session_id: "EXAMPLE-SOCIAL-COMMUNITY-TOXICITY" created_at: "2026-01-01T00:00:00Z" updated_at: "2026-01-01T00:00:00Z" version: 1 contributors:

• agent: "tutorial" program: "brennerbot.org" model: "static" contributed_at: "2026-01-01T00:00:00Z" status: "active"

Brenner Protocol Artifact: EXAMPLE-SOCIAL-COMMUNITY-TOXICITY

1. Research Thread

RT: Why do some online communities become toxic over time?

Context: Some communities devolve into harassment, norm violations, and chronic conflict, while others remain constructive. We want to discriminate between moderation/incentive explanations, selection/composition explanations, and platform-amplification explanations under realistic constraints (observational data, quasi-experiments).

Why it matters: If we can identify the dominant causal lever (moderation, selection, or amplification), we can design interventions that actually change trajectories instead of moralizing after the fact.

Anchors: §89, §161, §147, domain

2. Hypothesis Slate

H1: Moderation capacity / enforcement failure

Claim: Communities become toxic when moderation capacity and enforcement lag behind growth and conflict; norms drift because violations go unpunished. Mechanism: Low enforcement increases payoff for antisocial behavior; norm entrepreneurs set a new equilibrium; constructive members churn out, accelerating drift. Anchors: §89

H2: Selection effects / composition shift

Claim: Communities become toxic primarily because they attract and retain toxic participants; the main driver is who joins and stays, not within-community change. Mechanism: A small early toxicity signal changes who enters; homophily and churn shift composition; toxicity appears ‘caused’ by the community but is selection-driven. Anchors: §161

H3: Algorithmic amplification / incentive gradient (third alternative)

Claim: Platform ranking and engagement incentives amplify outrage and conflict; toxicity is an emergent effect of attention allocation rather than moderation or selection alone. Mechanism: Ranking models optimize engagement; inflammatory content gets surfaced; feedback loops raise conflict and normalize toxic styles even among initially neutral users. Anchors: inference, §147 Third alternative: true

3. Predictions Table

4. Discriminative Tests

T1: Natural experiment: enforcement shock (Score: 11/12)

Procedure: Identify a policy or staffing change that sharply increases enforcement (or a tooling rollout). Compare toxicity metrics pre/post within the community, with matched control communities that did not change enforcement at the same time. Discriminates: H1 vs H2 Expected outcomes:

• H1: Post-change toxicity declines relative to controls, even with similar membership composition.
• H2: No meaningful change unless membership composition shifts; controls look similar after adjusting for entrants. Potency check: Validate that enforcement actually changed (ban/removal rates, response times). Check for concurrent external shocks. Feasibility: medium Evidence-per-week score: LR=4, Cost=2, Speed=2, Ambiguity=3

T2: Amplification test: ranking perturbation (A/B or policy proxy) (Score: 10/12)

Procedure: Use a known ranking change, feed ranking, or visibility intervention (A/B if possible; otherwise proxy via platform rollouts). Measure whether toxic content visibility predicts downstream toxicity even controlling for composition and enforcement. Discriminates: H2 vs H3 Expected outcomes:

• H2: Once composition is controlled, ranking perturbations have limited effect on toxicity trajectories.
• H3: Visibility/engagement shifts causally change toxicity (conflict escalates when outrage is surfaced). Potency check: Ensure the ranking perturbation is plausibly exogenous (rollout timing, geography) and that measurement captures exposure, not just production. Feasibility: low Evidence-per-week score: LR=3, Cost=3, Speed=1, Ambiguity=3

5. Assumption Ledger

A1: Toxicity metric tracks the construct

Statement: Our toxicity metric (e.g., classifier, rule-violation rate, harassment reports) reflects meaningful harm rather than mere disagreement or topical intensity. Load: Load-bearing for all hypotheses Test: Audit samples with humans; measure false positives (heated but constructive) and false negatives (subtle harassment).

A2: Controls address confounding shocks

Statement: We can find comparison communities or within-community controls that share external shocks (news cycles, political events) so we don’t mistake global events for causal mechanisms. Load: Load-bearing for causal claims Test: Matched controls + interrupted time series; sensitivity analysis to unmeasured confounding.

A3: Scale check: individual vs group vs platform level

Statement: We are not mixing causal levels (individual traits vs group norms vs platform incentives). Each hypothesis must specify its operative level and how effects propagate across levels. Load: Load-bearing for coherent mechanism Test: Explicit level-split in the model; include measures at each level (user churn, norms, ranking exposure). Scale check: true

6. Anomaly Register

None registered.

7. Adversarial Critique

C1: Are we mistaking visibility for prevalence?

Attack: Perceived toxicity may rise because toxic content becomes more visible, not because more toxic behavior is produced. Measurement and exposure are intertwined. Evidence: Ranking systems can increase exposure of a minority of toxic posts; users experience a ‘toxic feed’ even if production is unchanged. Current status: Separate production metrics from exposure metrics. Add instrumentation for what users actually see. Real third alternative: true

C2: Norms may be multi-dimensional

Attack: ‘Toxic’ collapses many dimensions (rudeness, hate, off-topic, conflict). Different mechanisms may govern different dimensions; an aggregate metric can hide discrimination. Evidence: Some communities tolerate profanity but not harassment; others tolerate ideological conflict but not personal attacks. Current status: Split toxicity into subtypes and rerun discrimination; ensure interventions target the right subtype.