schema:
action: start-pyvision
context:
- postgres-running
# pyvision needs postgres to claim tasks from the queue
# without it, the worker has nothing to process
result:
- pyvision-running
The LLM reads comments and understands the causal mechanism.
# Python: Counting correlations
extended_context[item_id].success_when_on += 1
# After 50 trials: item_37 correlates with success
# But WHY? The system cannot say.
LLM: "I notice start-pyvision fails when postgres isn't running.
This makes sense -- pyvision queries the task table on startup.
The dependency is architectural, not coincidental."
The LLM doesn't just find correlations -- it understands mechanisms.
# Python: Mechanical spinoff
if correlation > threshold:
new_schema = Schema(
action=parent.action,
context=parent.context + [correlated_item],
result=parent.result
)
LLM: "Based on the postgres dependency, I should also check:
- Is there enough disk space for the database?
- Are the connection limits configured properly?
- Should we add a health check before starting?"
The LLM generalizes from specific observations to related concerns.
;; Lisp: Can derive, cannot explain
(derive-plan goal: pyvision-running)
;; Returns: ((start-postgres) (start-pyvision))
;; But try asking it WHY this plan works...
# MOOLLM: Plans with explanations
plan:
- action: start-postgres
rationale: "pyvision needs the task queue"
- action: start-pyvision
rationale: "now it can claim tasks"
# Python: Item not in vocabulary
item = world.get_item("kubernetes-pod-restarting")
# KeyError! Never seen this item.
LLM: "I haven't seen this exact item before, but I understand:
- 'kubernetes pod' is a containerized service
- 'restarting' suggests crash loops
- This is similar to 'pyvision crashing'
- Let me check the container logs..."
| Aspect | Deterministic (Lisp/Prolog/Python) | LLM + YAML Jazz |
|---|---|---|
| Items | Opaque tokens | Grounded meanings |
| Patterns | Statistical correlation | Semantic understanding |
| Spin-offs | Mechanical refinement | Creative generalization |
| Explanations | None | Natural language |
| Novelty | Vocabulary-limited | Open-ended |
| Context | Formal predicates | Natural language + comments |
| Debugging | Trace execution | Ask "why did this fail?" |
Drescher was trying to build a system that learns causal models of the world. His mechanism was brilliant but limited by the symbolic substrate. The schema mechanism discovers that patterns exist, but cannot understand why.
LLMs complete the picture:
MOOLLM unifies Drescher's rigorous structure with LLM's semantic understanding. The YAML provides the skeleton; the LLM provides the soul.
| Drescher | MOOLLM Skill |
|---|---|
| World state | YAML files in skill directory |
| Items | Fields in state files |
| Actions | Skill verbs and procedures |
| Schemas | Documented procedures with context/result |
| Extended Context | Prerequisites, dependencies |
| Extended Results | Side effects, outputs |
| Synthetic Items | Undocumented state the skill discovers |
| Composite Actions | Multi-step procedures |
| Spin-offs | Refined procedures from experience |
"If you can observe patterns, you can discover causality." "If you track correlations, you can spin off knowledge." "The YAML provides the skeleton; the LLM provides the soul."
Based on SOC occupation classification