Community Direction explained

What a topic weight is

Each category on HYPEXIO (AI, Biotech, Fusion, Longevity, Quantum, Climate, Space, Society) carries a topic weight between 0 and 100 at any moment. The total across all categories sums to exactly 100. The weight answers one question: what percentage of the current community attention sits on this topic?

Unlike classical trend lists, the weight is relative. If AI rises from 28 to 32, that does not mean "30 % more attention in absolute terms" but "AI pulled 4 percentage points out of other categories". The trade-offs are visible. A donut chart visualises exactly that, because donuts show proportions, not absolute values.

The week-over-week comparison rests on the same calculation with a 30-day baseline. If last week sat at 32 and the baseline at 24, the trend indicator switches to "UP". If the weight drops from 32 to 19 (baseline 24), it flips to "DOWN". Within +/-20 % the trend stays "STABLE".

Why activity, not views

Most platforms measure attention via view counts or watch time. That is easy to measure — and lying. A bot-farm run can produce 10,000 views without a single human reading. A headline can go viral because it polarises, not because the field is interesting. View counts blend engagement, algorithmic push, and pure curiosity.

HYPEXIO defines activity more narrowly and with weight: activity = 1 · votes + 3 · predictions + 5 · articles. Votes count less, because they are cheap. Predictions count more, because the user commits. Articles count most, because a new article is a full editorial commitment.

Not perfect, but robust against the usual manipulation levers. A bot can drop a thousand votes — that shifts a topic weight by maybe 0.4 points, and is downscaled by reputation weighting anyway. Mass-submitting predictions costs score points on later Brier resolve if the predictions were wrong. Manipulating articles would require editorial approval — not algorithmically exploitable.

The sum-to-100 normalisation

Each daily recompute runs like this: first measure activity per category over the last 7 days. Then divide by the sum of all category activities. Then multiply by 100. Then round to integer. Rounding remainders are distributed onto the last category so the total stays exactly 100 — the "conservation law of attention".

This has two interesting consequences. First: when AI explodes, something else must drop. The platform does not reward every viral spike — it surfaces where attention is shifting. Second: the system is invariant to absolute activity levels. Whether a week saw 100 votes or 100,000, the donut looks similar.

It follows: in an empty early phase (few votes, few predictions, few articles) the weights are unstable. Three extra votes can shift a topic by 5 points. Only at a few hundred activity points per week does the compass settle.

How to read the donut

The compass donut has eight segments, one per category. Each segment is proportional to its current topic weight. Hover (or tap on mobile) shows the exact value plus the comparison to the 30-day baseline. Cyan highlight marks the top three categories of the week.

Click a segment to land on the category page with all current entries. Directly below, the compass shows leader cards: one concrete article detail per top category with reality and hype score, plus a short reason why the weight currently sits where it sits ("EAST tokamak story collected 412 votes in 7 days").

The trend arrows (UP/DOWN/STABLE) show movement against the 30-day baseline. UP means: activity 7d / baseline 30d >= 1.2. DOWN means: ratio <= 0.8. Anything between is STABLE. Tracking trend arrows across multiple weeks surfaces the long-term shift in community interest.

Why this beats view counts

View counts mix three things: active reading intent, algorithmic recommendation, and randomness. Reading the aggregate signal as "interest" permanently overestimates what algorithmic push mechanics do. That is not malice — it is just lying.

Activity weighting requires a deliberate action: someone clicks vote, someone commits to a prediction, someone writes an article. Each action carries a small but real cost to the user. That filters out passive swipes.

The downside: in the early phase the compass is volatile. Anyone who cannot stomach week-to-week jumps should read the 30-day baseline instead. The platform shows both side by side.

For more depth, the methodology page carries the full formula, edge cases included (what happens at 0 activity in a category, how the baseline is computed, how old a contribution may be to count).

When the compass recomputes

The recompute job runs every day at 04:00 UTC, before the 08:00 UTC newsletter. One transaction, no LLM, cost cap near zero. To enable: set COMMUNITY_COMPASS_ENABLED=true in /opt/hypexio/.env; the cron pattern hangs on COMPASS_RECOMPUTE_CRON (default 04:00 UTC).

Each run writes one snapshot per language into community_compass_snapshots. This feeds the donut visualisation, the leader cards, the stats, and the trend arrows. Snapshots stay for 90 days — after that, historical topic weights remain without the snapshot body.

The job is deliberately LLM-free. Anyone who tried to interpret community data via LLMs usually came back with embedded biases. A direct SQL aggregation is transparent, reproducible, and cost-neutral.