P1, P2, P3 — and P4 never fires.
The trace, cycle by cycle:
- WM
{goal:tea, kettle:empty}— P1 matches (2 conditions), P4 matches (1). P1 wins on specificity →kettle=full. - WM
{goal:tea, kettle:full}— P2 matches (3 conditions, counting the NOT), P4 matches (1). P2 wins →water=hot. - WM
{goal:tea, kettle:full, water:hot}— P3 matches (2), P4 (1). P3 wins →tea=made,goalremoved. - No
goal=teain WM → nothing matches (even P4) → halt, tea in hand.
P4 — the doom-scrolling rule — matched every single cycle and never fired, because something more specific always outranked it. Now rerun the system under a different policy (“most recently added rule wins”, or random choice among matches) and this same rule set checks the phone forever. Nothing in the rules changed; only the arbitration did. That’s the lesson in one line: in any architecture where multiple things could run, the choosing policy is not plumbing — it is the behavior. Hold that thought for exactly one lesson: the blackboard architecture is what happens when conflict resolution grows up and gets a name (opportunistic control).
Why production systems earned four decades of loyalty from cognitive modelers:
- Modularity of knowledge. Each rule is an independent shard of know-how. Adding skill = adding rules, no rewiring — the property Hearsay-II will scale up to whole subsystems.
- Reactivity. Control re-derives from the situation each cycle, so the system handles interruptions gracefully — like the cook who deals with the boiling-over pot mid-recipe.
- Psychological teeth. With the right (learnable) rules and timing assumptions, production systems reproduce human learning curves and error patterns — the bet that becomes ACT-R.
The harness parallel: an LLM agent loop is a recognize–act cycle — context window as WM, match+conflict-resolution collapsed into one gloriously opaque step (the model reading its prompt and choosing a tool), act = tool call appending facts to context. What the classic systems make explicit — and what you can steal — is the arbitration layer: OPS5-era strategies like specificity, recency, and refractoriness (don’t re-fire on the same data — instantly recognizable to anyone whose agent has called the same tool with the same arguments five times).
Pitfall: dismissing this as “just a rules engine.” The claim Newell & Simon attached to it — the physical symbol system hypothesis — is that this kind of machinery is sufficient for general intelligence. Right or wrong (the field still argues; stage 7 hosts the fight), the architectures it spawned are the best-tested models of human cognition ever built.
Where this goes: next lesson, the payoff this track was built around — what happens when the “rules” become whole expert subsystems, working memory becomes a shared multi-level hypothesis board, and conflict resolution becomes an intelligent scheduler. Speech understanding, 1976: Hearsay-II.