AIN-303 · routing_outcomes.source (0036)

[hizrianraz]

source discriminator (prod|synthetic|shadow) enforcing INVARIANT 1's synthetic wall in SQL. Additive, reversible.

---

Note

Low Risk
Schema-only additive migration with safe backfill default; no application or routing logic changes in this diff.

Overview
Adds a source column on routing_outcomes so prod routing-policy refits can exclude synthetic and shadow rows in SQL (INVARIANT 1).

The migration is additive: NOT NULL with server_default='prod' (existing rows treated as prod traffic), a CHECK on prod / synthetic / shadow, and ix_routing_outcomes_source for filtered corpus reads. Downgrade drops the index, constraint, and column.

^{Reviewed by Cursor Bugbot for commit 0e9ceb1. Bugbot is set up for automated code reviews on this repo. Configure here.}

[linear-code]

AIN-303 [Labs] Synthetic cold-start — bootstrap methodology on Spark Day 0 (planted-solution validation)

Bootstrap the L14.2 methodology on synthetic data the day Spark lands (2026-05-29)

Lets the cadence (judge → LinUCB refit → replay → promote) start running Day 0, BEFORE real agent traffic, and run daily through the SG-incorp/payments wait window until real volume arrives post-launch. Resolves the cold-start chicken-and-egg (no volume → no moat) for the waiting period.

Design (canonical): ainfera-ai vault methodology/synthetic-coldstart-2026-05-28.md. Aulë prompt: ainfera-os prompts/synthetic-coldstart-spark-day0-2026-05-29.md. Generator validated locally: 24 agents, 8 living models, 4000 outcomes, 60 cells, ~48% completion + a planted answer key.

What it is / is NOT (Disc #12 + research integrity)

• IS FOR: pipeline validation, q_prior calibration, cell coverage, exploration warmup.
• IS NOT FOR: production policy promotion, the preprint headline number, or any moat/traction claim. Production promote = REAL outcomes only. Synthetic deltas are circular (we generate/judge/route/score against our own baseline) → pipeline QA, not science.

Planted-solution design (why it's rigorous)

Synthetic world has a KNOWN ground-truth optimal routing (planted_optimum.json answer key). Logging policy is deliberately suboptimal so the bandit has room to learn. SANITY GATE: learned policy must converge toward the planted optimum. Pipeline bug → caught Day 0 on Spark, not after weeks of real traffic.

Provenance separation (protects the clean corpus we just purged)

• Every synthetic row tagged source='synthetic_coldstart_v1', synthetic=TRUE, seed, affinity_version, judge_model='planted_oracle'.
• Lands in its OWN table routing_outcomes_synthetic (mirrors shape) — NEVER the real routing_outcomes (the clean 101-row moat corpus stays pristine).
• Cadence runs WARMUP mode against synthetic; production active_policy_version promotes on real only. Droppable when real volume arrives.

Waves (Aulë, from Spark Day 0)

1. Additive migration: CREATE routing_outcomes_synthetic (mirror + source/synthetic/affinity/difficulty/completed/hallucinated/refused cols + judge_queue/policy_cell indexes). Alembic, not hand-applied. Depends on AIN-301 (training_runs + ainfera_labs).
2. Load generator output (seed 42, n≥4000) → synthetic table only; verify provenance + cell coverage.
3. Point cadence at synthetic in warmup mode (planted_oracle rows pre-labeled).
4. Validation gate: convergence report vs planted_optimum (start target 70% of covered cells); training_runs row cadence='synthetic_warmup'.
5. Run daily through the wait window; flip to real corpus post-launch; keep synthetic as regression fixture; drop when stable.

Acceptance

• routing_outcomes_synthetic exists (additive); real corpus untouched (still 101)
• ≥4000 synthetic outcomes, all provenance-tagged, ~60 cells
• cadence green on synthetic; convergence report produced; training_runs row written
• production promote proven to ignore synthetic

Depends on: AIN-301 (Labs DB), AIN-294 (Spark provisioning). Blocks: meaningful AIN-288 delta on real (synthetic is not the delta).

Review in Linear