7.7 KiB
StarForth Physics Runtime
PHYSIOLOGICAL MONITORING PRINCIPLES
(Non-Perturbative, t≈0, History+Present Feedback Model)
Author: Captain Bob Purpose: Canonical rules for ALL AI assistants and contributors Scope: StarForth’s physics runtime, monitoring subsystem, inference engine, DOE integration
1. Overview
The StarForth VM is built as a physics-driven organism, not a traditional interpreter. It contains several continuous, interacting feedback loops:
- Heat accumulation & decay
- Rolling Window of Truth
- Hotwords cache dynamics
- Prefetch heuristics
- Transition graph formation
- Adaptive decay slope tuning
- Adaptive heartbeat interval control
These loops must exist and operate before any monitoring or metrics can be safely attached.
Monitoring is non-perturbative and observational only. No sensor may inject latency, allocations, reordering, contention, or architectural changes.
This document defines the fundamental principles for where, when, and how monitoring occurs.
2. Physiology Before Instrumentation
Instrumentation must never be built before the runtime’s physiology is wired.
Required order of operations:
- Build core runtime loops
- Heat & decay
- Window updates
- Prefetch logic
- Inference heuristics
- Tick/heartbeat loop
- Transition metrics
- Let the system run and reach first-cycle stability
- Only then attach:
- Sensors
- DOE metrics
- Event bus listeners
- Observability points
Why? Before the loops run, there is no signal to observe; metrics would produce garbage or perturb the VM.
3. Monitoring Begins at t≈0 (Earliest Stable Cycle)
3.1 Window Bootstrapping: Empty Window Is VALID
The Rolling Window of Truth must begin empty at VM startup.
This is intentional and required.
Principle:
The VM must be able to operate with no history and adapt “with what it has,” allowing inference mechanisms to form their earliest predictions from sparse data.
This means:
- An empty window is not an error condition
- No artificial seeding or fake signal insertion is allowed
- Inference must be able to:
- detect missing history
- treat first values as baseline
- adapt gradually as the window fills
- Early fluctuations are normal and desired
- Decay slope tuning and window-size tuning must behave conservatively when history is thin
- Prefetch heuristics must NOT assume a non-empty window
- No subsystem may depend on the window being pre-populated
This startup period is the system’s initial learning phase,analogous to:
- the warm-up cycle of a PID controller
- the first few samples of a strip-chart recorder
- the startup of a biological homeostasis loop
The window will fill naturally as the VM runs, and all feedback loops will stabilize accordingly.
Monitoring should begin at:
The earliest moment when all feedback loops have completed at least one internal cycle, but before any long-term adaption occurs.
This is referred to as t≈0:
- Not literal time zero
- But the first physiologically meaningful moment
At t≈0:
- The Rolling Window of Truth has its first value
- Heat has been incremented at least once
- Cache promotion may have occurred
- Decay slope has an initialized baseline
- Prefetch tracking exists
- Transition graph has ≥ 1 edge
- Heartbeat loop has run once
From this point forward, the VM has true physics to observe.
Where seed values are needed, they will be provided buy the builder when they configure their build.
4. Non-Perturbative Observation
Monitoring must never disturb the VM or affect:
- dictionary ordering
- bucket heat
- window width
- inference tuning
- lookup strategy
- word execution timing
- tick interval
- cache hits/misses
Absolute rules:
- No malloc/free
- No locks (use read-only pointers)
- No dictionary traversal
- No reordering
- No I/O to stdout during VM operations
- No touching any structure that the physics feedback loops mutate
Monitors observe. They never participate.
5. History + Present → Inference → Adjustment
This is the golden rule governing all inference cycles.
Every adaptive subsystem:
- decay slope tuning
- window width adjustment
- prefetch ROI
- hotword promotion heuristics
- adaptive heartbeat pressure
…must base its decisions on:
1. History (integral component)
Examples:
- Rolling Window of Truth
- Previous decay slope
- Heat momentum
- Prefetch historical hit/miss ratio
- Transition graph edges
- Long-term pressure average
2. Current Value (proportional component)
Examples:
- current heat spike
- current window entry
- instant tick-time pressure
- current transition
- current hit/miss event
3. Inference Process
- smoothing
- numerical weighting
- confidence scoring
- stochastic biasing (deterministic, seeded)
- bounded slope tuning
- bounded window expansion/contraction
4. Adjustment (output)
- updated decay slope
- updated window width
- updated prefetch heuristics
- updated heartbeat interval (if enabled)
- updated hotword promotion threshold
This mirrors a PID controller but tuned for StarForth’s information thermodynamics.
6. Monitoring Points Must Be Explicit and Side-Effect-Free
For every feedback loop (FL1–FL4), there should be explicit, isolated sampling points:
- heat updated
- window updated
- decay slope recalculated
- inference cycle completed
- prefetch stat updated
- transition recorded
- heartbeat tick completed
These sampling points must:
- Use only const VM* (no mutation)
- Read-only access to all internal structures
- Never influence scheduling or ordering
- Never produce allocations
- Never block the VM
7. Monitoring Feeds DOE, Not the VM
All monitoring output goes to one of three sinks:
-
Event Bus
- Optional, listener-based, non-intrusive
-
DOE Metrics Collector
- Single-line CSV output
- Deterministic
- No logs on stdout
-
Optional Debug Logging
- stderr only
- never in CICD
- never in DOE mode without explicit flag
Monitoring informs experimentation. It does not control the VM.
The VM’s physiology controls itself.
8. Monitoring Must Not Create New Feedback Loops
The system already has four core loops:
- FL1: Heat accumulation & decay
- FL2: Hotword promotion
- FL3: Inference → decay slope tuning
- FL4: Inference → window width tuning
Monitoring must not introduce:
- FL5 feedback contamination
- logging-based slowdown
- mechanical timing noise
- heat changes due to sampling
- cache busting due to instrumentation
Monitoring describes. It does not shape.
9. Summary Checklist
Mandatory:
- Physiology wired before instrumentation
- Monitoring begins at first stable cycle (t≈0)
- Monitors are non-perturbative
- Metrics use history+present → inference → adjustment
- Metrics never modify physics state
- DOE output is isolated and side-effect-free
Forbidden:
- Sampling inside locked dictionary regions
- Allocating memory during monitoring
- Reordering dictionary due to instrumentation
- Logging to stdout during VM operation
- Creating new feedback loops via metrics
10. Closing Notes
StarForth is not a “VM with some metrics.” It is a self-regulating cybernetic organism with measurable physics.
These monitoring principles are doctrinal. Any assistant, contributor, or future tool must follow them without modification.
Misunderstanding this document = breaking StarForth.