Boundary-Sharpening Review Map¶
Monitoring dashboard — instantiates Neighbor-Suppression Contrast Sharpening
Lays the raw field beside the sharpened output, marks which neighbors were suppressed, and scores whether the sharpening actually improved detection.
Boundary-Sharpening Review Map is the evaluation surface for the whole archetype. It puts the raw field and the sharpened result side by side, highlights exactly which neighbors were suppressed to get from one to the other, and attaches a quality score answering the question the sharpening itself cannot: did suppressing those neighbors actually help? Its distinguishing feature is that it sharpens nothing and decides nothing — it is a mirror, not an actuator. Every other mechanism here produces or changes the field; this one only measures the field's before and after, and its most important output is the one nobody else surfaces: the cost side, where a valid neighbor was wrongly dimmed.
Example¶
A trust-and-safety team ships a change so that a high-confidence policy-violating post also down-ranks the borderline posts clustered around it. Did that sharpen the real signal or bury innocent content? The Review Map shows it. On the left, the baseline field — every post by risk score, before suppression. On the right, the sharpened readout — the same field after neighbor down-ranking, with the demoted posts marked. Beneath both sits a scorecard computed on a human-labeled sample: precision and recall, before versus after. The map reports that precision rose from ≈0.71 to ≈0.86, that recall held, but that ≈3 of the demoted posts in the sample were correctly-labeled benign — over-suppression, made visible. That last number is the whole point: without the map, the precision win ships and the three buried posts are never counted.
How it works¶
- Pair before and after. The raw field and the sharpened readout are shown together, so the effect of suppression is legible rather than assumed.
- Mark the suppressed. Every neighbor that was dimmed is flagged, turning "the output looks cleaner" into "here is precisely what was removed."
- Score against ground truth. A quality metric computed on labeled data reports the detection or decision delta — the map's claim is evidence, not aesthetics.
- Re-check the losers. A sample of suppressed items is re-examined for false negatives, which is how over-suppression becomes a visible number instead of an invisible cost.
Tuning parameters¶
- Quality metric — precision, recall, F-score, or a decision-accuracy measure. The choice encodes which error the team is willing to trade; a precision-only metric will happily hide over-suppression.
- Ground-truth source — held-out labels, adjudicated samples, or delayed outcomes. Cheaper sources score faster but can bake in the same bias the sharpening has.
- Suppressed-set sampling rate — how many demoted neighbors are re-checked. Higher sampling catches more wrongful suppression but costs review effort.
- Aggregation window — per-event versus rolling. Tight windows catch a regression fast; wide ones smooth out noise but lag.
- Over-suppression alarm — the false-negative rate among suppressed items that trips a warning. Set it low and you flag often; high and you miss slow erosion.
When it helps, and when it misleads¶
Its strength is that it converts "looks crisper" into checkable evidence and is the only place in the archetype where over-suppression — the valid neighbor wrongly erased — becomes a measured quantity. It is also what the tuning loop reads to adjust anything at all.
Its failure modes are those of any metric surface. Optimize the sharpening to maximize the map's own score and the score stops measuring what it was meant to — the trap that Goodhart's law names.[n1] It sees over-suppression only among the items it re-samples, so valid neighbors outside the sample stay invisible; and being post-hoc, it reports damage that has already shipped. The classic misuse is displaying only the flattering before/after pair to justify a tuning already chosen. The discipline that keeps it honest is to score on held-out labels, sample the suppressed set rather than only the survivors, and always report the loss column next to the win.
How it implements the components¶
Boundary-Sharpening Review Map fills the display-and-detection components — what a monitor can render and measure:
baseline_response_map— it renders the raw, pre-suppression field as the "before" against which everything is judged.sharpened_boundary_readout— it renders the post-suppression boundary as the "after," the object actually under review.over_suppression_monitor— the quality delta plus the re-check of suppressed items is the over-suppression detector.
It does not perform the suppression (neighbor_suppression_rule → Center-Surround Filter), move any dial in response to what it finds (suppression_gain_parameter, adaptive_neighborhood_resizer → Contrast-Adaptation Tuning Loop), or let a user restore hidden context in the moment (context_preservation_guardrail → Context-Restoration Toggle).
Related¶
- Instantiates: Neighbor-Suppression Contrast Sharpening — the map is the archetype's evaluation and over-suppression-detection surface.
- Consumes: Center-Surround Filter supplies the sharpened readout it displays; a labeled sample supplies the ground truth it scores against.
- Sibling mechanisms: Contrast-Adaptation Tuning Loop · Center-Surround Filter · Over-Suppression Red Team · Alert Neighbor Dimming Rule · Competitive Priority Board · Context-Restoration Toggle · Non-Maximum Suppression Pass · Local Contrast Enhancement Transform · Inhibitory Feedback Circuit · Mutual-Exclusion Choice Gate · Ecological Spacing Buffer
Editorial Notes¶
Form Classification¶
Form family: Assessment, Review & Assurance
Rationale: Lays the raw field beside the sharpened output, marks which neighbors were suppressed, and scores whether the sharpening actually improved detection, making its operative form a bounded evaluation of existing evidence or work that produces a finding or disposition.
Independent corroboration: The frozen evidence defines Boundary-Sharpening Review Map as 'Lays the raw field beside the sharpened output, marks which neighbors were suppressed, and scores whether the sharpening actually improved detection', so its operative form is Assessment, Review & Assurance.
Nearest alternative: Monitoring, Sensing & Alerting — It performs a bounded before/after assessment of a sharpened output rather than repeatedly observing changing state.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Computer Science & Software Engineering
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Specialized
Rationale: Computer science is primary because the mechanism's raw-versus-sharpened field, neighbor suppression trace, and false-negative review directly derive from image processing and computer-vision edge evaluation.
Related originating lineages:
- Art & Aesthetics — Visual contrast and edge legibility contribute perceptual criteria for displaying whether a boundary appears sharper.
- Data Science & Analytics — Labeled evaluation sets and before-versus-after model metrics make the sharpening effect empirically reviewable.
- Statistics & Experimental Design — Precision, recall, sampling, and false-negative estimation supply disciplined comparison against ground truth.
Review resolution: Canny's foundational computer-vision work formalizes edge detection around detection quality and localization, while NIST evaluation methodology formalizes precision, recall, and review against labeled ground truth. This mechanism combines that image-processing lineage with statistical error auditing and a visual review surface, making computer science primary and the artifact a synthesis.
Attribution caveat: The artifact also resembles a data-science model-evaluation dashboard, but the neighbor-suppression operation and boundary-sharpening vocabulary are specifically image-processing concepts.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Researched adjudication after independent review; high confidence.
Sources consulted:
Notes¶
The map is a passive monitor over normal traffic: it can only see the over-suppression that actually occurs in the data it happens to observe. Finding the over-suppression that would occur under adversarial or rare inputs is a different job, done by the sibling Over-Suppression Red Team, which constructs the hard cases the map would otherwise wait to encounter.
[n1] Goodhart's law — "when a measure becomes a target, it ceases to be a good measure," a caution associated with economist Charles Goodhart. A sharpening tuned to maximize the review map's own score can improve the number while quietly degrading the real detection the score was meant to stand in for. ↩