Phase 1 (GATE) — Dataset + forgetting floor

[Tejas-JB]

PRD §6 Phase 1. This is a hard gate — Phase 2 cannot start until the exit criterion is met.

Deliverables

• src/dreamcue/data/generator.py — programmatic synthetic fact generator
  • ~600 learn facts, fixed seed
  • random 20% flag (random designation eliminates the easy-flagged-facts confound)
  • paraphrased probe per fact via handcrafted templates (NOT via the LLM under test — leak risk; see docs/decisions.md)
  • ~2000 interference facts on overlapping entities with conflicting values
• Learn-phase LoRA fine-tune to ≥90% probe accuracy on BOTH flagged and unflagged
• No-confound check: proportion z-test between flagged and unflagged probe accuracy, p > 0.1 → write to results/phase1-learn.json
• No-replay arm: continue training on interference stream, probe at checkpoints, save retention curve
• results/phase1-gate.json with the gate result

Gate

Flagged probe accuracy must drop ≥25 absolute points from learn-phase peak.

Backstep policy if gate fails (owner-confirmed)

1. Dial up interference: more facts, more entity overlap, higher LR.
2. If still flat, escalate to meta-llama/Llama-3.2-3B-Instruct.
3. Only then stop and report.

Every retry is logged in docs/phase1-tuning.md.