hitl-override-metrics-and-feedback-audit
Audits a human-AI partnership's health with override-rate and override-accuracy metrics (catching both automation bias and an under-performing model), and designs the feedback loop that turns human co
SKILL.mdhuman-ai-collaboration-design/skills/hitl-override-metrics-and-feedback-audit/SKILL.md
GitHubHITL Override Metrics & Feedback Audit
Purpose
Turns "we have human oversight" from a design claim into a measured, ongoing fact. A HITL process that was well-designed at launch can quietly degrade into rubber-stamping, or can reveal a model that was never actually fit for the task, and neither failure mode is visible without tracking the right metrics over time. This skill defines those metrics and the audit cadence, and designs the loop that captures human corrections as an asset rather than letting them evaporate after each individual decision.
Anchored in research
- Su Belagodu — the override-rate diagnostic (below ~5% signals automation
bias risk; the team should investigate whether that reflects genuine
model accuracy or reviewers disengaging from real judgment) — see the
sourcing caveat in
../hitl-maturity-and-confidence-routing/SKILL.md's "Anchored in research" section, which applies here too. - Joshua Ebner (AI engineering & AI integration strategy practitioner) — the "AI Flywheel" framing: cumulative human feedback compounds into a competitive advantage that can't be replicated just by buying the same underlying model. He is a real, identifiable practitioner in this field; the specific quotes and framing attributed to him in the source material for this skill could not be independently verified word-for-word — treat the AI Flywheel concept as a reasonably well-grounded practitioner framing, and any specific quote as unconfirmed paraphrase.
- The general "data flywheel" mechanism already used elsewhere in this
repo (Collins, Jim — Good to Great, 2001, applied to data in
../../../data-strategy-and-literacy/skills/data-role-diagnosis/SKILL.md) — this skill applies the same self-reinforcing-loop logic specifically to human correction data rather than to data assets generally.
Method
- Track the override rate as an ongoing metric, not a launch-time check. Override rate = the share of AI outputs or recommendations that a human reviewer changes, rejects, or overrules. Calculate it on a rolling basis (e.g. weekly), broken down by task type — an aggregate number across dissimilar tasks hides which specific task type is actually the problem.
- Read the override rate against both failure directions, not just
one:
- Below roughly 5% — investigate before treating this as good news. It can mean the model is genuinely excellent, or it can mean reviewers have stopped exercising real judgment and are approving by default (automation bias). Check a random sample of "approved" decisions by hand against the actual right answer to tell the two apart — don't infer from the rate alone.
- Above roughly 30% — the model is very likely not yet fit for the
task at its current confidence threshold. Either the confidence-score
routing in
../hitl-maturity-and-confidence-routing/SKILL.mdneeds re-calibrating (more cases should escalate before this point), or the underlying model/prompt needs improvement before this task is a good automation candidate at all.
- Measure override accuracy, not just override rate. An override being frequent doesn't tell you whether the human's correction was actually better — spot-check a sample of overridden decisions against an independent ground truth or a senior reviewer's judgment. A high override rate paired with low override accuracy points to a confused or under-trained review process, not (only) a weak model.
- Design the feedback-capture mechanism before assuming corrections feed back into anything. A correction that lives only in the individual reviewer's head or an unstructured comment field doesn't compound. Structure what's captured: what was wrong, what the correct answer was, and (where possible) a category tag for the type of error — this structured record is the raw input to prompt refinement or fine-tuning, i.e. the AI Flywheel.
- Close the loop deliberately — schedule a recurring review (not just
an inbox of corrections) where structured feedback is actually
triaged into: a prompt/guardrail change (see
../ai-accuracy-guardrails-and-grounding-design/SKILL.md), a fine-tuning or retrieval-corpus update (an engineering task, out of this skill's scope), or a routing-threshold change (see../hitl-maturity-and-confidence-routing/SKILL.md). A flywheel that only accumulates feedback without ever acting on it isn't actually turning. - Feed the audit results into prioritization — for a product manager
deciding where to invest next, task types with the strongest AI
Flywheel effect (feedback clearly and repeatably improves the model)
and reasonably clean data deserve more investment than task types where
feedback doesn't seem to move the needle; use an ROI-vs-Difficulty view
similar to
../../../ai-strategy-and-governance/skills/ai-opportunity-portfolio/SKILL.md's scoring logic to make this trade-off explicit rather than intuitive. - Produce a structured audit output: override rate by task type and trend over time, override accuracy from the spot-check sample, flagged automation-bias or under-performance risks, and the specific actions the triage in step 5 produced.
- Re-run this audit on a fixed cadence, not only when something has already gone wrong — a healthy human-AI partnership at launch can drift in either direction (toward complacent rubber-stamping or toward reviewer distrust that quietly re-inflates cost) without a single dramatic failure to trigger a review.
What this skill does NOT do
- Doesn't design the routing rule itself — see
../hitl-maturity-and-confidence-routing/SKILL.md; this skill audits whether that design is actually working. - Doesn't implement the technical fine-tuning or retrieval-corpus update that structured feedback might justify — that's an engineering task; this skill produces the structured input for it.
- Doesn't confirm figures or performance data from memory — uses the
organization's own logged override data, or marks an assumption clearly
(
[assumption — verify]) if that data doesn't exist yet, in which case building the logging capability is the actual first step. - Doesn't replace
../../../ai-strategy-and-governance/skills/ai-output-curation-and-quality-control/SKILL.md's step 7 ("track quality over time") — this skill is the deeper, fuller audit methodology that step points to.
Refinement notes
Areas to keep deepening with real practice:
- your own calibrated override-rate bands by industry/task risk profile — the 5%/30% figures above are starting heuristics, not validated universal thresholds
- a concrete feedback-triage template (into
../../references/) - real before/after examples of a flywheel effect that measurably improved a model's usefulness over a defined period
Once this section is filled in and validated in practice, update the
maturity field in skills_index.json to draft, validated, or
canonical (see ../../../meta/maturity_levels.md). Don't add new
fields to the frontmatter — name and description are the only ones
allowed (see ../../../meta/frontmatter_schema.md).
Continue from here
- Before this in this pack:
../hitl-maturity-and-confidence-routing/SKILL.md,../ai-accuracy-guardrails-and-grounding-design/SKILL.md - Related skill in this pack:
../expert-agency-and-apprenticeship-protection/SKILL.md— a healthy override rate depends on reviewers who still have the expertise to override well; that skill protects the pipeline that produces them. - Related skill in another pack:
../../../ai-strategy-and-governance/skills/ai-output-curation-and-quality-control/SKILL.md— deepens that skill's step 7. - Related skill in another pack:
../../../ai-strategy-and-governance/skills/ai-opportunity-portfolio/SKILL.md— for prioritizing which task types deserve continued AI investment based on this audit's findings. - This pack's shared guardrails:
../../CLAUDE.md
References
../../references/hitl-partnership-heuristics-research.md— full sourcing and grounding-strength notes for this pack../../references/— the pack's shared background material../../CLAUDE.md— the pack's shared guardrails