Skip to content

Intersection Matrix

Analytic tool — instantiates Structural Filter Intersection Audit

Cross-tabulates candidate outputs against every parallel filter at once, exposing the joint survivor set and the exclusions that several filters overdetermine together.

An Intersection Matrix is a grid: rows are the candidate outputs (or output classes), columns are the parallel filters, and each cell records whether that output would pass, soften, delay, or fail under that filter's reconstructed rule. Reading across a row shows an output's joint fate; reading down a column shows a single filter's selectivity. Its defining move is to compute the intersection explicitly — an output reaches the surface only if it passes every column — which surfaces the finding no single-filter analysis can: overdetermined exclusions, the outputs killed by several filters at once. That is exactly why removing any one filter changes the visible surface so little, and it is the difference between this tool and a plain catalogue of filters.

Example

A media-criticism team wants to understand why in-depth coverage of a major advertiser's labor practices almost never airs on a cable network. They build a matrix: rows are candidate story types, columns are the filters — parent-company interest, advertiser sensitivity, availability of official sourcing, defamation risk, produceability on a 24-hour cycle, audience-attention pull. Marking each candidate's fate per filter, the advertiser-labor story shows three independent "fail" cells: advertiser sensitivity, legal risk, and thin official sourcing. The matrix makes the point unmistakable — the exclusion is overdetermined. Even if the network stopped deferring to advertisers tomorrow, the story still dies on sourcing and legal. That single reading redirects the reform from "lean on advertisers" toward sourcing and legal backing, and it flags which story classes sit at a fragile single-filter margin — the ones worth protecting — versus a hopeless multi-filter margin.

How it works

  • Rows are candidate outputs; columns are the filters drawn from the filter inventory.
  • Each cell applies a filter's reconstructed pass/fail rule to that output — not whether it did pass, but whether it would.
  • Survivors are the logical AND across a row; the matrix makes that conjunction visible instead of assumed.
  • Counting fail-cells per excluded output grades overdetermination; counting survivors per column grades each filter's selectivity.

Tuning parameters

  • Row granularity — individual items versus output classes; fine rows catch specific casualties, coarse rows reveal structural pattern.
  • Cell resolution — binary pass/fail versus graded pass / soften / delay / downrank; graded cells capture transformation, not just outright rejection, at the cost of harder scoring.
  • Filter-set boundary — how many columns you admit; too few misses the real intersection, too many makes everything fail something and manufactures an over-filtering artifact.
  • Criteria source — whose reconstruction fills the cells (owner-stated, producer-inferred, or observed rule); disagreement among these sources is itself a finding.
  • Counterfactual toggle — recomputing survivors with one column removed, to test whether an exclusion is overdetermined or hangs on a single decisive filter.

When it helps, and when it misleads

Its strength is converting a vague "the coverage feels narrow" into a specific claim about which outputs die and why, and it uniquely exposes overdetermination — the reason single-filter reforms so often disappoint — while localizing the fragile margins where one filter is decisive and reform has the most leverage.

Its cells are reconstructions, so a matrix is only as honest as its pass/fail criteria; confident-looking cells can quietly encode guesswork, lending false precision to the whole grid. It is easily run backwards — rows and columns chosen to guarantee a predetermined "they're suppressing X" pattern. And a map of survivors-versus-filters says nothing about whether each filter is legitimate: a perfectly real intersection can be entirely proper on grounds of quality, law, or privacy. The propaganda model of Herman and Chomsky proposed exactly this kind of multi-filter intersection for the news, and it is the archetypal use and misuse of the idea.[1] The discipline is to fill cells from reconstructed criteria plus rejected-item evidence, mark the uncertain ones, and hand the legitimacy question to a separate standard.

How it implements the components

  • surviving_intersection_model — the matrix is this model: the explicit conjunction of filters that defines the survivor set and grades how overdetermined each exclusion is.
  • filter_pass_fail_criteria — every cell operationalizes a filter's reconstructed pass/fail rule against a real output; the matrix is where those criteria stop being abstract and get applied.

It applies the filters but does not inventory them — parallel_filter_inventory and filter_owner_and_incentive_map are Filter Stack Map's — nor supply the candidate_output_universe and rejected_or_transformed_output_sample it charts (those come from Rejected-Item Sampling), nor judge whether any filter is legitimate (filter_legitimacy_standard, Filter Rationale Register).

Notes

The matrix models whether an output survives, never whether the filtering is right — pairing it with a legitimacy standard is what keeps it from sliding into conspiracy theory. Its signature contribution is overdetermination: the thing a single-gate analysis structurally cannot see, and the reason so many well-meaning single-filter reforms leave the surface unchanged.

References

[1] The propaganda model of Edward Herman and Noam Chomsky (Manufacturing Consent, 1988) argued that mass-media content passes through a series of filters — concentrated ownership, advertising revenue, reliance on official and corporate sources, organized "flak," and a guiding ideology — whose joint operation predicts which news survives without any explicit censor. It is the archetypal intersection model; used well it is a hypothesis to test against rejected-item evidence, not a verdict to assume. ↩