PRD — On-device continual-learning loop

Thesis: every correction a user makes is free, high-signal training data that never has to leave the Mac. Capture it, refresh a LoRA on it overnight, gate it, promote it. The model gets better from real usage — and privacy isn’t a compliance checkbox, it’s the default.

This is the Trajectory steal (trajectory.ai, docs): “Instrument → Understand → Steer → Learn.” Trajectory and Applied Compute both sell continual learning to enterprises and have to bolt on SOC 2 + “you approve every update” + audit trails to make it palatable. On-device, those are structural givens — the wedge competitors pay overhead to fake.

Why / What

Problem: TinyGPT’s factory is an offline batch loop. A specialist ships once and is frozen until someone manually re-runs the pipeline. The market (Trajectory, Applied Compute) has moved the loop online — improve continuously from production signal.
Users: anyone running a local specialist/planner (Pace, dictation post-proc, a per-task assistant) who corrects its output. Their correction is ground truth.
In scope: capturing corrections, turning them into training pairs, an automatic gated LoRA refresh, user-gated promotion. All local.
Out of scope (for now): any cloud upload, federated learning, RLHF/online-RL with a reward model (Trajectory does text-reward — we start with supervised corrections), multi-user aggregation.

What already exists (do NOT rebuild)

The offline pipeline is shipped — the continual loop wraps and automates it:

AgentTrajectory / .atraj (native-mac/Sources/TinyGPTModel/AgentTrajectory.swift, Agent.swift, AgentLoop.swift) — token-preserving rollouts.
tinygpt traces-to-data (TracesToData.swift) — trajectories → SFT pairs.
tinygpt sft (SFT.swift) — LoRA/DoRA fine-tune (DoRA default; magnitudes persist since the TGLA-v2 fix).
tinygpt eval-gate (EvalGate.swift) — the promotion gate (BFCL/τ-bench/lm-eval).
tinygpt bake-lora (BakeLora.swift) — merge adapter for deployment.

What’s new (the delta)

Correction capture. .atraj records completed rollouts; it does not record when a user edits / rejects / retries / overrides an output. Add a correction event type (original output, corrected output, intent kind) and hook it into the points where the user can correct: dictation edit, tool-call reject, action override. Append to a local correction store. The corrective text is the target — Trajectory’s “not just a binary reward but the actual text” insight.
Curation. Extend traces-to-data to turn correction events into (input, corrected-output) SFT pairs (and/or DPO pairs: chosen=corrected, rejected=original). Dedup; optionally reject-sample low-confidence corrections.
Scheduled refresh. Periodic (overnight / idle, under caffeinate) LoRA/DoRA fine-tune on accumulated corrections. Sequential, not Trajectory’s concurrent sampling/training pools — single-GPU contention on a Mac means sequential + caffeinate (see the parallel-training lesson). Mix in a replay sample of the original SFT data to resist catastrophic forgetting (ratio = TBD).
Gate + auto-rollback. The refreshed adapter must pass eval-gate AND not regress vs the current adapter before promotion. Regression ⇒ discard, keep current. This is the no-quality-regression rule made automatic — non-negotiable for an unattended loop.
User-gated promotion. Default: auto-promote only if the gate passes, with a notification (“your dictation model improved on 47 of your corrections; using it now”). Optional always-confirm mode. Nothing ever uploads.

Design (the loop)

serve/agent ──(user corrects)──> correction store (local, on-device)
                                        │  accumulate to threshold N
                                        ▼
                          traces-to-data (corrections → SFT/DPO pairs)
                                        │  + replay sample of base SFT
                                        ▼
                          sft (LoRA/DoRA refresh, overnight, caffeinate)
                                        │
                                        ▼
                          eval-gate (must pass + not regress)
                            │ pass                 │ fail
                            ▼                       ▼
                   promote (user-gated)        discard, keep current

Risks / gates

Quality regression / overfitting to recent edits → eval-gate + rollback (mandatory)
- replay mix. The single most important guardrail.
Feedback loop — model trains on corrections to its own outputs; could drift. Mitigate by gating on a fixed held-out eval, not just recent corrections.
Sparse/noisy corrections — how much before a refresh is worth running? (TBD, measure.)
Compute — overnight LoRA on a Mac is feasible (M1 Pro 16GB secondary w/ throttle per the training-hardware note); schedule when idle.

Open questions (fill after learning / measuring — do NOT pre-answer)

Minimum correction volume N to trigger a refresh? — TBD, measure on real usage.
DoRA vs LoRA for incremental refresh? — likely DoRA (more lift), confirm it doesn’t overfit on small correction sets.
SFT pairs vs DPO pairs (chosen/rejected) from corrections — which generalizes better? — TBD.
Replay ratio (corrections : original SFT) to balance plasticity vs forgetting? — TBD.
Auto-promote-if-gate-passes vs always-user-confirm — product call. — TBD.

Phased plan (smallest shippable first)

Capture only — ✅ SHIPPED (2026-06-24). CorrectionEvent + append-only JSONL CorrectionStore (TinyGPTIO/CorrectionEvent.swift, pure Foundation), tinygpt record-correction CLI (--intent/--original/--corrected/…, --list), defaulting to ~/.tinygpt/corrections. Unit-tested (CorrectionStoreTests, runs under swift test, no Metal). Ingestion: tinygpt record-correction CLI AND serve POST /v1/corrections (✅ added 2026-06-24 — --corrections-dir, atomic O_APPEND store, 0600 perms; clients like Pace can capture directly). Phase 1 complete.
Curation — ✅ SHIPPED (2026-06-24). tinygpt corrections-to-data --out … [--format sft|dpo] [--intent …] [--replay base.jsonl --replay-ratio r]. Pure curation in TinyGPTIO/CorrectionCuration.swift (sftPair/dpoTriple/CorrectionCurator, unit-tested); SFT rows match the traces-to-data ChatML shape. Corrections without an input (can’t ground a pair) are skipped + reported. Replay mix stride-samples the base SFT file (deterministic) to resist forgetting.
Gated refresh — scheduled overnight sft + eval-gate + auto-rollback. The core loop, unattended-safe.
Promotion UX — user-gated apply + notification.

External reference

Trajectory — continual learning platform, “Instrument/Understand/Steer/Learn”, Continuous LoRA (split sampling/training pools): https://docs.trajectory.ai/introduction · https://trajectory.ai/field-notes/multi-lora-training-for-continual-learning
Applied Compute — “Improve” via production data + online RL: https://www.appliedcompute.com