Crate airframe_observe 

airframe-observe

FSE-based inference observation layer, powered by dzero.

The Core Inversion

Conventional inference capture is observer-first (broken pattern):

vault_seed:   forward_pass() → capture hidden_states
candle_probe: forward_pass() → capture logits
fse_policy:   forward_pass() → scan output_text
Cost: O(N_observers × M_selectors)

This crate is selector-first (FSE/dzero architecture):

ObservationSession::new(plan)  ← compile once
forward_pass emits facts        ← single pass
All observers receive data simultaneously via dzero broadcast
Cost: O(M_selectors), independent of observer count

Usage

use airframe_observe::{ObservationSession, InferenceFact};

// 1. Build a session with registered observers
let mut session = ObservationSession::new();
session.register_vault_oracle();
session.register_candle_compare();

// 2. During forward pass — emit facts via convenience helpers
let hidden: Vec<f32> = vec![0.1; 2048];
session.emit_layer_output(0, 1, &hidden);

let logits: Vec<f32> = vec![0.01; 32000];
session.emit_final_logits(1, &logits);

// 3. Run to fixpoint — all observers receive their data
let result = session.saturate();
assert!(result.saturated);

Patent Notice

This selector-first, compile-once, single-pass observation architecture applied to AI inference is covered by a pending US patent by Michael A. Kuykendall. All rights reserved. Contact: michaelallenkuykendall@gmail.com

Re-exports

pub use facts::alpha_key_of;

pub use facts::HaltReason;

pub use facts::InferenceFact;

pub use isf::GenerateOutput;

pub use isf::ISFState;

pub use isf::InferenceSaturationFabric;

pub use observer::CandleCompareObserver;

pub use observer::VaultOracleObserver;

pub use output::ObservationData;

pub use session::ObservationSession;

Modules

facts

InferenceFact — the domain fact vocabulary for inference observation.

isf

Inference Saturation Fabric (ISF)

observer

Observer — anything that wants to receive inference data.

observers

Built-in observers for the vault pipeline.

output

ObservationData — the value broadcast to observers when a selector fires.

plan

ObservationPlan — the compiled FSE execution plan for inference observation.

quant_formula

Canonical GGML/GGUF quant dequantization formulas — the spec-cited math core.

session

ObservationSession — the live runtime for a single inference pass.