# Prevented Work Is Invisible to Velocity Metrics ## Claim Velocity metrics count work that was produced, so labor that stops defects from existing registers as no contribution at all — not by miscalibration but by construction. <!-- atomicity: 1 idea ✓ --> ## Evidence **Source**: [[3. Resources/Readwise/Articles/AI Productivity Is Burning Out Your Best Engineers]] — Sreenivasa Reddy (AT&T) on the "invisible validator" > [!quote] > Invisible validator — an engineer whose primary contribution is catching/correcting AI-generated defects, work that leaves no trace in velocity metrics. The distribution is what makes it a structural problem rather than an accounting one. AI assistance makes juniors ship faster and lets seniors prototype quicker; both groups' gains land squarely in the metric. The cost of those gains — reviewing plausible-looking output, catching the security defect, correcting the subtly wrong abstraction — falls on mid-level engineers whose entire contribution is that a bug never entered the system. There is no artifact. No ticket closes, no commit lands, no line count moves. The team's throughput rises, the L4/L5 engineers work harder than they did before, and the dashboard reports them as the least productive people on it. That is the mechanism behind the attrition. It is not that the work is unpleasant; it is that the work is unobservable to every system that decides promotion, staffing, and sprint capacity. ## Scope & conditions - **Holds** for any metric defined over produced artifacts — story points, commits, PRs, tickets, lines, deploy frequency — against any labor whose output is a non-event. - **Sharpens** wherever generation gets cheaper without review getting cheaper, because the ratio of prevention labor to production labor rises while the measurement stays fixed. - **Breaks** where prevention produces its own artifact: a written test, a lint rule, a filed bug, a documented invariant. Converting prevention into an artifact is the whole fix. - **Weakens** in teams small enough that everyone can see what everyone does. It is a measurement problem, so it scales with distance between the work and the person evaluating it. - **Does not follow** that velocity metrics are useless — they measure what they measure. The error is treating them as a measure of contribution. ## Links - [[3. Resources/Permanent Notes/Automation Relocates Toil Into Supervision]] — because extends: this is what happens after the relocation. The supervision is real work, and the measurement system does not have a column for it. - [[3. Resources/Permanent Notes/The Bottleneck Moved From Code Generation to Review and Judgment]] — because extends: if review is now the binding constraint, then measuring generation is measuring the part that stopped being scarce. - [[3. Resources/Permanent Notes/Silent Failures Cost More Than Loud Ones]] — because extends: prevented defects are silent by definition, so the labor that prevents them inherits the detection problem of the thing it prevents. - [[3. Resources/Permanent Notes/Maintenance Is Most of the Work but Gets No Prestige]] — because extends: the same invisibility, one layer up. Maintenance at least leaves commits; prevention leaves nothing. - [[3. Resources/Permanent Notes/A Metric Is Vanity Only Against Your Own Outcome Data]] — because contrasts: that note says a metric earns meaning from outcome data. This one says some contributions produce no data at all, which outcome-linking cannot rescue. - [[3. Resources/Permanent Notes/Open Source Depends on Teaching Labor That Has No Formal Home]] — because extends: the open-source cognate. Teaching and reviewing are both load-bearing and both structurally uncounted. - [[3. Resources/Permanent Notes/The Reverse Centaur Pattern - Humans Forced to Work Like Machines]] — because applies-in: absorbing an uncredited cleanup queue generated by a machine is the reverse-centaur arrangement in its ordinary workplace form. - [[AI & the Future of Knowledge Work MOC]] — because applies-in. - [[3. Resources/Permanent Notes/A Reliable Cleanup Crew Removes the Reason to Stop Making the Mess]] — because extends: this is the organisational consequence of the invisibility. If only rescue registers, the reward structure funds a cleanup crew instead of the capability that would retire the work. - [[3. Resources/Permanent Notes/Decomposing a Capability Gap Into Tickets Makes It Closeable and Permanent]] — because contrasts: prevention produces no artifact to count, remediation produces too many, and both distortions push the same way — toward funding cleanup over capability. - [[3. Resources/Permanent Notes/A Capability Demonstration Is Not an Operating Model]] — because applies-in: the obligations a job holds between tasks are mostly non-events, so a substitution decision made on measurable output is made on the half of the role that registers. - [[3. Resources/MOCs/DevRel MOC]] — because the DevRel function's core output is largely prevented work: the migration that stayed uneventful, the question that stopped being asked, the defect that never shipped. - [[3. Resources/Permanent Notes/A Release Lead's Output Is Mostly Subtraction]] — because extends: the same invisibility one altitude up, where an unshipped feature leaves even less trace than a prevented defect. - [[3. Resources/Permanent Notes/A Count Stops Measuring Effort Once the Counted Thing Is Cheap]] — because contrasts: that note is the same broken-metric family from the other direction — the count still registers output but has stopped meaning anything because the output got cheap (a denominator collapse), where this note's failure is an omission (the work never registers at all). ## Implications - The practical fix is to make prevention leave an artifact — require that a caught defect becomes a test or a lint rule, and count *that*. It converts an invisible act into a durable one and pays twice. - Any team adopting AI assistance should expect its productivity gains and its burnout to land on different people, and should look at who is absorbing review load before reading the throughput number as good news. - This is a specific, mechanical instance of Goodhart: the metric did not get gamed, it simply never had visibility into the contribution that now matters most. - For open-source maintainers the same argument applies to triage and plugin review — the WordPress.org review queue is prevention labor measured, if at all, by volume rather than by what it stopped. ## Open Questions - Can prevention be measured directly — escaped-defect rate attributed to reviewers, rather than caught-defect counts that invite gaming? - Does the burden actually concentrate at L4/L5, or is that an artifact of who was surveyed? The claim's force depends on the concentration being real. - If prevention is converted into artifacts wholesale, does the artifact count become the new target and reintroduce the same problem one level up? ## Source - [[3. Resources/Readwise/Articles/AI Productivity Is Burning Out Your Best Engineers]] — the invisible-validator framing and the distribution of gains versus costs across levels