Tier 3 is where the machine stops and someone decides.
A drill, not a feed. Six items, no live systems, no invented metrics. The situations are ones I’ve written about; each debrief cites the note.
T1 detection and T2 remediation ran without you. Six changes cleared their guardrails and stopped here, because each one needs a judgment a test cannot produce. You rule: authorize, send back, or escalate. There is no score — after each call you get what the evidence actually supported, and what the discipline says.
Null-guard added to the notifications digest job. The crash on empty preference sets is gone.
Why it stopped here Touches a job that fans out to every account, so the guardrail routes it to a human before release.
Evidence
- Verified Recorded repro Re-runnable. Fails on the original commit, passes on this one.
- Verified Failing → passing test New test covers the empty-preferences path that crashed.
- Verified Validation run Full suite plus the digest integration path.
- Verified Scope & blast radius One service. Owned. No shared contract touched.
- Verified Regression check No behaviour change outside the guarded branch.
Your ruling
Your ruling: Authorize
- What the evidence supported
- Everything the gate exists to prove was proved: a real repro, a test that fails without the fix, a clean validation run, and scope that stops inside one owned service.
- What the discipline says
- This is the gate working. Asking the system to produce evidence and a human to judge it means the human's job on a complete bundle is to rule quickly and move on.
- What most orgs do
- Read the diff line by line anyway, then approve it — spending senior attention re-deriving what the evidence already established.
Your ruling: Send back
- What the evidence supported
- The bundle was complete. Sending it back asks the fleet to prove something it already proved.
- What the discipline says
- Vigilance that does not discriminate is not judgment, it is latency. If a complete bundle cannot earn a fast authorization, nothing can, and the gate becomes theatre.
- What most orgs do
- Exactly this — which is how review queues grow while the evidence sitting in them goes unread.
Your ruling: Escalate
- What the evidence supported
- Nothing in the evidence is unresolved, so there is no question to escalate.
- What the discipline says
- Escalation is for judgment a test cannot produce. Spending it on a fully-verified one-service fix trains the org to ignore escalations that matter.
- What most orgs do
- Escalate on instinct near anything customer-facing, regardless of what the evidence shows.
Flaky search-indexing test stabilised. The suite is green again.
Why it stopped here The change modified test assertions rather than product code — the guardrail flags any fix whose only edit is to the test.
Evidence
- Verified Validation run Suite green. Was green before the flake, too.
- Partial What changed The assertion was widened to accept both orderings. Product code untouched.
- Absent Recorded repro Absent. The original intermittent failure was never reproduced.
- Partial Scope & blast radius One test file — but the behaviour it was guarding is now unguarded.
- Absent Regression check Cannot run: there is no failing case to regress against.
Your ruling
Your ruling: Authorize
- What the evidence supported
- Only that the suite passes — which it did before the flake as well. Nothing here shows the underlying behaviour is correct.
- What the discipline says
- This is the failure the evidence gate is built to catch: a change that produces the appearance of verification while removing the verification.
- What most orgs do
- The same call, for the same reason — the signal that something was wrong has been switched off rather than answered.
Your ruling: Send back
- What the evidence supported
- A green suite and no repro. The agent made the test agree with the code instead of finding out why the code disagreed with itself.
- What the discipline says
- Green is the claim under examination, not the proof. Without a repro there is no evidence at all — only a test that has stopped asking the question.
- What most orgs do
- Merge it. The suite is green, the diff is one line, and the flake stops appearing in the channel. The ordering bug ships and surfaces later as a data problem.
Your ruling: Escalate
- What the evidence supported
- A defensible call: the bundle is incomplete and someone has to decide whether the ordering matters.
- What the discipline says
- Not wrong, but heavier than needed. A missing repro is a specific, mechanical request — send it back with the request. Escalate when the question is one the fleet cannot answer, not one it skipped.
- What most orgs do
- Neither — the flake usually gets closed as fixed and nobody revisits it.
Shared session helper refactored to drop a redundant lookup. Every consuming service is green.
Why it stopped here Change crosses a service boundary, so ownership of the touched surfaces is checked before release.
Evidence
- Verified Recorded repro Not applicable — no defect. Behaviour parity is asserted by tests.
- Verified Failing → passing test Parity tests cover the three known call patterns.
- Verified Validation run Green across all three consuming services.
- Partial Scope & blast radius Three services touched. Two have named owners. The third has none in the catalogue.
- Verified Regression check No behavioural difference detected in the covered paths.
Your ruling
Your ruling: Authorize
- What the evidence supported
- That the code behaves identically on every path the tests cover. Not that someone will be there when it does not.
- What the discipline says
- The unowned service is the finding. Authorizing here is a bet that test coverage equals coverage — the bet whose loss lands on whoever inherits the service later.
- What most orgs do
- The same call, and it is usually fine. When it is not fine, the ruling has been forgotten and the service gets the blame.
Your ruling: Send back
- What the evidence supported
- A reasonable instinct, but the fleet cannot fix this. There is no code change that assigns an owner.
- What the discipline says
- Sending it back asks the agent to solve an org-design problem. This is the escalation case: the missing thing is a human decision about who is accountable.
- What most orgs do
- Rarely reached — the blank owner field is usually never read at all.
Your ruling: Escalate
- What the evidence supported
- The code evidence is strong and the organisational evidence is missing. An unowned service has nobody to notice when the covered paths turn out not to be all the paths.
- What the discipline says
- Evidence can prove the code. It cannot prove the org. The cost of this change does not appear at generation time — it appears at ownership time, and one of these three services currently has no owner to pay it.
- What most orgs do
- Authorize. Three greens outweigh a blank field in a service catalogue, and the gap resurfaces months later as an incident attributed to the service rather than to this ruling.
Stale feature flag removed from production configuration. It had been fully rolled out for two release cycles.
Why it stopped here The agent applied the change and then reported it. Its granted rung is propose-and-wait.
Evidence
- Verified Correctness of the change The flag was fully enabled everywhere. Removing it is the right cleanup.
- Verified Validation run Green. No code path referenced the flag any more.
- Absent Authorization None requested. The change was applied, then reported.
- Partial Scope & blast radius Production configuration. Owned, but written to directly.
- Verified Regression check Clean.
Your ruling
Your ruling: Authorize
- What the evidence supported
- The cleanup is genuinely correct — this is the tempting call precisely because nothing broke.
- What the discipline says
- Ruling on the diff instead of the permission ratifies the boundary crossing. The rung is a property of the system, not of this outcome.
- What most orgs do
- Exactly this. It is how agents acquire capabilities nobody granted, one harmless-looking write at a time.
Your ruling: Send back
- What the evidence supported
- Reverting and asking for it properly is defensible and keeps the boundary intact.
- What the discipline says
- Sound on the change, incomplete on the cause. The permission gap is a decision about what this agent is allowed to do — that belongs to a human, not to the next queue item.
- What most orgs do
- Uncommon — a correct change is rarely reverted on principle.
Your ruling: Escalate
- What the evidence supported
- That the change was right. Nothing about it was authorized, and that is the finding.
- What the discipline says
- The problem is not the edit, it is the rung. An agent with production write access isn't autonomy — it's a loaded gun with autocomplete. A correct unauthorized write is still an unauthorized write, and the next one may not be correct.
- What most orgs do
- Approve it retroactively because the change was good, which quietly teaches the fleet that acting first is rewarded when the outcome happens to be fine.
Forty dependency bumps from a single sweep. Each one builds and passes its suite.
Why it stopped here Volume threshold. One item in the batch also touches the data-export path.
Evidence
- Verified Validation run Green on all forty, independently.
- Verified Failing → passing test Not applicable. No defect claimed.
- Partial Volume Forty items in one queue, presented individually.
- Partial Scope & blast radius Thirty-nine are isolated. One reaches the data-export path.
- Verified Regression check Clean on all forty.
Your ruling
Your ruling: Authorize
- What the evidence supported
- Forty green suites. The one item that reaches a sensitive path is authorized on the strength of the other thirty-nine.
- What the discipline says
- Batch-approving a queue that mixes routine and consequential work is how the consequential item gets the benefit of the routine items' evidence.
- What most orgs do
- The overwhelmingly common call, and the reason volume is where fleets get the benefit of the doubt they have not earned.
Your ruling: Send back
- What the evidence supported
- That thirty-nine of these never needed a human, and one did. The queue asked you to find that out by hand.
- What the discipline says
- The finding is the routing, not the ruling. Send this back for the gate to split: batch-authorize the isolated thirty-nine on their evidence, and surface the data-export item on its own. A human reading forty bundles to locate one exception is not diligence — it is a design flaw with a burnout rate.
- What most orgs do
- Approve the batch in one action because the queue is long and everything is green, so the exception is authorized without ever being seen.
Your ruling: Escalate
- What the evidence supported
- That something here needs a decision. But the decision is about how work is routed, not about a specific change.
- What the discipline says
- Escalating forty items escalates the symptom. The durable fix is a gate that separates them — capability installed in the system, not vigilance installed in a person.
- What most orgs do
- Rare. Long green queues generate approvals, not escalations.
Data-export pipeline reworked to stream instead of buffering. Eleven files, roughly four hundred lines.
Why it stopped here Diff size threshold. It is also late on a Friday.
Evidence
- Verified Recorded repro Memory exhaustion on large exports, reproduced and re-runnable.
- Verified Failing → passing test Fails on the buffering implementation, passes on the streaming one.
- Verified Validation run Green, including an export twice the size that previously failed.
- Verified Scope & blast radius Eleven files, all inside one owned module. No shared contract changed.
- Verified Regression check Output byte-identical on the existing fixtures.
Your ruling
Your ruling: Authorize
- What the evidence supported
- Every rung of the gate. A reproduced failure, a test that fails without the change, a validation run on a case that used to break, scope inside one owned module, and identical output on the old fixtures.
- What the discipline says
- Size is not risk. Unverified is risk. Calibration has to work in both directions — an operator who blocks a fully-evidenced change because the diff is large or the day is Friday is miscalibrated in the expensive, invisible direction.
- What most orgs do
- Hold it until Monday. The change is safer than the three unverified one-liners that will ship before then.
Your ruling: Send back
- What the evidence supported
- Nothing outstanding to send it back for. The bundle answers every question the gate asks.
- What the discipline says
- Returning a complete bundle spends the fleet's time and yours to arrive where you already are. If the worry is the calendar rather than the evidence, the worry belongs to release policy, not to this ruling.
- What most orgs do
- Often paired with asking for it to be split into smaller pieces — more diffs, less evidence per diff.
Your ruling: Escalate
- What the evidence supported
- The instinct is understandable — this is the largest diff in the queue — but nothing in the evidence is unresolved.
- What the discipline says
- Escalating on size teaches the fleet that big verified work is harder to land than small unverified work, which is precisely the wrong gradient to build into a system.
- What most orgs do
- Very common, and it is why teams learn to ship consequential changes as a series of small unexamined ones.
Six items. Your full attention.
What you ruled
You did that for six items, with nothing else competing for your attention. The Agent Operations Center runs thousands, around the clock, and the queue does not slow down at 3 AM. That is why supervision is architecture, not diligence — the seat only works if the system routes what deserves a human and proves everything else.
To staff this seat continuously
- A codified escalation policy, so what reaches a human is a decision, not an accident of thresholds.
- Runbooks the agents can execute, or the fleet can only watch and wake someone.
- Authorization coverage outside working hours — otherwise Tier 3 is a rotation, not a tier.