Beyond Ninety Days

33.1 Months Four Through Six: Operate and Instrument #

The first ninety days close the merge boundary and prove one oracle. This quarter is where the platform work that makes the second, third, and fourth work classes cheap actually happens.

Replace standing credentials with ephemeral issuance. Task-scoped, audience-scoped, carrying an actor chain that survives the hop. Long-lived tokens on developer machines are what turned a compromised dependency into a thousand-repository credential leak in Section 40, and the remediation the affected vendor shipped was exactly this substitution.

Move architectural constraints out of documentation and into blocking checks, including a duplication threshold. House rules that live in a style guide govern nobody; house rules expressed as machine-checkable lint scale to any volume. This is also the cheapest available answer to the failure mode in Section 35, where roughly a fifth of rejected fixes were rejected for not using the organization’s own helper APIs—conventions the agent had no way to learn.

Pilot diff-scoped mutation testing on agent-authored modules. Manage the unproductive-mutant rate before gating on it. Section 37 is the evidence: mutation-guided generation killed mutants at roughly six times the rate of its coverage-guided predecessor, and half its useful tests added no line coverage at all.

Add a held-out verification signal for every work class above tier A2, and begin trending the gap between visible-suite and held-out-suite results. Almost no organization computes this, and it is the only early indicator of optimization toward the visible check.

Restrict network access and git history during acceptance and qualification, so that acceptance measures engineering rather than retrieval.

Emit and verify provenance. Signed artifacts, release manifests that distinguish agent-authored components, SBOMs covering AI components. Verification at consumption, not merely production.

Route model and tool calls through governed gateways with per-team quotas and cost attributed per agent and per task. Cost attribution is the item most often deferred and the one that is hardest to add later, because it requires the call path to have been instrumented from the beginning.

Test separation of duties by attempting to violate it, quarterly, and record the result as evidence rather than as an assertion.

Test kill switches with a measured time-to-effect, including at least one agent running inside a third-party platform where the switch is somebody else’s to throw.

Begin the seeded-defect probe on review and measure the catch rate. No published study measures human review effectiveness against agent-authored code, so this number will not come from the literature and has to be produced internally.

33.2 Months Seven Through Eighteen: Qualify for Stage 3 #

Two quarters of trended operating evidence per work class. Merge rate, revert rate, defect escape, review cost, conflict rate, and incident involvement. This is the requirement that cannot be compressed, and attempting to compress it is the most common way organizations arrive at Stage 3 by accretion rather than by decision.

Continuous validation against production traffic, alerting on distribution shift rather than on error rate alone.

An auto-merge policy defined by change class, with a measured false-negative rate. Not a belief about which changes are safe. Section 38 shows that a temporal control—cohort gating, working-hours constraints, and post-deployment failure correlation—can substitute for per-change verification at sufficient scale and sufficiently low blast radius per change. It is a legitimate design and it is not the same thing as verification. An organization adopting it should say which one it is buying, and should publish internally the failure rate that the case study’s source does not.

Fleet-level identity, cost attribution, and observability, with the generated-to-analyzed telemetry ratio instrumented. Telemetry that is emitted and never read is a cost center wearing the costume of a control.

Independent assurance against this framework’s evidence requirements rather than self-assessment. The distinction that matters is whether someone outside the delivery organization can reproduce the evidence.

Funded consolidation work driven by measured duplication and reuse trends. Generation is elastic and consolidation is not; without a standing budget for it, duplication accumulates at the rate agents produce it.

Deliberate management of the junior pipeline and of comprehension-preserving usage patterns, with the instrumentation to know whether it is working. Nobody has published that answer, so the organization will have to produce its own.

33.3 What This Period Does Not Buy #

It does not buy a defect-escape rate that can be compared against anyone else’s, because no organization publishes one—a gap Section 42.6 names directly. Internal trend is the only comparison available, which is a further reason the baselines from Section 32.2 matter more than they appear to at the time.

It does not buy a cost model with external reference points. Of the eight cases in Part VI, one explicitly excludes compute and continuous-integration cost from its analysis, one reports wall-clock and no monetary figure, and the rest report nothing at all.

And it does not buy Stage 3 by elapsed time. Two quarters of evidence is a floor, not a schedule.

33.4 Failure Signatures #

• Evidence assembled after the decision. Metrics reconstructed to justify a stage change already made are the clearest sign that the progression is running backward.
• Platform work deferred in favor of more work classes. A fifth work class standing on a thin substrate performs like the first four, and for the same reason. Adding classes is visible progress; deepening the substrate underneath them is not, which is why the second is the one that gets postponed.
• Instrumentation that nobody reads. The generated-to-analyzed ratio applies to an organization’s own telemetry as much as to its agents’ output.
• Consolidation permanently deprioritized. Duplication trends are slow, visible, and easy to defer for four consecutive quarters.

33.5 What This Rests On #

The credential, provenance, and gateway items derive from documented failure modes in Part VI and from the control requirements in Sections 17 and 18. The mutation-testing and held-out-signal items rest on measured findings.^113,114,170 The two-quarter evidence requirement, the month boundaries, and the ordering within each period are this framework’s construction. No organization has published a staged qualification timeline with outcome data against it, and the temporal-versus-logical control substitution in Section 33.2 is named as an open problem in Section 23 precisely because it is unmodeled.