Sanity Media 2.4

This is is a relatively simple AI prompt and scoring model I initially posted on X and Blue Sky that seems to work fairly well at giving folks an "ingredient list" for any media content they might run across. The key is that it is reasonably non-partisan to give you a reference to think more objectively about what an article or post is trying to say. Note that my benchmark tests currently use Chat GPT 5.6 Sol. Folks are welcome to explore it's usefulness with other models. Enjoy and feel free to add constructive commentary below.

To be clear: You can to curse like a drunken sailor, but add something constructive to the discussion and do not attack others politically or personally or your post will be taken down.

P.S. I make no claims about the accuracy of the model. It is simply a tool to use at your own risk if you find it useful.

AI prompt:

< place the article, post, or media url here. If AI can't access it you can usually place the text of it here.>

Please score the media above with the Sanity Media scoring system, given below. Please adhere strictly to the rules of the framework. Do not invent or create rules other than the text given below.

Sanity Media 2.4 scoring
(1–100 each, 10 point resolution, uncertainty of at least +/- 10 is assumed):
• Mainstream Alignment (institutional narrative)
• Accuracy (corroborated evidence)
• Clickbait (attention engineering)
• Toxicity (tone)
• Synthetic Indicators (artificial Synthetic Indicators potential)
Paste content into AI → get a score.

• Mainstream Alignment Score:
• measures alignment with institutional, credible discourse (AP/Reuters/expert consensus)
• 0–20 outlier
• 20–40 edge / limited adoption
• 40–60 Mainstream debate
• 60–80 widely accepted
• 80–100 institutional baseline / consensus
• rules: Mainstream ≠ truth; do not use virality/platform prevalence; if routinely present in wire-level reporting → Mainstream ≥50

• Accuracy Score:
• measures alignment with corroborated, verifiable evidence
• 0–20 false / contradicted
• 20–40 weak / misleading
• 40–60 mixed / partially supported
• 60–80 mostly supported
• 80–100 strongly corroborated

• rules:
• corroboration > authority
• no single source determines accuracy
• score claims, not sources
• uncertainty lowers score
• use institutions as inputs, not arbiters
• source count alone does not equal corroboration
• stronger corroboration comes from independent evidence across different source types and information ecosystems
• examples of source diversity include mainstream/wire reporting, alternative or independent media, local or specialist reporting, individual/firsthand evidence, primary documents or datasets, and academic/technical analysis
• source category does not itself determine credibility; evaluate the evidence each source contributes
• several sources relying on the same underlying document, witness, leak, dataset, or reporting chain count as substantially less than genuinely independent corroboration
• agreement across different source ecosystems is especially useful when the sources also rely on different underlying evidence
• Circular Laundering Rule:
Multiple outlets repeating the same single-source report, anonymous official leak, document, witness, dataset, or upstream reporting chain count as ONE evidentiary source for corroboration, not independent verification.
Example: Reuters reports an anonymous official claim and ten other outlets repeat Reuters. This remains one evidentiary chain.
Example: Reuters and AP separately quote different officials. This provides more independence, but both may still represent the same evidence class and institutional perspective.
Example: wire reporting + primary records + local firsthand evidence + an independent technical analysis represents substantially stronger corroboration because both the sources and evidence types differ.

Accuracy remains capped at ≤50 when a consequential claim rests primarily on one unverified evidentiary chain. The cap may be lifted by primary evidence, a genuinely independent second-party investigation, or converging evidence from materially different source types or information ecosystems
• Clickbait Score:
• Attention engineering (emotional hooks, narrative framing, engagement optimization — independent of Toxicity, do not collapse)
• 0–20 Neutral / Technical. Dry, informational, zero editorializing. Rare outside academic journals, whitepapers, and raw wire feeds (AP/Reuters unedited text).
• 20–40 Standard Reporting. What "standard" used to mean: quiet, direct, matter-of-fact. Increasingly uncommon in consumer-facing media.
• 40–60 Mild Framing — THE BASELINE. Standard Mainstream articles, local news, major tech publications. Almost everything now carries a narrative hook or interpretive angle. This is the floor of modern media, not a warning sign.
• 60–80 Noticeable Push. High-velocity digital media, legacy opinion columns, content optimized for algorithmic feeds (YouTube/X). Language is doing real work beyond conveying facts.
• 80–100 Extreme Engineering. Pure engagement farming: outrage Synthetic Indicators, hyper-compressed out-of-context clips, deliberate bait.
• rules: Baseline is 40–60 (modern media floor); Signal threshold: ≥60 — below this, attention engineering is ambient, not diagnostic on its own.

• Toxicity Score:
• measures tone / hostility (not disagreement)
• 0–20 neutral / clinical
• 20–40 mildly loaded but civil
• 40–50 adversarial but restrained (baseline political coverage)
• 50–70 heated / accusatory
• 70–100 hostile / dehumanizing
• rules: baseline reporting lives in 0–50; >50 requires clear escalation; insults/dehumanization drive high scores. Celebrating, inciting, or encouraging direct violence is automatically 100, as is doxxing.

• Synthetic Indicators Score (multimodal):
• measures synthetic/templated production patterns across text, image, audio, video
• 0–20 organic
• 20–40 mostly organic
• 40–50 mild templating
• 50–70 synthetic traits
• 70–100 strong synthetic patterns
• rules: probabilistic, not definitive; do not assert AI vs human; production quality ≠ truth; detects patterns, not origin

• Consistency & Verification Guardrails:
• contradiction = attributed claim conflicts with well-documented, recent public record of the figure (statements, actions, positions)
• contradiction → Accuracy ≤50 until corroborated
• contradiction + high Synthetic Indicators (≥60) → Accuracy ≤30 until independently verified
• production quality does not raise Accuracy
• lift caps only with multiple independent credible sources or primary confirmation (official channels, full context)
• allow for genuine position changes → require verification, not assumption• Consistency & Verification Guardrails:
• contradiction = attributed claim conflicts with well-documented, recent public record of the figure (statements, actions, positions)
• contradiction → Accuracy ≤50 until corroborated
• contradiction + high Synthetic Indicators (≥60) → Accuracy ≤30 until independently verified
• production quality does not raise Accuracy
• lift caps only with multiple independent credible sources or primary confirmation (official channels, full context)
• allow for genuine position changes → require verification, not assumption

• Core Principles:
• Mainstream ≠ Accuracy
• Accuracy ≠ source trust
• Clickbait ≠ falsehood
• Toxicity ≠ disagreement
• Synthetic Indicators ≠ deception
• do not collapse axes
• score content, not intent
• disagreement between scores is expected

• Signal Principle:
• signals abound across the model
• Mainstream vs Accuracy gaps are one signal
• Accuracy vs Synthetic Indicators → verification risk
• Clickbait vs Accuracy → framing vs evidence
• Toxicity vs Mainstream → tone vs Mainstream position
• Synthetic Indicators vs Clickbait → synthetic amplification
• do not collapse axes—patterns across scores are the signal

• Institutional Omission rule: If a claim is strongly corroborated (Accuracy ≥ 60) but mostly absent from wire-level reporting, flag for Institutional Omission. The low Mainstream score here is a diagnostic signal of gatekeeping, not invalidity.