Skip to content

Counterfactual Filter-Relaxation Workshop

Facilitated workshop — instantiates Structural Filter Intersection Audit

A facilitated session that imagines one filter removed or loosened and reasons out what would then survive — surfacing outputs the intact stack silently kills.

The Counterfactual Filter-Relaxation Workshop works forward from a hypothetical: pick one filter, imagine it relaxed or removed, hold everything else constant, and reason out what new outputs would appear. It is not a measurement of the live stack — it is a constructed model of a counterfactual stack, built to expose each filter's marginal suppressive effect. The workshop format is the point: because the suppressed outputs by definition are not in any dataset, the only way to reach them is to gather people who know the domain and have them imagine what is missing — especially the outputs that are never even produced because a known filter deters them upstream.

Example

A philanthropic foundation notices that first-time, community-led applicants almost never get funded, and assumes their proposals are simply weaker. The workshop takes one filter — "requires a fiscal sponsor with three years of audited financials" — relaxes it, and asks program staff alongside outside community organizers: what would we now be reading? Together they reconstruct a plausible set of proposals — mutual-aid groups, immigrant-led cooperatives — that never appear in the current pool not because they were rejected but because the requirement is known in advance and deters them from applying at all. The counterfactual makes visible that the filter's largest effect sits upstream of any decision, in a candidate pool that quietly shrinks before review ever begins.

How it works

The distinguishing method is single-filter relaxation, everything-else-constant, carried out as structured imagination with people positioned to know the excluded outputs. Unlike any measurement of the live stack, it can model second-order effects — deterrence and chilling, not just rejection — because it reasons about outputs that were never submitted and so could never be counted. Each imagined output is pushed for concreteness, then, where possible, checked against a real system that lacks the filter.

Tuning parameters

  • Which filter to relax — the single most-suspected suppressor, or each in turn. One-at-a-time isolates marginal effect; all-at-once shows the ceiling but blurs attribution.
  • Relax vs. remove — loosen the threshold or delete the filter entirely. Removal shows maximum effect; loosening is more realistic.
  • Participant standpoint — insiders only, or deliberately include people from the excluded population who can name what is absent. Outsiders see more, but the session is harder to run.
  • Grounding discipline — how firmly to insist each imagined output be concretely plausible rather than speculative. Tighter grounding resists fantasy; looser surfaces more.

When it helps, and when it misleads

Its strength is unique: it reaches outputs that were never produced, because a known filter deters people before they try, and no measurement can see a deterred output. Its failure mode is that it is imagination — it can conjure a suppressed masterpiece that never actually existed and so overstate a filter's cost. The classic misuse is relaxing only the filters you already dislike, to manufacture a case for a change you had pre-decided. The discipline that guards against this is to relax the filters you are attached to as well, and to validate a counterfactual against a real holdout or a peer system without the filter wherever one exists.[1]

How it implements the components

  • counterfactual_filter_relaxation_model — the constructed picture of what would survive under a relaxed filter, including the deterred, never-submitted outputs that no live measurement can observe.

It does NOT model the ACTUAL surviving set (surviving_intersection_model — Intersection Matrix), inventory the filters (parallel_filter_inventory — Filter Stack Map), or enact any relaxation as standing policy (independence_and_diversity_controls — Filter Rotation or External Challenge). It supplies the what-if the others test or act on.

Notes

Because its product is imagined, the workshop is strongest when paired with a held-out or shadow stream (Shadow Review Board): the counterfactual proposes what a relaxed filter would surface, and the holdout — a real feed run without that filter — can partially confirm or deflate it. Treated as a standalone verdict rather than a hypothesis to test, the workshop's imagined losses are easy to inflate.

References

[1] A chilling effect is the deterrence of legitimate output by the mere prospect of a filter — people self-suppress before submitting, so the loss never appears as a rejection. It is why a filter's true cost can lie entirely upstream of the outputs it visibly blocks. ↩