Skip to content

Latest commit

 

History

History
181 lines (117 loc) · 4.04 KB

File metadata and controls

181 lines (117 loc) · 4.04 KB

Adaptive Core — Architecture

This document describes the conceptual and operational architecture of Adaptive Core. It is not an implementation guide. It is a semantic map of the system.

The code remains the primary specification.


1. Design Goal

Adaptive Core is designed to make failure explicit, bounded, and irreversible.

The system does not attempt to optimize outputs, recover indefinitely, or conceal degradation. Instead, it enforces a controlled descent into a terminal state when anomalous conditions persist.

This is intentional.


2. System Model

Adaptive Core is a pure state machine.

At every step:

Input:

  • RawMetrics (entropy, divergence, repetition)
  • Previous immutable SystemState
  • Absolute wall-clock time
  • Static configuration

Output:

  • New immutable SystemState

There are no side effects. There is no hidden state. There is no learning.


3. Time Model

Time is modeled on two independent axes:

3.1 Wall-Clock Time

  • Represents external, real-world time
  • Used for: - anti-stutter protection
    • starvation detection
    • ontological reset after prolonged absence

Wall-clock time is assumed monotonic. Violations are guarded against defensively.

3.2 Logical Time

  • Represents internal control evolution
  • Advances only when a control step is performed
  • Used for:
    • mode timeouts
    • cooldown duration
    • policy semantics

Logical time is monotonically non-decreasing.

In FALLBACK, logical time may advance, but control physics are frozen.


4. Modes (Control States)

The controller operates in exactly one of four modes:

STANDARD (STD)

  • Normal operation
  • Energy may be spent to escape anomalies

EMERGENCY (EMG)

  • Entered when repetition exceeds threshold
  • Consumes discrete energy on entry
  • Intended as a short-lived escape regime

COOLDOWN (CDN)

  • Entered after EMERGENCY timeout
  • Gradually restores energy
  • Acts as a damping phase

FALLBACK (FBK)

  • Terminal, absorbing state
  • No transitions out are permitted
  • Control physics are frozen
  • Sampling configuration is minimal and deterministic FALLBACK represents ontological failure, not temporary safety mode.

5. Energy Model

Energy is:

  • Discrete
  • Bounded
  • Spent only on EMERGENCY entry
  • Partially restored during COOLDOWN
  • Partially restored (single bounded +1, once per lifecycle) after prolonged absence (>24h); see §6 Ontological Reset

Energy exhaustion under sustained anomaly leads to FALLBACK.

There is no energy regeneration loop that guarantees recovery.


6. Ontological Reset (Energy-Conservative)

After prolonged absence of interaction, the system may perform a single, bounded energy recovery.

This reset:

  • Grants at most one unit of energy (δ ≪ Eₘₐₓ)
  • Is permitted only once during the entire system lifetime
  • Is tracked via an immutable reset_used flag in SystemState
  • Does not constitute recovery from failure
  • Exists solely to prevent permanent starvation due to initial inactivity

This design preserves irreversibility under sustained anomaly while acknowledging real-world usage patterns.

Formally: ∑ᵢ wᵢ ≤ 1, where wᵢ indicates reset activation at step i.


7. Metrics and Smoothing

Raw metrics are externally provided and normalized.

Internal state tracks smoothed versions of:

  • entropy
  • divergence
  • repetition

Currently:

  • Only repetition participates in emergency logic
  • Other metrics are preserved for forward compatibility

State evolution is deterministic given inputs.

8. Failure Philosophy

Failure is a valid and expected outcome.

The system:

  • Surfaces degradation
  • Bounds escape attempts
  • Refuses infinite recovery loops

Once FALLBACK is reached, the system remains there.

This is not pessimism. This is honesty.


9. Formal Properties (Informal)

The system satisfies:

  • Referential transparency
  • Deterministic transitions
  • Bounded resources
  • Explicit terminal state
  • Suitability for formal specification (e.g. TLA+)

The implementation is an executable specification of these properties.


End of architecture.