Llm campaign drift gate

LLM Campaign Drift Gate

Problem

Batch-scoring campaigns (exposure measures, classifiers, extraction runs) call provider aliases that can be silently repointed to a new model at any time. Resuming a half-finished campaign after the alias moves splices two different scorers into one variable, with the version boundary correlated with whatever orders the chunks (time, firm id) — a silent confound. Providers can also RETIRE the old model entirely, making the original campaign uncompletable.

Context / Trigger Conditions

• Resuming a scoring run more than ~a day after its last paid chunk
• "Top up credits and finish the run" requests
• Any incremental scoring against an existing response cache
• Symptom of a missed gate: a step-change in scores at a resume boundary

Solution

Before ANY production spend on resume, run a two-part gate (~$0.30–2):

1. Canary (the decisive check): sample ~100 already-cached items, re-send their EXACT stored prompts fresh, compare fresh vs cached scores. Gate: ≥97% all-field exact match and no systematic directional shift. Write the comparison in a standalone script — never through the pipeline's cache layer, which would overwrite production entries.
2. Gold re-validation: re-score the gold/validation panel fresh and compare agreement metrics to the prior validation (e.g. median F1/κ within ~0.03, no domain dropping >0.10).

Also capture response.model on every gate call — pipelines rarely store it, and it is the only direct evidence of a repoint. Check the provider's /models endpoint: if the old model id is gone, no rollback exists.

3. If the canary fails, diagnose BEFORE concluding — two mandatory follow-ups:
   • Date the suspected flip against the provider's changelog before inferring a model splice. response.model on fresh calls identifies today's model only; if the alias already pointed there when the cache was written, there is no family splice and the mismatch needs another explanation. (Verified failure mode: an alias that had served the "new" model for months was misread as a fresh repoint.)
   • Fresh-vs-fresh canary to separate serving drift from temperature-0 nondeterminism: re-score the same items a second time. Drift signature = fresh2-vs-fresh1 agreement high and symmetric while both fresh runs disagree with the cache at a higher rate in the SAME signed direction. Noise signature = fresh-vs-fresh disagrees about as much as fresh-vs-cache, with no directional bias.
   • Supporting forensic: compare raw-response formatting fingerprints (JSON pretty/compact ratio, key order) between cache and fresh — a heterogeneous or shifted style distribution corroborates a serving change when no model id was recorded.

Key subtlety (why both checks): a new model or revision can validate AGAINST GOLD as well as the old one (κ holds or improves) while still disagreeing with the old scores on 10–30% of items, concentrated in borderline-heavy fields. Gold agreement does not license splicing — the gate fails on the canary alone. And alias stability is not serving stability: the same alias serving the same model family can still drift across days via silent serving revisions; a canary-failed resume is a seam either way, and the decision (resume with a documented seam vs re-score the universe) belongs to the budget owner.

Verification

The gate script logs: fresh response.model ids, canary exact-match rate, per-field mismatch counts with signed direction, and the gold-metric deltas. GO only if both checks pass.

Example

T1 exposure_v2 resume, 2026-07-16: canary returned 71% exact (gate ≥97%) with a litigation-concentrated negative shift, yet holdout median κ improved 0.607→0.644. First interpretation — "alias repointed to a new model family" — was WRONG: the provider changelog showed deepseek-chat had served v4-flash since April, months before the campaign. The fresh-vs-fresh follow-up then isolated the true cause: fresh2-vs-fresh1 93% exact/symmetric/litigation 0, both fresh runs vs cache 71–72% with litigation −12 identically — a serving revision within the same model across a 2-day gap, corroborated by a shifted JSON-formatting fingerprint. Total diagnosis cost ~$0.30; the resume-vs-rescore decision went to the budget owner with the seam quantified.

Notes

• Design campaigns for this failure: per-response content-addressed cache + append-only checkpoint makes "re-score everything under the new model" a clean cache-rotation, not a data loss.
• If the cache key embeds the alias string rather than the resolved model, record actual response.model in run reports — the cache cannot tell you later which model produced an entry.
• One campaign = one model. Budget and schedule so the universe completes within days, or accept that a provider release can force a full re-score.
• See also: [llm-gold-bound-failure-check] for the companion pre-campaign check — whether a validation-gate failure is fixable by prompt at all, or bound to the gold construct.