rkitoAgent False
Positive Rate
% of Lens findings on agent-authored PRs later dismissed by an architect — compared with the same rate for human PRs.
This page describes the intended design. It requires agent-authorship attribution to separate agent and human dismissal rates, and neither the attribution nor a dismissal comparison exists yet. It doesn't appear on your dashboard yet, and no org currently has a live value for it.
on agent-authored PRs
÷ total Lens findings
on agent-authored PRs)
× 100
AFPR has no fixed thresholds. It is read relative to the equivalent rate on human-authored PRs.
What it signals
AFPR shows whether Lens prompts over-fire on the structural patterns agents tend to produce. Agent code is typically more defensive, more verbose and more conditional than human code. A Lens written with human code in mind can flag those patterns even when the design intent is met.
A high AFPR on agent PRs does not mean the code is architecturally wrong. It may mean the Lens was authored around a specific human expression of a rule rather than the structural intent behind it.
That makes AFPR a Lens calibration metric. Prompts with a high AFPR on agent code need to be refined to detect the structural intent being evaluated, not the specific human expression of it.
How rkito produces it
AFPR is designed around two inputs: which PRs were agent-authored, and which Lens findings an architect later dismissed. Agent-authored PRs need to be identifiable from commit metadata — co-author signatures, agent labels, session IDs, or MCP access logs correlated with branch creation. Like the other agentic metrics, AFPR only gives a statistically valid signal once a meaningful fraction of PRs are agent-authored.
Who this metric is for
AFPR points to the specific Lenses whose findings on agent code keep getting dismissed. Those are the prompts to rewrite around structural intent rather than a human coding style.
If Lenses over-fire on agent code, agent conformance numbers look worse than the architecture really is. AFPR is the check before drawing conclusions from agent-vs-human comparisons.
Findings that are routinely dismissed teach teams to ignore Lenses. Keeping AFPR close to the human rate keeps CDA findings worth reading.
How this metric gets misused
Do not read a high AFPR as evidence that agent code is worse. AFPR counts findings an architect dismissed — by definition, those were not real violations. A high AFPR is a problem with the Lens, not with the agent.
Do not read AFPR on its own.
The number only means something next to the equivalent dismissal rate on human-authored PRs. A Lens that over-fires on everyone is a general calibration problem, not an agent-specific one.
What to do at each signal
Step-by-step response playbook for each signal state.
Lens prompts are working well for both human and agent code.
No calibration action needed. Monitor periodically as agent toolchain patterns evolve.
Not yet in rkito: agent PR attribution, which this comparison depends on. Until it exists, the agent and human rates cannot be separated.
Lenses are over-firing on agent code. Find which ones, and rewrite them around structural intent.
Navigate to /changecontrols/designaudits, filter by agent-authored PRs (once attribution is available), and identify which Lenses are dismissed most frequently. Open each at /steerings/lenses to recalibrate the prompt. Not yet in rkito: agent PR attribution and an automated AFPR comparison view.
Recalibrate over-firing Lens prompts via CHI at /changecontrols/changeintentions so they detect the underlying structural pattern rather than a human-specific expression of it — agent code is more defensive and verbose.
The agent-vs-human conformance gap. AFPR tells you how much of a negative AHCR may be Lens calibration rather than real drift.
Whether agents that loaded Lenses produced zero violations — only meaningful when those Lenses are calibrated for agent code.
High LPP with a high false positive rate points to redundant Lenses rather than genuine coverage.
Are your Lenses fair to agent code?
AFPR depends on agent attribution that rkito is still building. Connect your repo now to start recording the design audits and Lens findings it will be calculated from.