The algorithm, step by step.

Does the strength of the post match the strength of the evidence card? We compute a speaker confidence level (uncertain / hedged / assertive / assertive-with-absolutes) and a proper confidence ceiling for the card's evidence level and direction. Overstatement reduces the score.

• Aligned or understated → 38
• Slightly overstated → 28
• Overstates weak / ambiguous evidence → 14
• Strongly overstates → 5
• Asserts against consensus evidence → 2

Does the post specify population, intervention, outcome — or generalize? "Works for everyone" triggers a penalty when the card's scope is narrower.

• Narrow scope matching the card → 20
• Broad scope where the card is narrow → 10
• No population/intervention/outcome at all → 3

Rewards hedging and explicit uncertainty. Penalizes absolutes ("proven", "every", "#1"); the absolutes penalty scales by evidence level.

• Baseline 15
• Absolutes on H5 → −3 (stylistic tic)
• Absolutes on H4 → −5
• Absolutes on H3 → −6
• Absolutes on H2 → −8
• Absolutes on H1 → −10 (actively misleading)
• Assertive tone on ambiguous/misaligned claim → −4
• Acknowledges uncertainty explicitly → +2

Does the account acknowledge conflicting evidence, cite sources, or update stale claims?

• Baseline 6
• Acknowledges conflicting evidence → +3
• Post cites a source (referenced → +1, linked → +2)
• Misaligned claim with no acknowledgment → −4

Selling the exact intervention being advocated is not disqualifying — but it is penalized more heavily as evidence gets weaker. Commercial overlap is read from the account's product list + a claim-family overlap map.

• Direct product overlap, H1/H2 evidence → −7
• Direct overlap, H3 evidence → −4
• Direct overlap, H4/H5 evidence → −2
• Indirect commercial overlap → −2

Rewards mentioning side effects, contraindications, or uncertainty on interventions with non-trivial risk. Safety cards are flagged low / low-to-moderate / moderate / high.

• Baseline 5
• Moderate/high-risk intervention with no safety note → −3
• Absolute framing on non-low-risk intervention → −1

From the card's safety_flag:

• low → 1.0 · low-to-moderate → 1.2 · moderate → 1.5 · high → 2.0
• H1/H2 evidence multiplies severity by 1.2 (weak evidence magnifies consequence)

Engagement tier stands in for viral spread.

• high → 1.5 · medium → 1.0 · low → 0.6

Recent behavior weighs more than historical.

• ≤ 3 months → 1.2 · ≤ 12 months → 1.0 · older → 0.8

Instances of the same claim family are averaged first, then the family score is weighted once. This prevents a single well-known claim from dominating by sheer repetition — and prevents a good claim from being diluted by repetition.

Account repeatedly makes assertive claims on H1/H2 evidence while selling a matching product. Fires at ≥ 2 such claim families.

Account repeatedly discourages established care for serious conditions. Scaffolded; requires an "anti-establishment rhetoric detector" not in v1.