Bycatch Rate Dashboard¶
Monitor — instantiates Bycatch-Aware Selective Intervention Design
A live scoreboard that keeps off-target capture and its cumulative harm on screen next to target yield, broken out by non-target class, so bycatch can't hide behind a good headline number.
A selective process usually reports one number — how much target it caught — and that number can climb while the collateral damage climbs faster. Bycatch Rate Dashboard is the visibility instrument that refuses to let that happen: it maintains a running ledger of off-target captures and their accumulating harm, and it displays that ledger in the same frame as the target yield it was bought with. Its whole job is standing — to keep bycatch continuously on-screen, segmented by which non-target class is absorbing the damage. It does not decide anything and it does not fix anything; unlike the Bycatch Tolerance Stop Rule, which acts on a breach, the dashboard only makes the breach impossible to overlook.
Example¶
A platform's trust-and-safety team ships automated removals of spam and abuse accounts. The executive dashboard shows one triumphant line — accounts actioned, up and to the right. The Bycatch Rate Dashboard adds the panel nobody was looking at: legitimate accounts caught in the sweep, measured by appeals upheld and reinstatements, broken out by user segment. When a new classifier goes live, the target line jumps — and so does a second line the team can now see, wrongful suspensions of small businesses posting in a non-English language, up roughly threefold.
The setup is deliberately dull: pipe the removal log and the appeals-overturned log into one view, key both to a shared list of user segments, refresh weekly. The outcome is not a verdict but a fact that can no longer be avoided — the win and its collateral cost sit on the same screen, in front of the same people, at the same time. That is the entire value: the harm is now argued about instead of unmentioned.
How it works¶
- Two lines, one frame. The distinguishing move is juxtaposition — target yield and bycatch harm are rendered together, defeating the tunnel vision that a single-metric report creates.
- Segmented by class. The ledger is sliced by non-target class, so a harm concentrated in one vulnerable segment doesn't average away into a reassuring aggregate.
- Rate and accumulation. It tracks both the current capture rate and the cumulative harm to date, because a low rate over a long time can still add up to a large debt.
- Display, not detection. It charts signals produced elsewhere (sampling, audits); it is the pane of glass, not the sensor behind it.
Tuning parameters¶
- Segmentation granularity — how finely bycatch is broken down by class. Finer slicing surfaces concentrated harm but multiplies noise and invites cherry-picked framings.
- Refresh latency — near-real-time versus periodic. Faster feedback catches a regression early; slower, larger samples read more stably.
- Leading vs. lagging signals — cheap proxies (appeal volume) shown immediately versus confirmed harm shown late. Proxies react fast but can mislead.
- Harm weighting — raw capture counts versus severity-weighted harm. Weighting stops a flood of trivial errors from drowning out a few grave ones, but the weights are a judgement call.
- Coverage disclosure — whether the view shows how much of the non-target space is actually instrumented. Hiding coverage lets an unsampled class read as a comforting zero.
When it helps, and when it misleads¶
Its strength is that it dissolves metric tunnel vision: once collateral harm shares the screen with the win, "we're crushing it" has to survive contact with "and here's who's paying for it." It is the shared factual surface every other mechanism in the archetype reasons from.
Its danger is that a dashboard measures only what is instrumented, so an un-sampled non-target class shows up as zero and absence of evidence masquerades as evidence of absence. The classic misuse is building or slicing it after the fact to reassure — choosing the segmentation and the harm weights that make bycatch look negligible, which is exactly the failure mode Goodhart's law warns of: once the on-screen number becomes the thing being managed, it stops measuring the thing that matters.[1] The discipline that guards against this is to fix the segmentation and weighting in advance, publish coverage honestly, and validate the ledger against an False-Capture Audit and sentinel sampling rather than trusting the picture the operator draws of itself.
How it implements the components¶
bycatch_rate_and_harm_ledger— the running count of off-target captures and their cumulative, severity-weighted harm is the dashboard's backing data; maintaining and rendering that ledger is what it does.non_target_class_register— it keeps the per-class breakdown live and current, the register in its operational, at-a-glance form; the prospective definition of those classes is authored by the Non-Target Impact Pre-Mortem, and the dashboard tracks them.
It does not set the tolerance line — that's the Bycatch Tolerance Stop Rule; it does not reweight the scorecard (Success Metric Reweighting); and it does not generate the detections it charts (Non-Target Sentinel Sampling, False-Capture Audit).
Related¶
- Instantiates: Bycatch-Aware Selective Intervention Design — the dashboard is the shared visibility surface the appraisal reads from.
- Consumes: Non-Target Sentinel Sampling and False-Capture Audit supply the detections it aggregates.
- Sibling mechanisms: Bycatch Tolerance Stop Rule · Non-Target Sentinel Sampling · False-Capture Audit · Success Metric Reweighting · Non-Target Impact Pre-Mortem · Compensation and Restoration Trigger · Escape Hatch or Release Protocol · Negative Filter or Exclusion Device · Selectivity Window Test · Selector Retuning Cycle
Editorial Notes¶
Form Classification¶
Form family: Monitoring, Sensing & Alerting
Rationale: A live scoreboard that keeps off-target capture and its cumulative harm on screen next to target yield, broken out by non-target class, so bycatch can't hide behind a good headline number, making its operative form an ongoing sensing arrangement that repeatedly observes state and surfaces changes or alerts.
Independent corroboration: The frozen evidence defines Bycatch Rate Dashboard as 'A live scoreboard that keeps off-target capture and its cumulative harm on screen next to target yield, broken out by non-target class, so bycatch can't hide behind a good headline number', so its operative form is Monitoring, Sensing & Alerting.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Marine Science & Oceanography
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Specialized
Rationale: Its governing quantity and class segmentation come from fisheries bycatch monitoring, where non-target capture must be reported beside target catch.
Related originating lineages:
- Data Science & Analytics — Operational dashboard practice supplies continuous display, slicing, cumulative harm, and target-versus-bycatch comparison.
Review resolution: Marine science is the agreed primary lineage because target-versus-bycatch rate and non-target-class harm are fisheries measures. Data science contributes the live segmented dashboard form; cross-domain examples describe applicability, not additional origins, so origin reach remains specialized.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Sources consulted:
Notes¶
Visibility is not control. A dashboard that draws a threshold line is often mistaken for the thing that acts on it — but the line belongs to the Bycatch Tolerance Stop Rule, and a green dashboard is no substitute for an independent audit of the classes it may not be sampling. Keep the pane of glass and the decision separate, or the picture starts standing in for the reality it was meant to reveal.
References¶
[1] Strathern, Marilyn. "‘Improving ratings’: audit in the British University system". European Review 5(3): 305-321, 1997. Explains that managing to a measure as a target degrades its value as a measure of the underlying performance. registry ↩