Minimal AI stack

This document describes a minimal, layered AI instrument architecture.

The goal is stable behavior under declared regimes, with explicit retrieval scope and full lineage.

Layer 0: Substrate#

• Model runtime (e.g., ONNX runtime)
• Model artifacts:
  • generator (LLM)
  • optional embedder
  • optional reranker

Rule: models do not decide policy. They generate or score.

Layer 1: Regime gate#

A regime object is created at session start and attached to every turn.

• Default regime: rtt=1 | coherence=declared | drift=bounded | paradox=structural

Rule: no turn executes without a regime snapshot.

Layer 2: Lineage ledger#

An append-only ledger records every event required to reproduce outputs.

Rule: if it is not in the ledger, it did not happen.

Layer 3: Retrieval instrument#

Retrieval is bounded context acquisition, not “memory magic.”

Retrieval scope is explicit:

• none | local | session | prior_sessions | web

Rule: retrieval scope MUST be declared per turn (or defaulted) and logged.

Layer 4: Prompt compiler#

Prompt assembly is treated as compilation.

Inputs:

• regime snapshot
• task envelope
• constraints
• retrieved context
• conversation slice (bounded)

Output:

• prompt package + prompt hash

Rule: prompt assembly MUST be deterministic given inputs.

Layer 5: Post-processing policy#

Post-processing enforces the declared regime.

Minimum checks:

• coherence: did we stay in the declared frame?
• drift: did we exceed bounds or fail to label speculation?
• paradox: did we log contradictions as objects when present?

Rule: post-processing MUST be logged as an event.

Layer 6: UI portal#

Minimal UI exposes the instrument controls without restricting capability.

Always visible:

• regime line
• retrieval scope
• lineage indicator (exportable ledger)

Optional toggles:

• include prior sessions in retrieval
• tighten/open drift
• audit mode (rtt=2)

Rule: no hidden mode switches.

Upgrade discipline#

Upgrades are lineage events, not silent replacements.

A new model artifact implies:

• new model_hash
• new tokenizer_hash (if applicable)
• new runtime_version binding

Rule: the instrument can evolve, but lineage must remain continuous.