Crate auracle_session 

auracle-session

The two-loop engine (the reference: The two loops) every frontend drives:

• Patch loop (fast, silent): vetted prior draws fill a pool; once a posterior exists, a short typed-MH walk on π_β ∝ p_grammar · exp(β·E[u_θ]) moves the pool toward the user’s taste (engine::Engine::refine). Local refinement on that target, not a draw from it.
• Taste loop (slow, human-paced): feedback appends to the observation log as raw φ; the posterior re-fits from it, standardizing at fit time (engine::Engine::fit_posterior). Between fits each vote is folded in by importance reweighting, so the next question responds to the last answer.
• Acquisition between them: BALD — expected information gain about θ (engine::Engine::next_duel), which measurably beats the dueling Thompson rule it replaced and ties uniformly-random pairing (engine::Acquisition carries the numbers).

The M4 gate is this crate’s closed-loop test: engine + synthetic user, end-to-end through the real grammar → render → vet → features pipeline, asserting the learned taste ranks genuinely-preferred patches on top.

Re-exports

pub use calib::calibration;

pub use calib::Calibration;

pub use calib::Forecast;

pub use calib::ProvenanceScore;

pub use calib::ReliabilityBin;

pub use engine::phi_names;

pub use engine::tilt_weights;

pub use engine::Acquisition;

pub use engine::BankEntry;

pub use engine::Candidate;

pub use engine::Contribution;

pub use engine::DuelChoice;

pub use engine::EditOutcome;

pub use engine::Engine;

pub use engine::Explanation;

pub use engine::ImplicitEvent;

pub use engine::LineageEvent;

pub use engine::Origin;

pub use engine::Profile;

pub use engine::RefineKeep;

pub use engine::RenderPolicy;

pub use engine::SessionConfig;

pub use engine::SessionState;

pub use farm::draw_seed;

pub use farm::Draw;

pub use farm::PreFeaturized;

pub use map::MapPoint;

pub use map::TasteMap;

pub use naming::claim_name;

pub use naming::NameScale;

pub use surrogate::SurrogateFitness;

pub use surrogate::QUARANTINE_FITNESS;

Modules

calib

Prequential calibration: is the model’s confidence honest?

engine

The two-loop session engine (the reference: The two loops).

farm

The stateless render farm: an indexed pool-draw stream, and the off-engine unit of work it hands out.

map

The taste map: a 2D embedding of everything the user has heard, with posterior utility attached — taste rendered as territory (islands of glow across patch space) rather than a single preference vector.

migrate

Migrating profiles written before raw-φ logging.

naming

Musical names for patches, read off the measured features.

surrogate

The learned taste as a fugue-evo fitness: the surrogate that lets the machine evolve thousands of candidates silently (the reference: The two loops).