How the Pulse works — and where each tier can lie
The Profoundd Public Sentiment Pulse is a multi-tier dashboard that measures the gap between what mainstream coverage is telling you and what other signals are showing. Every tier has documented biases. We surface them rather than hiding them. This page is the honest methodology disclosure that every confident claim on a topic page links back to.
The litmus test for any topic page
For any topic page to publish a confident "what the public thinks" headline, at least 3 of 4 tiers must be reporting data that points the same direction:
- Random-sample poll from a Pollster with verified track record (Tier 2)
- Behavioral / outcome signal — votes, real-money market moves, demographic shifts (Tier 0)
- Multi-platform social signal where left-coded AND right-coded platforms agree (Tier 1)
- Profoundd audience signal aligning with one of the above (Tier 6)
When only 1 or 2 tiers align, the topic page shows WEAK SIGNAL. The system does NOT manufacture a consensus number — the disagreement IS the headline.
Where each tier can lie
Tier 2 — Polls
We do not treat any single poll, or a blended poll average, as ground truth. We score each pollster against measured outcomes (Tier 0). Known structural problems:
- Response-rate collapse. Live-caller poll response rates went from ~9% in 2000 to under 1% today. The 1% who do respond are not representative.
- Mode effects. Online-panel polls and live-caller polls disagree by 4-8 points on the same question. Mixing them in an "average" hides this signal.
- Weighting opacity. Pollsters weight by demographics and recalled vote. Recalled-vote weighting is a thumb on the scale and assumptions are rarely published.
- House effects. Rasmussen runs Trump-favorable ~4 points vs. median; Marist runs ~3 points Dem-favorable. These are STABLE biases, not noise. The pollster scoreboard surfaces them per pollster.
Tier 1 — Social platform sentiment
We never combine left-coded and right-coded platforms into a single number. We report them as paired signals with the spread visible. The dominant problem:
- Social-media users are not representative. Roughly 10% of people post about politics; the other 90% are invisible to this tier. Posters skew younger, urban, college-educated, and more politically engaged.
- Platform algorithms amplify engagement-bait. What we measure is filtered through "what platforms chose to surface," not "what people are saying."
- Reddit's downvote-brigade dynamic distorts the visible top-of-thread sentiment. We use "controversial" sort + upvote-ratio weighting to discount brigaded posts. The limitation is documented, not eliminated.
- LLM scoring carries left-bias in training data. We use VADER (deterministic lexicon) as the floor, plus multiple LLM scorers (Claude + Grok) on contested items and store the spread. High-spread items are flagged "contested," not blended.
Tier 3 — Prediction markets
Polymarket and Kalshi probabilities are "money on the line" signals that often beat polls on binary questions (elections, major event resolutions). Limits:
- Thin liquidity outside flagship questions reduces signal quality on niche topics.
- Most policy questions don't get bettable markets.
- Markets can be moved by whales; we don't filter for this at v1.
Tier 4 — Mainstream coverage
This is the comparison pole, not a truth signal. Operator-curated mainstream-source list (AP, Reuters, NYT, WaPo, WSJ, ABC, CBS, NBC, CNN, Fox, etc.) — the curation IS the editorial choice. Sentiment scoring on articles uses our existing single-LLM tagging plus a VADER backfill for transparency.
Tier 5 — Latent forecasting
Modal-language extraction (anger, fear, hope, dread, certainty) per topic over time, plus Hawkes-process attention-curve forecasting. Output: "topics likely to surface in mainstream within 7 days." Every forecast is auto-scored against eventual mainstream-coverage spike — our own accountability matches the discipline we apply to pollsters.
Tier 6 — Profoundd audience signal
Search queries, page views, AI-question topics, and feedback from Profoundd readers. Self-selected, smaller than the country, and editorially aligned with Profoundd's perspective. Most useful as: topic-emergence leading indicator (your audience leads the curve on issues they care about), Bob editorial calibration, and triangulation when other tiers disagree.
What this system cannot tell you
Even with all 6 tiers running cleanly, the Pulse cannot reliably tell you "what Americans actually think." To get closer to that, you need demographically weighted academic surveys (Pew Research, ANES, GSS) and behavioral signals (voter registration shifts, donation patterns, migration). Those reference layers are documented as Phase 1.1 roadmap; they materially expand what the system can represent.
Foundational references
- GDELT Global Knowledge Graph — global news tone, free
- cardiffnlp/twitter-roberta-base-sentiment-latest — sentiment model for social text
- BERTopic — emerging-topic discovery, Grootendorst 2022
- Tetlock 2015 Superforecasting — the philosophy of scoring forecasters against outcomes
- Tumasjan 2010 "Predicting Elections from Twitter"; Asur & Huberman 2010 "Predicting the Future with Social Media"; O'Connor 2010 "From Tweets to Polls"
- Clif High archive on archive.org — original web-bot methodology (take with skepticism)