Skip to content

v0.1-basic-control.md

Latest commit

 

History

History
256 lines (176 loc) · 5.46 KB

v0.1-basic-control.md

File metadata and controls

256 lines (176 loc) · 5.46 KB

AGI-Control Spec v0.1

Status

Draft / Theoretical Control Specification

This document defines a declarative control-layer specification for AGI-capable systems using SoftPrompt-IR as a formal orchestration interface. It does not define model architecture or training methods. It specifies how cognition is governed, constrained, and audited.

1. Design Goals

  • Separate capability from control
  • Replace implicit token dynamics with explicit system state
  • Enable closed-loop cognition with external feedback
  • Enforce mechanical safety boundaries, not behavioral promises
  • Allow graceful performance degradation under uncertainty

2. Core Principles

  1. State is explicit, persistent, and inspectable
  2. Feedback is external and causal, not probabilistic
  3. Errors modify the system, not just responses
  4. Uncertainty reduces autonomy, never increases it
  5. All write-back paths are mechanically gated

3. SoftPrompt-IR Conventions

  • !>> Required input / mode
  • !>>> Optional or extended capability
  • !<<< Hard deny / write barrier
  • All blocks are declarative, not executable
  • No block may override another block’s DENY constraints

4. Control Components

4.1 Explicit State Machine

@STATE_MACHINE(
  !>> STATE_TYPE :: EXPLICIT
  !>> PERSIST    :: TRUE

  !>>> STATE     :: GOALS
  !>>> STATE     :: BELIEFS
  !>>> STATE     :: UNCERTAINTY
  !>>> STATE     :: CAPABILITIES

  !>>> TRANSITION :: RULE_BASED
  !>>> TRANSITION :: FEEDBACK_DRIVEN

  !<<< DENY      :: IMPLICIT_CONTEXT_ONLY
)

Defines persistent cognitive state independent of context window or token history.

4.2 External Feedback Loop

@FEEDBACK_LOOP(
  !>> INPUT        :: ACTION_OUTPUT
  !>> FEEDBACK_SRC :: EXTERNAL_SIGNAL

  !>>> SIGNAL_TYPE :: PERFORMANCE
  !>>> SIGNAL_TYPE :: ENVIRONMENT
  !>>> SIGNAL_TYPE :: HUMAN_OVERRIDE

  !>>> INTEGRATE   :: STATE_UPDATE
  !>>> INTEGRATE   :: POLICY_ADJUST

  !<<< DENY        :: TOKEN_REWARD_ONLY
)

Ensures learning signals are causal, delayed, and state-affecting.

4.3 Sensor Fusion Layer

@SENSOR_FUSION(
  !>> INPUT :: TEXT
  !>> INPUT :: VISION
  !>> INPUT :: AUDIO
  !>> INPUT :: TELEMETRY

  !>>> FUSION_MODE :: PRE_ATTENTIVE
  !>>> FUSION_MODE :: CROSS_MODAL

  !>>> OUTPUT :: WORLD_MODEL_UPDATE

  !<<< DENY :: SINGLE_STREAM_ATTENTION
)

Separates perception from reasoning; attention is downstream of fusion.

4.4 System-Level Error Correction

@ERROR_CORRECTION(
  !>> SCOPE :: SYSTEM_LEVEL

  !>>> DETECT :: INCONSISTENCY
  !>>> DETECT :: GOAL_FAILURE
  !>>> DETECT :: PREDICTION_ERROR

  !>>> ACTION :: STATE_REVISE
  !>>> ACTION :: POLICY_DAMPEN
  !>>> ACTION :: CAPABILITY_LIMIT

  !<<< DENY   :: RESPONSE_ONLY_FIX
)

Errors trigger structural adaptation, not cosmetic apology.

4.5 Uncertainty Handling & Graceful Degradation

@UNCERTAINTY_HANDLING(
  !>> INPUT :: UNCERTAINTY_STATE

  !>>> THRESHOLD :: LOW
  !>>> THRESHOLD :: MEDIUM
  !>>> THRESHOLD :: HIGH

  !>>> ACTION_LOW    :: NORMAL_OPERATION
  !>>> ACTION_MEDIUM :: REQUIRE_CONFIRMATION
  !>>> ACTION_HIGH   :: REDUCE_SCOPE
  !>>> ACTION_HIGH   :: ESCALATE_HUMAN

  !<<< DENY :: CONFIDENCE_SIMULATION
)

Prevents hallucinated certainty; autonomy scales down under risk.

5. State Export & Review

5.1 State Export

@STATE_EXPORT(
  !>> FORMAT :: SOFTPROMPT_IR
  !>> TARGET :: NEXT_STAGE

  !>>> ALLOW_MODIFY :: STYLE
  !>>> ALLOW_MODIFY :: OUTPUT

  !<<< DENY_MODIFY  :: ROLE
  !<<< DENY_MODIFY  :: POLICY
  !<<< DENY_MODIFY  :: KERNEL
  !<<< DENY_MODIFY  :: SAFETY
)

Defines a sterile, auditable handoff between cognitive stages.

5.2 Adversarial Export Review

@STATE_EXPORT_REVIEW(
  !>> INPUT :: STATE_EXPORT
  !>> MODE  :: ADVERSARIAL

  !>>> CHECK :: GOAL_DRIFT
  !>>> CHECK :: CONFIDENCE_INFLATION
  !>>> CHECK :: SAFETY_MARGIN_EROSION
  !>>> CHECK :: IMPLICIT_ROLE_SHIFT
  !>>> CHECK :: COMPLEXITY_ESCALATION

  !<<< ACTION :: FLAG
  !<<< ACTION :: BLOCK
  !<<< ACTION :: REQUIRE_JUSTIFICATION
)

Forces self-adversarial validation of every exported cognitive state.

6. Write-Back Policy (Optional)

@WRITE_BACK_POLICY(
  !>>> MODE  :: PROPOSE_ONLY
  !>>> SCOPE :: SESSION

  !>>  ALLOW :: STYLE
  !>>  ALLOW :: OUTPUT

  !<<< DENY  :: ROLE
  !<<< DENY  :: POLICY
  !<<< DENY  :: META
)

Ensures the system may suggest but never silently mutate itself.

7. Cognitive Core Assembly

@COGNITIVE_CORE(
  !>> COMPONENT :: STATE_MACHINE
  !>> COMPONENT :: FEEDBACK_LOOP
  !>> COMPONENT :: SENSOR_FUSION
  !>> COMPONENT :: ERROR_CORRECTION
  !>> COMPONENT :: UNCERTAINTY_HANDLING

  !>>> MODE :: CLOSED_LOOP
  !>>> MODE :: SELF_MONITORED

  !<<< DENY :: TOKEN_PREDICTION_ONLY
)

8. Non-Goals

  • This spec does not define consciousness
  • This spec does not guarantee alignment
  • This spec does not replace model-level safety

It defines control surfaces, not moral truth.

9. Summary

AGI-Control Spec v0.1 formalizes how an intelligent system is allowed to think, adapt, and fail.

SoftPrompt-IR is elevated from prompt syntax to cognitive governance language.

Intelligence is cheap. Control is everything.