Deferred feedback batching
Deferred feedback batching is the plan's lowest-latency learning tier, which records feedback now and trains later.
Project status: Research proposal/design in the inspected project sources; no implementation or completed local efficacy run identified. This entry describes the source audit of 14 September 2026; historical measurements retain their original dates.
Mechanism
Append the prompt, response, and feedback to the trajectory log and return before taking a gradient step. A later idle or threshold-triggered process chooses the actual learning method and drains eligible records.
Implementation and controls
Plan T0.1 describes log-and-batch as an engineering composition rather than a loss. Current trajectory logging and idle replay supply pieces of the design, but that does not establish a complete per-event deferred-mode contract with exactly-once drain semantics. Such a contract must specify acknowledgements, eligible records, duplicate handling, and visibility of applied commits.
Evidence and evaluation
The cited project document records this candidate and its intended experiment. It does not provide a completed local result for this method. Published-paper results mentioned by that document are background, not Trainfer measurements.
Limitations and interpretation
Acknowledging recorded feedback does not mean weights have changed. Process-local replay caps and retries can alter exposure. Latency figures in the tier menu are design targets, not measurements for the current hardware/model.
Sources
- trainfer: trainfer/PLAN.md — checkout audited
1c6391f3773b. - trainfer: trainfer/engine/replay.py — checkout audited
1c6391f3773b. - trainfer: trainfer/controller.py — checkout audited
1c6391f3773b.