M0 — Trace contract (chatty-trace)

Linear: AGE-5 · Crate: chatty-trace · Promotion: pending

No single paper — this is the contract layer every optimizer reflects on.

Why it exists

AFlow, GEPA, ACE, and DGM all optimize against traces, not final answers alone. Chatty already ships a ReAct loop, ATIF v1.6 export, and token budgeting. M0 adds:

Per-module attribution — which prompt produced each step (GEPA credit assignment)
ATIF round-trip — Deserialize path so reflection reads trajectories back in
FeedbackFn — scalar + natural-language feedback per module (GEPA’s µ_f)

Chatty mapping

Paper concept	Chatty location	Status
Trajectory `(thought, action, observation)*`	`chatty-trace::Trajectory`	Reserved — human writes
`Action::Language` (ReAct’s Â = A ∪ L)	`chatty-trace::Action`	Reserved
Per-step `ModuleId`	Extension over `AtifStep`	To build
ATIF export	`chatty-core` exporters (Serialize-only today)	Round-trip in M0
Trace retention	`Recorder` trait, no-op in release	To build
Train vs validation rollouts	`RolloutBudget` over `token_budget/`	To build

Capture attaches to the existing loop via rig-agent hooks + rig-tap — no hand-rolled agent loop.

Eval protocol

Stage A: HotpotQA run produces a Trajectory that round-trips; two seeded runs on a warm cache are byte-identical.

Stage B: Harbor runs consume exported ATIF; attribution must survive export/import.

Production landing

Mechanism	Likely promotion
No-op `Recorder` in release	Default
Full step retention	Setting (`TrainingSettingsModel.atif_auto_export` already exists)
`FeedbackFn` implementations	Offline / optimizer only

Reserved symbols (human only)

Trajectory, Step, Action, Outcome, FeedbackFn — see RESERVED.md.

Depends on

AGE-32 rig 0.42 upgrade
AGE-22 walking skeleton findings

Chatty Documentation