Skip to content

PATTERN Cited by 1 source

Closed feedback loop for AI features

Summary

An architectural discipline for employee-facing (and by extension customer-facing) AI features: every recommendation, ranking, or classification surfaced by the model must be independently reproducible, explainable at the component level, and paired with a feedback channel that allows the human consumer to correct, accept, or challenge the output — with that feedback flowing back into training/evaluation.

Three component primitives:

  1. Reproducibility. The human can rerun the inputs and recover the same output. No hidden randomness, no non-deterministic retrieval, no mystery "model revision."
  2. Explainability. The output comes with evidence — which candidates were considered, which rules fired, which signals drove the ranking. Not necessarily full model interpretability; at least sufficient to check the plumbing.
  3. Feedback channel. The human can mark the output correct, wrong, or useless — and that signal aggregates into retraining data, confidence-threshold tuning, or eval-set additions.

Canonical wiki reference

Meta's web-monorepo RCA system (2024-06; sources/2024-08-23-meta-leveraging-ai-for-efficient-incident-response) names the discipline verbatim:

"we ensure that all employee-facing features prioritize closed feedback loops and explainability of results. This strategy ensures that responders can independently reproduce the results generated by our systems to validate their results."

Meta pairs this with confidence-thresholded refusal (patterns/confidence-thresholded-ai-output):

"We also rely on confidence measurement methodologies to detect low confidence answers and avoid recommending them to the users — sacrificing reach in favor of precision."

Closed-feedback-loop + explainability + confidence-thresholding are named as the three-part discipline that keeps an AI ranker trustworthy in an incident-response context where misleading engineers is very expensive.

Why this discipline, specifically

Three failure modes it mitigates:

  1. Silently wrong recommendations. Without a feedback channel, a model that silently drifts keeps emitting confident-sounding top-5s after its training distribution has shifted. With a feedback channel, the drop in acceptance rate surfaces the drift.
  2. Ownership of the model evaporates. An AI feature with no feedback channel is a black box whose original engineers move on. Closed feedback means the next team knows what's failing and why.
  3. "Trust without verify" incidents. In high-stakes contexts (incident response, production-debugging, medical/legal/finance), engineers following a wrong recommendation can escalate damage. Explainability + reproducibility let them verify quickly; the feedback channel captures the correction.

Contrast: open-loop AI features

Open-loop AI features (the anti-pattern) have any of:

  • Recommendation without shown reasoning — "the model says top-5 is X, Y, Z, trust us."
  • No way for the user to mark the output wrong.
  • Feedback collected but not flowing back to training/eval.
  • Reproducibility gated by production access (only the ML team can re-run).

Meta's stance: "sacrificing reach in favor of precision" — if the closed-loop mechanism isn't there, the feature shouldn't ship to engineers.

Implementation hooks

  • Show-your-work UI. Alongside the top-5 ranked changes, show: which candidates were considered, which retriever rules were triggered, which signals boosted each survivor. This is what "explainability" means operationally, even when the ranker is an LLM.
  • Accept/reject on every recommendation. The minimum feedback primitive. Aggregate to per-model acceptance rate; alarm when it drops.
  • "Why not the real root cause?" For RCA specifically: when the human finds the real root cause after the fact, allow them to tag the correct change — this flows into the SFT training set for the next model revision.
  • Model + revision stamping. Every output tagged with the model ID + training-data vintage. Reproducibility depends on this; drift-detection depends on this.

Caveats

  • Feedback infrastructure is not free. Collecting, aggregating, validating, and recycling feedback into training is its own engineering system. Meta does not disclose the shape of its feedback-collection mechanism in this post.
  • Explainability ≠ interpretability. Meta's framing is "components can be independently reproduced" — inputs to the retriever, candidate set, ranker outputs. This is plumbing-level explainability, not model-weight interpretability. Teams hoping for attention-head-level interpretability should calibrate expectations.
  • Precision-over-reach is a product decision. Confidence thresholding turns into silence on low-confidence cases. Whether that's the right trade-off is workload-specific; for incident RCA, silence-when-unsure beats confident-and-wrong; for other workloads, the opposite can be true.
  • Feedback loop closure latency. A feedback tag today might influence the model months later via the next SFT retrain. For fast-moving distributions, that loop closure can be too slow. Meta does not disclose its retrain cadence.

Seen in

Last updated · 319 distilled / 1,201 read