How Future Index calculates its numbers

Probability

The source's own most recent published probability for this question, unchanged.

Each source is polled on a schedule and every reading is stored with the instant the source says it applies to. The published probability is the newest of those readings, rounded and otherwise untouched.

Nothing is rescaled, re-weighted, or reconciled against any other source. What is shown is what that one source published.

What this does not say

• It is one source's view of one question, not a consensus and not an average.
• A probability from a source with very little participation behind it is still published, with its confidence and its participation figure beside it.

Change over time

How far the probability has moved over each window, in probability points.

For each window the baseline is the newest reading at or before the start of that window, and the change is the current probability minus that baseline. The baseline's own timestamp is published beside the figure, because the newest reading a window back is rarely exactly one window old.

A baseline older than the limit below is not used. The change is reported as unavailable and a chart draws a gap.

What this does not say

• An unavailable change means there was nothing to measure from, which is a different statement from no movement. The two are never rendered the same way.
• A change is a difference between two readings from the same source. It is never a difference between two sources.

Variability

How much the probability has been moving about within each window.

Each change between consecutive readings is scaled by the time elapsed between them and squared; the published figure is the square root of the mean of those values, in probability units per square-root hour.

Scaling by elapsed time is what makes the figure independent of how often the source happened to be polled. Without it, a source polled twice as often would appear twice as variable.

What this does not say

• It describes how much a number moved, never why it moved and never whether the movement was justified.
• Below the minimum sample stated below, the figure is left empty rather than estimated from two points.

Confidence

How well supported this source's number is — never how large, how recent, or how interesting the number is.

Four inputs are scored from 0 to 1: participation behind the contract, the age of the newest reading, the distance between the two quoted sides, and the time remaining until the question settles. Each is published individually with the share of the score it carried.

The four are combined as a weighted geometric mean rather than an average, so one input near zero pulls the whole score down instead of being averaged away by three healthy ones. A contract with nothing staked on it is not two-thirds trustworthy for being recent, tightly quoted and close to settlement.

An input a source does not publish — many forecaster aggregates quote no two-sided spread — has its share redistributed across the inputs that are present, so every source's confidence is scored out of the same total. A source that publishes no participation figure at all is scored zero on that input rather than having it redistributed: absent evidence of support is not evidence of support.

Each input is rounded before the inputs are combined, so the breakdown a reader is shown reproduces the score printed beside it.

What this does not say

• Confidence has no access to the probability it describes. A dramatic move on a low-confidence contract stays low confidence.
• It says nothing about whether the source is right. It is a statement about how much evidence stands behind one published number.
• It is not agreement with anyone else. Nothing here reads a second source.

How easily this could be moved

A band read from participation alone: how much it would take to move this contract.

Depth is compared against two thresholds, on its own scale — dollars for a market, contributors for a forecaster aggregate — and the result is reported as low, elevated, or high.

A source that publishes no participation figure lands on elevated rather than low, with the reason stated. Low is a claim that a contract is expensive to move, and nothing supports that claim without a number.

What this does not say

• Nothing here detects a manipulation pattern, and the band must not be read as an accusation. The claim is narrower and checkable: a contract this thin is cheap to move.
• The two scales are never compared to each other. A dollar figure and a contributor count are not the same quantity.

Freshness

Whether the reading behind a number is current, out of date, or missing entirely.

Every value carries the instant it was observed and the instant it was calculated. A value is current while its newest reading is within the number of expected poll intervals below, and marked as possibly out of date after that.

Where a source has published nothing at all there is no value to mark, so no figure is produced. A page narrows; it does not fill the space with an old number wearing a current label.

What this does not say

• The poll interval a reading is measured against is the one the scheduler chose for that source at the time, which varies with demand, with the source's own rate limits, and with its health.
• Out of date and unavailable are different states and are drawn differently. A chart shows a gap where a source was down; it never draws a line across it.

Coverage

Which sources stand behind a number, and which do not.

Every published value names the source it came from, that source's class, and the participation or depth figure behind it. Because one contract is one source's question, coverage here is always a single source, stated rather than implied.

The full registry of sources — what each one is, the permission it runs under, and whether it is permitted to run at all — is published at /v1/sources. A source that is present in the registry and not contributing is visible there, which is what makes an empty figure readable rather than mysterious.

What this does not say

• No language or geography claim is attached to a forecast contract. The sources do not state one, and inventing one would be worse than reporting none.
• A source that is listed and not running contributes nothing, and no other source stands in for it.

Precision

How far each published number is rounded, and why the rounding is part of the method.

Rounding is applied before a value is stored, not on the way to a screen, so the stored value and the published value are the same number and a stored value can be recomputed from its stored inputs exactly.

The precision below is far finer than any source publishes and far coarser than a floating-point result, which is deliberate: fifteen digits invites a reader to believe the last ten, when the input was quoted to two decimal places.

What this does not say

• Precision is not accuracy. A probability published to six places is a source's own two-decimal quote, carried without loss.

The attention figure

How much a topic is being discussed and looked up right now, compared with how much that same topic normally is.

Each source is scored against that topic's own recent history on that source — not against other topics, and not against other sources. The scores are then combined into a single figure from 0 to 100, where the middle of the scale is the topic's own normal level.

A figure above the middle means the topic is busier than it usually is. A figure below it means it is quieter than it usually is. Every published figure carries a plain-language label beside it saying which.

What this does not say

• It says nothing about how many people in the world are paying attention. A topic that is permanently busy sits in the middle of the scale for as long as it stays that way, and a small topic having an unusual week can score higher than a large one having an ordinary one.
• It is not comparable between topics as a level. Two topics both reading 80 are each unusually busy for themselves; that is all it says about the pair of them.
• It says nothing about whether the attention is warranted, welcome, or accurate. More attention is neither good nor bad here.

Turning one source's activity into a score

Each source's departure from that topic's own history is placed on the same 0 to 100 scale, so platforms that count entirely different things can be combined.

A source reports a count — comments, stories, page views. The count itself is not comparable with another source's, so what is used is how far it sits from that topic's usual level on that source, measured in the spread of its own history.

That departure is mapped onto the scale by a curve that is centred on the topic's normal level and flattens at both ends. It always rises: a larger departure always produces a larger figure, and there is no point at which two different departures produce the same number.

Where a source sees too few items on a topic for that comparison to mean anything, it produces no score at all rather than a confident-looking figure over counts that are usually zero. The source still counts toward coverage and is named in the limitations.

What this does not say

• A departure from normal is not an explanation. It says the level changed, never why, and never that anything caused it.
• A platform's own changes — an outage, a redesign, a moderation decision — move its counts without anything happening in the world. Having several independent sources is what guards against reading one of those as a real change.

Discussion and news are counted separately

How much the public discussed a topic and how much was published about it are two different measurements, and only the first can carry the figure.

Sources are placed in one of two groups by what kind of source they are, which is recorded when the source is reviewed rather than decided per calculation. Discussion platforms and reference activity form the public group; news outlets and official releases form the context group.

The context group has a fixed maximum share of how far the figure may move from the middle of the scale. That share does not grow when the public group is quiet or absent: a topic that only news sources reported on stays close to the middle of the scale and cannot be published as elevated.

Within a group, no single source may carry more than a set share, so one busy platform cannot decide the figure on its own. A group containing one source is the exception the arithmetic cannot avoid, and the confidence score is what reports it.

What this does not say

• The context group is capped because volume of coverage is a measurement of newsrooms, not of the public. Twenty outlets carrying one wire report is one occurrence reported twenty times, and it is not twenty independent signs of public attention.
• The cap does not make coverage volume worthless — it remains part of the figure and is published as its own share, so a reader can see exactly how much of the number it accounts for.

Visits to this site are one of the sources

How much this site's own pages were read counts towards a topic's attention, and is deliberately held to a small part of it — a topic must not become prominent here merely because it was shown prominently here.

Only a page about one subject is counted. A list, a ranking or anything that displays several topics at once records nothing for any of them, so appearing on such a page is never itself the evidence.

Visits are grouped into periods and counted to a limit within each. Once that limit is reached, further visits to the same page in the same period change the figure by nothing — so a page that is being read heavily produces the same number as one being read steadily.

What is counted may account for no more than a set share of the figure. Where no other source has anything to say about a topic, the rest of that share is held at the middle of the scale rather than passed to this one, so visits here alone cannot move a topic up the scale.

It also earns none of the increase described in the section below: this site reading its own pages is not an independent source agreeing with another one.

A search counts towards the subject it matched, never towards the words that were typed. Nothing is recorded about who visited, when within the period, or in what order — the whole of what is kept is a subject and a total.

What this does not say

• It undercounts, and by more the busier a page is. Pages are cached, so a page being read continuously reaches the counter far less often than it is read, and there is no way to correct for that without identifying visitors — which this product does not do.
• It is a measurement of this site's readers, who are not the public. It is included because interest here is interest, and it is capped because interest here also moves with what this site chooses to show.
• It cannot be compared with the same count from anywhere else. A visit here and a page view on an encyclopaedia are both called page views and are not the same quantity.

Independent sources agreeing

When several unrelated public sources see the same departure at the same time, the figure reaches further up the scale than one source alone can take it.

Each additional public source that produced a score increases how far the public group's departure is allowed to move the figure, up to a fixed limit.

The increase applies to the public group only. Agreement between news outlets is not treated as independent evidence, for the reason given in the section above.

What this does not say

• Sources being independent of each other is an assumption about the platforms, not something measured. People discuss the same story in several places, and this treats those as separate observations because there is no way here to tell that they are not.
• The figure is still bounded at 100. A topic already at the top of its own scale does not go higher for being seen by a fourth source.

How well founded the figure is

A separate score from 0 to 1, built from how many independent sources contributed, how much history there was to compare against, how much activity sat behind it, and how concentrated that activity was.

The four parts are combined so that a weak one drags the whole score down rather than being averaged away by strong ones. A topic scored from a single source with barely any history is not two-thirds trustworthy for having plenty of activity.

The history part reads the *least* observed of the contributing sources, not the best. A topic is only as well founded as its newest source, and sources differ: a page-view history can be assembled from published archives, while a discussion platform's history only starts accumulating when we begin reading it.

The concentration part is what the risk band beside the figure is set from. Where no contributing source publishes an account count — a page-view total has no accounts behind it that anything can inspect — that part is left out, its weight is shared among the others, and the risk band reads elevated rather than low.

Confidence never sees the figure it describes. A dramatic departure with one thin source stays low confidence, and the code cannot do otherwise because the value is not passed to the calculation that produces this score.

What this does not say

• A high confidence score is a statement about the evidence behind the figure, not a claim that the figure is important or that the attention is deserved.
• The risk band says how much of the evidence was discounted for coming from few accounts. It is not a finding that anybody coordinated anything, and nothing in this system looks for one.

How current the figure is

Each source is judged against its own update rate, and the figure reports the worst of them.

Sources publish at different rates and that is not negotiable by asking more often — one of them serves daily totals and rejects an hourly request outright. Each source's reading is therefore judged late or current against its own rate, so a daily source is not marked out of date for being a few hours old.

The figure carries the worst state among the sources that actually went into it. A source that answered without producing a score — because it saw too little of the topic to say anything about it — is named in the coverage below and in the limitations, and does not decide how current the figure is: a source too sparse to score is a source with little recent data almost by definition, so letting it set the state made the state a fact about the quietest source rather than about the number.

Lateness is counted from the end of the period a reading covers, not from its start. A reading is not behind while the period it describes is still running.

A reading for a period that has finished is not made less current by being recalculated afterwards. One measurement arrives late by nature — how much weight a platform's items eventually attracted keeps changing for more than a day after the period they were posted in — so those periods are read a second time and the reading is corrected. How current a reading is stays measured from when the period was first read, because what changed was our knowledge of it and not the period.

The clocks the contributing sources were measured on are published alongside the figure, because a number blended from a daily source and an hourly one updates on both.

What this does not say

• A current figure is not a live one. The fastest source here updates in minutes, not seconds, and the slowest updates once a day.
• A source that has stopped reporting entirely is named separately, in the coverage below, rather than quietly lowering this figure.

Which sources are behind the figure

Every published figure names the sources that contributed, the share each one carried, and the sources that were expected and reported nothing.

Each contributing source is published with its own score and the share of the figure it accounts for. Where a public source contributed, those shares add up to the whole; where none did, they add up to the context group's cap and the shortfall is the missing public group.

A source that has reported on this topic before and did not this time is named as missing. A source that answered but saw too little to score is named separately, because those are different facts and only the first is an outage.

No combined count of items is published. The sources count different things — page views, comments, stories — and a total of them would be a number with no unit presented as though it had one. Each source's own count is published beside it instead.

What this does not say

• Coverage is a list of what was read, not a claim that it is everything worth reading. The sources here are the ones whose terms permit this use and whose data this system ingests.
• A missing source lowers confidence and is named, but the figure is still published. One failed source narrows what is shown; it does not blank the topic.

Rounding and versions

What is stored is rounded before it is stored, and every stored figure carries the version of the method that produced it.

Figures are rounded to the precision below before they are written, so a published breakdown reproduces the figure beside it rather than nearly reproducing it. Published shares are adjusted so that they still add up to the whole they are shares of — three equal sources of a group round to shares that would otherwise total slightly more than the group.

The version is stored on every figure. When a threshold on this page changes, the version changes with it, and every figure calculated under the old version keeps saying so.

What this does not say

• The precision here is well beyond what the underlying counts support. It exists so that a change to the method is visible in review, not because the last decimal place means anything.

Four figures, and deliberately no fifth

The World Pulse is four separate readings — how much expectations are moving, how much attention there is, how the discussion feels, and how much is in doubt. They are never added together.

Each of the four is calculated on its own, from its own sources, and carries its own confidence, its own coverage and its own timestamp. Nothing in this system combines them, and there is nowhere in the published shape to put a combined figure.

That is a design commitment rather than an omission. A single number would have to decide, silently, how much a shift in expectations is worth against a change in tone — a judgement nobody has made and nothing here could defend. The four are shown together because they are read together, not because they are parts of one quantity.

Each figure may be missing on its own. One failed source narrows the set of readings and never blanks the rest, and a missing reading says which of the four it is and why, rather than being shown as a zero.

**Every one of them is a comparison with its own past, not a position on an absolute scale.** Each is computed as a raw quantity in its own units — how much of the tracked world is unusually active, how far expectations moved, how the discussion leans — and then placed by how far that quantity departs from what the same quantity has usually been over the last two weeks. The middle of every scale therefore means *about what this normally is*, and the top means *further from normal than this has been*.

The distance is measured in robust deviations: the centre of the comparison is a median and its width a median absolute deviation, so a single extraordinary day does not widen the yardstick every later day is measured with. It is the same arrangement each individual subject's own figures use, which is why a subject three deviations above its own normal and a world three deviations above its own normal read the same number.

A figure with too little of its own history behind it is **not published at all**, and says so. A middling number from something with no idea what usual looks like would be the one reading that is certainly wrong.

What this does not say

• The four are not comparable with each other. An attention figure of 70 and an uncertainty figure of 70 have nothing to do with one another; each is on its own scale with its own meaning.
• Because each figure is compared with its own recent past, a state that persists stops reading as unusual. A world that has been loud for a month reads as ordinarily loud, and the figure describes the change rather than the level. What it cannot say is *this has been extreme all year*.
• The comparison reaches back two weeks and no further, so it describes departures from the recent past rather than from any long-run normal — and the past it uses is this product's own record, which is shaped by which sources were ingesting at the time.
• None of them is a forecast. They describe what is being expected, watched, said and disputed right now — never what will happen.
• A figure moving is not an event. These are readings taken repeatedly, and readings move.

Outlook — how much expectations are moving

How much the probabilities on the forecast questions this product tracks have changed over the last day, measured as how many of them moved and how far.

Where the Forecast Movement figure comes from — the short account

Each question is judged against itself. A question's own recent history says how much it normally moves in a day; the movement over the last day is measured in multiples of that, so five points of probability counts as a large move on a question that usually barely shifts and an ordinary one on a question that swings routinely.

Two halves are then combined in equal parts. The first is how many questions moved beyond the threshold below — the breadth of the change. The second is how far the average question moved, weighted by how well supported each question's number is, on a scale that flattens at the top so that one dramatic question cannot decide the figure alone. Both halves are published separately as well as blended.

A question with no reading from a day ago, or with too little history to say what an ordinary move is for it, is left out and counted. The count is published, because a figure taken over eight questions out of four hundred is a different statement from the same figure over all four hundred.

What this does not say

• **It has no direction.** It says how much expectations moved, never whether they moved towards something better or worse. That would require deciding which outcomes are good, which this product does not do.
• It is not a level. A high figure does not mean anything is likely or unlikely; it means the numbers have been moving.
• It says nothing about why anything moved. A settlement, a news event and a large trade all look the same from here.
• Questions from different sources are never compared or combined. Each question is one source's contract, and this figure counts movements rather than reconciling them.

Attention — how much of the tracked world is unusually active

How many tracked topics are departing from their own normal, and how far those particular topics have gone. Topics behaving normally are counted in neither and cannot move the figure.

Where the Attention figure comes from — the short account

Each topic's attention figure is calculated first, against that topic's own history on each source. A topic above the upper line below is counted as unusually active for itself; one below the lower line is counted as unusually quiet. A topic between them is ordinary and takes no part in this figure at all.

Two quantities are measured from that. **Total elevation** is how far past the upper line the departing topics sit, added up rather than averaged — so a hundred subjects behaving normally contribute exactly nothing and cannot move it. **The loudest few** is the same measure over the three most unusual subjects alone, which exists because no figure that treats two hundred subjects alike can show a single extraordinary one: that subject is one of two hundred whatever the weights are.

Each of the two is compared with its own recent past before they are combined, weighted as below. Comparing them first is what makes the weights mean what they say — the two run at very different sizes, and adding them raw would make one of them the whole figure.

An unusually quiet world needs no separate term: it is one whose total elevation is far below what it usually is, which the comparison says on its own.

What this does not say

• **It was an average of every tracked topic's figure until method version 2.0.0, and that made it partly a measurement of this product's own topic list.** Adding a hundred ordinary topics to the list pulled the average toward the middle, and retiring a few pushed it away, with nothing in the world having changed. Ignoring the topics that are behaving normally is what removes that. Figures published under the older version are on a different scale and are not points on the same series.
• **Between 2.0.0 and 5.0.0 it barely moved at all, and that was a defect rather than a calm world.** The two halves it was built from were subtracted from one another and both sat at their ceiling on almost every reading, so they cancelled: measured over a week in September 2026 the published figure never left a five-point stretch around the middle of its hundred, and an event taking one subject to the top of its own scale moved it by under half a point. Figures published under those versions are not points on the same series as these.
• It is built from relative figures, so it cannot say how much attention there is in absolute terms. If the departing topics are small ones, this still reads high — they are unusual *for themselves*.
• It describes the topics this product tracks, which are a chosen set. See the scope note below.
• More attention is neither good nor bad. Nothing here reads a high figure as alarming.

Sentiment — the balance of the discussion that could be read

The positive, neutral and negative split across every discussion item that was scored, published as a split and as a single figure saying how that balance compares with its own usual level.

Where the Mood figure comes from — the short account

Items are pooled within each platform and the platforms are then combined equally. Two comments on one platform are two of the same thing and add up; a comment on one platform and a post on another are not, and letting the busier platform's item count decide the overall figure is exactly what this arrangement avoids.

The overall balance of the discussion is the quantity, and the published figure is how far that balance sits from what it has usually been over the past two weeks. The middle of the scale means *the discussion leans about as it normally does*, not *the discussion is balanced* — a conversation that is steadily grim reads in the middle, and the split beside it is what says so.

The balance itself is published as well, so nothing about the split is recoverable only through the figure. That is what the plain rescaling this replaced was for; what it could not do was move. A mean over thousands of items has almost no room to vary, and the published figure spent a whole week inside 1.6 points of its hundred — which is arithmetic rather than a quiet world.

Items whose language the classifier cannot read are excluded rather than counted as neutral, and the figure describes the discussion that could be scored.

What this does not say

• **This is one online community's discussion, not a public.** One discussion platform is available to this product today; the other is not reachable under its operator's current terms. The people who post there are not representative of anyone, and the figure names the platform it came from.
• It measures the tone of what was written, not what anybody believes, and not how many people hold a view. A loud minority and a quiet majority are indistinguishable from here.
• The classifier is a model with its own errors, and its identity is stored with every figure so that a change of model is visible rather than looking like a change in the discussion.
• The single figure is a summary of the split *and* a comparison with the past. Where the actual balance matters — and it usually does — the split is what to read.
• Because it is a comparison, a persistently negative discussion reads as ordinary. This figure describes whether the tone has shifted, never how dark it is in absolute terms.

Uncertainty — how much is in dispute or in motion

Built from how unsettled forecast prices are, how far communities differ, how quickly attention is moving, and how weakly founded the other three figures are.

Where the Uncertainty figure came from — the short account

Each part is scored from 0 to 1 and they are combined by the weights below. A part that cannot be measured at all is left out and its weight is shared among the rest, rather than being scored zero — not measuring something and measuring it to be absent are different statements.

The parts are added rather than multiplied, which is the opposite of how the confidence score below is built, and the difference is deliberate. Forecast prices being settled does not make a violent swing in attention less unsettling, so one calm part must not be able to pull the whole figure down.

The last part is a reading of the other figures rather than of the world: it is high exactly when the rest of the vector rests on little. It is what stops a pulse assembled from one working source out of three from reading as a calm day.

What this does not say

• **Disagreement between forecast sources is not measured at all.** It is the input this figure would most obviously want, and comparing contracts across sources is out of scope for this product entirely — see the boundary on the outlook figure above.
• Uncertainty is not bad news. A high figure means the signals are moving or disagreeing; it says nothing about whether anything harmful is happening.
• It is not a probability, a risk level or a forecast of volatility. It is a summary of how much the things this product measures currently disagree with each other and with themselves.
• Community disagreement contributes nothing today, because only one community is being read.
• **When a whole layer goes dark this figure can drop rather than climb, and it is left that way on purpose.** The last part is computed from how well founded the figures that *were* published are, and withdrawing the weakest founded of them raises that average and lowers the doubt: measured on a test world, taking the attention layer away moved this figure from 34.09 down to 29.40. That is the defect this figure was withdrawn for, and the repair belongs to what replaced it — the data conditions block counts every reading this product defines rather than only the ones it managed. Changing the arithmetic here instead would mean a withdrawn figure quietly measuring something new under the same name, which is exactly what a deprecation window is for preventing. A client still reading it receives the quantity it was written against, defect included.

Turbulence — how fast the signals are moving

How rapidly and unusually the things this product measures are changing, whichever direction they are changing in.

Where the Turbulence figure comes from — the short account

Five parts, each scored from 0 to 1 and combined by the weights below: how far the forecast questions moved against their own normal, how unsettled their prices are, how fast attention shifted across topics, how far the tone of the discussion turned, and how many of those kinds of signal were moving at the same time.

The first, third and fourth reuse figures that are published in their own right rather than recomputing something similar, so a reader who asks why turbulence is high gets an answer that is literally the same arithmetic as the figure they can already see.

A part that cannot be measured is left out and its weight is shared among the rest. The parts are added rather than multiplied: forecast prices being settled does not make a violent swing in attention less turbulent.

The last part is a count rather than a magnitude, and it is a part with its own weight rather than a multiplier on the others — so several kinds of signal quietly agreeing that nothing is happening cannot be scaled up into a storm.

What this does not say

• **It has no direction and no valence.** A calm world can be a negative one and a turbulent world can be a positive one; this figure says nothing about either. A high reading is not bad news.
• **Nothing about how well this product can see can change it.** If several sources go dark, this figure narrows — the parts that cannot be measured drop out — and it does not rise. How much is currently unobservable is published separately, under Data Conditions, and never inside a reading of the world.
• It measures the signals this product reads, not the world. A rapid change nobody discusses, nobody trades on and nobody reports is invisible here.
• It says nothing about why anything is moving, and nothing here can express a cause.

Disagreement — how far independent observers differ

How far apart independent communities are about the same subjects. Published only where more than one independent observer exists, and absent otherwise.

Where the Disagreement figure comes from — the short account

How widely independent communities' overall tone is spread on the same subjects, scaled against the reference below.

**An observer is a community, and every discussion source that reported on a subject contributes its own.** A platform organised into named rooms contributes each room; a platform that is one undivided conversation contributes itself, and counts as an observer like any other — so two platforms differing about a subject is a disagreement this figure can see. Each platform then carries an equal share of the comparison, and within a platform its rooms carry as much as they each contributed, so the figure follows how far apart the communities are rather than how much any one of them posts.

**It is the spread across all of the communities, not the gap between the two furthest apart.** Those are different claims and only the first is what this figure says. Six communities reading a subject much the same way and one reading it very differently produces the same gap as a genuine two-way argument, and the gap also widens on its own as more communities are read — the furthest of many is further than the furthest of few, whatever the world is doing. Spread has neither property: a community that agrees with the rest brings it down.

**Below two independent observers no figure is published at all**, with a stated reason. There is no arithmetic anywhere in this method that can produce a figure from one community — not a small one, not a zero.

Subjects where the communities agreed are counted in, not left out. A subject is included as soon as more than one community has been read on it, whatever they said; excluding the ones that turned out to agree would decide the answer by choosing the question.

Disagreement between forecast sources is not part of it. Comparing two sources' answers requires deciding that they are answering the same question, which this product does not do; it is left out of the list rather than left in it scoring nothing, because a part that is permanently unmeasurable reads as a pipeline that keeps failing.

What this does not say

• **This figure is thin and was absent entirely until a second discussion community became reachable.** It rests on the subjects where more than one of them was read at all, which on this deployment is a small share of them, and each community's tone on such a subject can rest on a handful of comments. The published measurement quality figure beside it is the honest reading of that.
• **Where it is absent, that is the measurement.** A zero would be the strongest claim this product could make — that the world's observers agree — inferred from having asked one of them.
• Communities are not populations. Two online communities differing tells you about those two communities.
• Absence of measured disagreement is never evidence of agreement, here or anywhere else on this site.

Data conditions — how well this product can currently see

How well founded the available measurements are, how much of the intended source universe is contributing, how current it is, and how broadly the sample is drawn. None of this is a reading of the world.

**Measurement quality** is how well founded the readings are — the confidence each of them worked out for itself, added up and divided by the number of readings this method defines rather than by the number that survived. A reading that could not be made counts as nothing rather than dropping out of the average, so making fewer readings can only lower this figure. It is the one part of the old uncertainty figure that described this product rather than the world, published in its own right and the right way up: it rises when the evidence is solid rather than when it is thin.

**Coverage** counts the sources contributing against the sources this product intends to observe, and the denominator comes from the source register rather than from what arrived. That is the difference that matters: a source which is configured, expected and silent contributes nothing anywhere, so a figure computed from what arrived could never notice it — and noticing it is this block's whole job.

**Freshness** is the worst state among everything that contributed, so one lagging layer is visible rather than averaged away.

**Representativeness** is how broadly the sample is drawn, over four factors: how many independent communities were read, how many different kinds of evidence stand behind the figures, how much of the scored discussion is outside its own largest language, and how evenly the subjects are spread across the world's regions. A factor with nothing to measure is left out rather than scored — a set of subjects that are all global in scope is not a narrow one — but a factor with nothing to measure *because the sample is missing* scores zero, since a sample with no public discussion in it is not drawn broadly across communities or languages by any reading.

What this does not say

• **None of these figures may change a reading of the world, and none of them does.** The world figures are calculated first and handed to this afterwards, so there is no expression in which one could reach the other. If this product's sight worsens, the world figures narrow and say what is missing; they do not move.
• **Nothing here may improve when this product sees less, and until method version 4.0.0 two of these figures could.** Both were averages over what was published, and an average whose denominator is what survived cannot notice what did not: withdrawing the least well founded reading raised the quality figure, and losing the whole discussion layer removed the two representativeness factors this product is narrowest on and so raised the level. Measured on a test world, the discussion layer going dark took quality from 64.84 to 68.90 and representativeness from moderate to broad — the block whose whole job is to report this product's blindness reporting better sight at the moment it acquired some. Both denominators are now fixed, and every source and every layer is removed in turn on each build to check that no figure here rises.
• **Measurement quality is not representativeness.** A classifier can be highly confident about one community's comments while that community is a poor stand-in for anyone. Those are different claims, and they are shown side by side precisely so that a high quality figure next to a narrow sample reads as what it is.
• There is no history and no change figure here, because nothing stores yesterday's data conditions.
• Coverage counts sources, not what the sources cover. Ten feeds of the same wire report are ten sources and one witness.

How well founded each figure is

A separate score from 0 to 1 for each of the four, built from how much of what is tracked could be measured, how much of it there was, and how well founded each underlying figure was.

The three parts are combined so that a weak one drags the whole score down rather than being averaged away by strong ones. A figure computed from four questions out of four hundred is not two-thirds trustworthy for those four being excellent.

The support part reads the confidence each underlying figure published for itself, weighted by the share it carried. Confidence here is therefore inherited as well as computed — a pulse built from figures that were themselves unsure is unsure.

Confidence never sees the figure it describes. A dramatic reading with thin support stays low confidence, and the code cannot do otherwise, because the value is not passed to the calculation that produces this score.

What this does not say

• A high confidence score says the evidence behind a figure is solid. It does not say the figure is important, or that what it describes matters.
• It cannot detect a systematic problem shared by every source — a platform-wide outage that leaves consistent-looking data still scores well.

How current the figures are

A figure is marked as possibly out of date when most of what it rests on is older than the limit for its layer.

Each figure is a fold over the readings behind it — hundreds of contracts, or every topic being tracked. How current it is, is measured as how much of it rests on readings past the limit for its layer, counted by how much of the figure each reading accounts for rather than by how many readings there are.

The share is the rule rather than the single oldest reading, because these figures fold hundreds of things at once. Taking the oldest would let one lagging contract out of nine hundred describe all nine hundred, and the chance that nothing at all is lagging falls as more of the world is tracked — so the marker would end up reporting how much is being watched instead of how current it is.

Nothing is lost by that. A layer that has stopped publishing has all of its weight behind the limit, so it is still marked, and for the reason a reader would give. A single source going quiet inside a layer is named separately, in the coverage below.

The limits differ by layer because the layers publish on clocks a factor of eight apart. A single limit would either mark the pulse out of date every afternoon or never mark it at all.

Age is measured now, against the period a reading covers — not taken from what the reading said about itself when it was written. A reading written late says so for the rest of its life, even while it carries the most recent data there is; what it knew about its own source's update rate is what the separate limits above are for.

What this does not say

• A current figure is not a live one. The fastest layer here updates in minutes and the slowest once a day.
• Quiet is not the same as behind. A topic nobody discussed for a day produces no new reading, and the weighting is what keeps such a topic from marking a figure out of date on its own — but a layer that is quiet everywhere and a layer that has stopped look alike from here, and the coverage below is where they come apart.

Rounding and versions

Figures are rounded before they are stored, and every stored figure carries the version of the method that produced it.

Rounding happens before storage, and the parts of the confidence score are rounded before they are combined, so a published breakdown reproduces the score beside it rather than nearly reproducing it. Published shares are adjusted so that they still add up to the whole they are shares of.

One version covers all four figures, because they are one composition rather than four methods published together: the uncertainty figure reads the other three, and versioning them apart would let a stored figure claim a method its own inputs were not calculated under.

When any number on this page changes, the version changes with it, and every figure calculated under the old version keeps saying so.

What this does not say

• The precision here is far beyond what the underlying data supports. It exists so that a change to the method is visible in review, not because the last decimal place means anything.

Biggest movers

Which subjects' figures changed most over a period, one measurement at a time and never across measurements.

One tab is one measurement. Attention is ranked against attention and sentiment against sentiment; the panel has no way to place a change in one beside a change in the other, because the two are not the same quantity and there is no exchange rate between them.

Rows are ordered by the size of the change. On most measurements the direction does not affect the order, so a fall of ten places above a rise of six.

On the measurements listed below, a fall is not shown at all — only a rise is. Attention is scored against the busy side of a subject's own history and not against the quiet side, so 'quieter than usual' would be a size measured against a comparison nobody established. The quiet side also has a floor the busy side does not: an hour in which a normally busy subject was mentioned by nobody is as far down the scale as this product can go, and that happens constantly. Ranked by size alongside rises, those hours would be most of this panel most days.

A change smaller than the threshold below is not listed. With enough subjects something has always moved by a rounding error, and a panel headed 'biggest movers' showing one would be saying nothing while looking as though it were.

Confidence does not move a row up or down the list. A change we are not confident in is listed at its size like any other and carries its confidence where it is drawn, so that what the panel is ordered by and what it is headed by are the same thing. Suppressing such a row would hide what the confidence figure exists to disclose; ranking it beneath smaller ones would mean the first row was not the biggest change.

What this does not say

• A large change is not an important event. It is a change in one measurement of one subject, and the reason for it is not something this panel knows.
• The comparison is against the same subject earlier, never against other subjects. Two subjects both listed here moved a lot for themselves; that is all the pair of them have in common.
• Where only rises are shown, an absence from the list is not a statement that nothing fell. Subjects do go quiet, and this panel is not where that is reported, because it is not measured well enough to be ranked.

• attention

Which measurements are listed here only when they went up. On these, a subject going quiet is not published as a change at all — see the section's own limits for why the two directions are not treated alike.

Fixed by this product's own specification.

Trending subjects

Which subjects several independent kinds of source are unusually busy with at the same time — deliberately not a measure of how much is being published.

Five inputs are scored separately: public discussion, people looking a subject up, reading about it on this site, a change in what is expected about it, and how many independently-sourced reports have covered it. Each is scored from nothing to full against that subject's own history, so a small subject having an unusual week can score above a large one having an ordinary one.

The five are combined using fixed shares that never grow. An input with nothing to say scores nothing for that input and its share is not handed to the others — so a subject that only newsrooms have written about reaches the coverage share and stops there, and cannot lead this panel.

Coverage is counted as independently-sourced reports, never as articles. Twenty outlets carrying one wire report is one report carried twenty times.

The score is then increased for each additional kind of source that is also moving, up to a limit. It is a multiplier rather than an addition, so several kinds of source agreeing that nothing is happening still produces nothing.

Finally the score is reduced where the activity behind it came from unusually few accounts. That is a statement about concentration and never about who the accounts belonged to.

What this does not say

• Trending is not importance. A subject can matter enormously and be entirely unsurprising, in which case nothing here will list it.
• It cannot see a subject nobody in these sources discusses. A conversation happening somewhere this product does not read is invisible here, and this panel would report it as quiet rather than as unmeasured.
• Reading about a subject on this site is one of the five inputs, which means this panel is partly measuring what this site has been showing. That is why it carries the smallest share of the five.

Where sources disagree

Subjects where how much was published and how much was discussed are far apart, or where two communities feel materially differently.

Both sides are required. One side reporting is an absence rather than a disagreement, and publishing it against an assumed-normal other side would turn a source being unavailable into a finding.

Each side has to be behind a minimum number of items before it is published. The number of items is shown either way, so a reader can see how thin a comparison is.

The card states which side is ahead and nothing about why. 'More was published about this than was discussed' is a description of two figures; anything about what that means would be a claim about the world that nothing here measured.

Where two or more communities discussed a subject, how far apart they are is reported as well, against two conditions rather than one: how far the extremes are apart, and how spread out the communities are overall, so that a single outlier among many is not reported as a split. Where only one community was measured, no figure is published — a sample that cannot disagree with itself has not measured agreement.

Those two numbers are the ones the cross-signal card on the same page uses, read from where that method declares them rather than restated here. Two thresholds for one measurement would let the two panels disagree about one subject in front of one reader.

What this does not say

• It compares how much, not what was said. A gap between coverage and discussion says nothing about whether the two are saying the same thing.
• Today the discussion side is one community, so the second half of this panel publishes nothing at all. That is a limit of which sources are available rather than a finding about agreement.

The world by category

The same four figures the top of the homepage publishes, calculated over the subjects in one category.

Each figure is the World Pulse calculation run over the subjects in that category rather than over everything tracked, so the four numbers on a category and the four at the top of the page under the same chip are one calculation described twice. How each of them is arrived at is in the World Pulse sections of this page, and the version that produced it is published beside it.

A category with fewer measured subjects than the minimum publishes no figure at all, and is still shown with its subject count. One subject wearing a category's name is not a category.

Each figure states how many of the category's subjects are measured in the layer behind it, so a blank figure can be read as thinly measured rather than as broken.

The figures are calculated as the page is built and stored nowhere, so no change over time is reported for them: a change is measured against an earlier published value of the same figure, and for a category there is none.

What this does not say

• Every subject's figure is relative to that subject's own history, so a category's figure is how unusual its subjects are right now — not how much attention the category draws. A category of normally quiet subjects all having an unusual week will read higher than a permanently busy one having an ordinary one.
• The categories are a fixed list chosen when the product was designed. A subject can belong to more than one, and it then contributes to each — the figures are not shares of a whole.
• A subject counts toward a category's figure when it is in that category, and a forecast question counts when any subject it is linked to is. A question linked to subjects in two categories therefore contributes to both.

Precision

How precisely the figures on these panels are published.

A score is published to the same number of decimal places as the figures it is folded from, so nothing on these panels implies a precision its inputs do not have.

Shares of a score are published to more places than the score itself, because several small shares are compared against each other rather than read on their own.

What this does not say

• Rounding is applied at publication, not during the calculation, so two figures that round to the same number were not necessarily equal.

Reading one item

How a single comment or post becomes a reading of positive, neutral or negative, and what is recorded about the model that read it.

Each item's text is cleaned of quoting and formatting, its language is identified, and what is left is scored by a classifier. Nothing about an author is read, and the text itself is never stored — both platforms this reads permit deriving figures from their content and not keeping it.

Where a classifier reports a distribution across grades rather than one grade, the whole distribution is used: an item split almost evenly between the top two grades is mildly positive rather than confidently so.

Where the language of an item cannot be identified confidently, it is counted and not scored. Filing it under the most likely language would put a reading into a figure the reader is told is about a language it is not.

The identity and version of both models are stored with every figure they produced, as part of the version that figure carries. The version published on this page is the calculation's alone, because the process that serves this page does not run the models — the figure itself names the pair that read it.

What this does not say

• A classifier reads the words, not the meaning. Irony, quotation and reported speech are read at face value, and there is no measurement here of how often that is wrong.
• The classifier running today has not been calibrated against anything. Every figure it produced says so, in a mark on the figure and in the version it carries.
• Neutral is a reading. Being unable to read something is not one, and the two are never combined.

Which items are counted

The items that reach a subject's figure, and what happens to the ones that do not.

Items are set aside in a fixed order: near-duplicates of each other, then items whose shape suggests spam, then items not confidently about the subject, then items in a language that could not be read. Each item is set aside for the first reason that applies, so the reasons add up to the number of items fetched exactly and a reader can check the arithmetic.

A spam-shaped item is reduced in weight rather than removed. Removing it would move the published split by exactly the amount it carried, with nothing on the figure recording how much that was.

The count of items considered is per subject and not per period. An item that reached three subjects was screened three times, once for each, because whether an item is about a subject is a question asked of a pair rather than of the item alone.

What this does not say

• Nothing here judges an item's quality or its truth. The only questions asked are whether it is a duplicate, whether it looks automated, whether it is about the subject, and whether it can be read.
• An item set aside as spam-shaped is a statement about a pattern, never about who wrote it. No account is profiled and no author history is read.

How much one item counts

Every item carries a weight, so that a period is not decided by whichever item happened to be popular.

An item's weight rises with how much engagement it drew, on a curve that flattens: past the reference below, more engagement adds very little. Without that, a single item with a thousand responses would count a hundred times one with ten.

A second, separate limit then applies: no single item may hold more than a fixed share of a period's total weight. The two do different jobs and neither replaces the other — the curve bounds a busy period, where hundreds of items make any one of them unable to dominate arithmetically, and the share limit bounds a quiet one, where a single submission genuinely can be most of an hour.

Trimming one item lowers the total, which lowers the ceiling, which can bring a second item over it — so the limit is re-applied a bounded number of times. Where fewer items were scored than the share allows for, the ceiling is raised to an equal share, because a ceiling nothing can reach would drive every weight towards nothing.

Where the limit bound anything, the figure says so.

What this does not say

• Engagement is how much a platform's readers responded to an item, which is not how many people agreed with it. A heavily argued-with comment and a heavily agreed-with one look the same here.
• The weight is used inside one platform's own figures. It is never used to weigh one platform against another — platforms are combined equally, further up.

What is published for a period

The split between positive, neutral and negative, the communities behind it, and a single tone figure that is a plain restatement of the split.

What is published for one subject, one source and one period is the whole split, together with the emotional grouping where the classifier could produce one, the languages the items were in, and the communities they came from. Each community's own tally is kept rather than averaged away.

The single tone figure is the weighted mean of the item readings, on a scale from −1 to 1. It is a summary of the split rather than a replacement for it, and both are published together so that a reader can see the split a summary came from.

A period in which nothing could be scored produces no figure at all. A distribution over no items is not a neutral distribution.

What this does not say

• It is the tone of what was said on the platforms this product may read, in the period stated. It is not what people think, and where only one community was measured the figure describes that community and nothing wider.
• The tone figure and the split can look like a rating of the subject. They are neither: a negative reading about a disaster is a description of how it was discussed.
• There is no single 0–100 sentiment figure per subject anywhere in this product, and that is a decision rather than a gap. See the boundary at the foot of this page.

How far communities are apart

Whether the communities discussing a subject in a period agreed with each other, where there is enough from each of them to say.

Two figures are computed across the community averages: how far the extremes are apart, and how spread out the communities are overall. The second is what other parts of this product read, because a range is decided by two communities and discards every other one.

A community enters the comparison only once it has said enough for its own average to be a measurement. Below that its items still count towards everything else — the split, the sample size, the tone figure — and it simply gets no vote in a statement about how far communities are apart.

Where fewer than two communities clear that bar, no figure is published and the reason is stated on the row: one community measured, or several too thin to compare. The two point at different repairs — a wider set of sources, or a wider period — and telling them apart is the whole reason they are separate marks.

Nothing is ever published as zero disagreement on the strength of a single community. A sample that cannot disagree with itself has not been found to agree.

What this does not say

• It compares the tone of what different communities said, never the substance. Two communities can be equally positive about entirely different things.
• A community here is a room on a platform, not a group of people. Two rooms on federated platforms are two communities even where the same people are in both.
• On the shortest period this product publishes, the figure is usually absent, because a single hour rarely gives two communities enough items each. That is the honest state rather than a fault — the daily period is where the figure becomes possible.

Change against the subject's own history

How far a period's tone sits from what that subject's discussion has usually been, in its own units and in standard deviations.

A period is compared against the periods of its own width immediately before it — hours against hours and days against days. Comparing a day against the hours inside the days before it would report a difference that is a property of the two widths rather than of the discussion.

The comparison is published both in tone units and as a number of standard deviations of that history, so a small change on a subject whose discussion is normally steady is not read as the same thing as a small change on a volatile one.

Where too few periods have been measured, no comparison is published and the figure is marked as having no history yet. It is never published as no change.

There is no threshold here for what counts as a large change, and that is deliberate: no set of periods has ever been labelled, so a threshold invented here would become the number every alert downstream was built on.

What this does not say

• A large change against a subject's own history is not a large change in the world. A quiet subject's ordinary week can produce one.
• The history is this product's own record of that subject, so it begins when this product began reading. A subject added last week has no history and no comparison.

The periods a figure is published for

An hour, a calendar day, and the day up to now — the last two assembled by adding up the hours, because the text they were read from is gone.

The hour is what both sources are read at. The two day-long periods are built by adding together the stored hourly figures: what was classified was text, the text is not kept, and so a wider figure cannot be produced by reading anything again. It is the one calculation in this product whose input and whose output are the same table.

Two day-long periods rather than one, because they answer different questions. The calendar day is a day that finished and does not move again, which is what a comparison across different kinds of source needs. The day up to now slides forward every hour, which is what a page describing today needs — a calendar day is frozen between midnights.

A wider period is assembled only once the hours inside it have had time to arrive, and is then written once and never revised. The delay below is what makes that safe: one assembled early would be missing its last hours for good. What that delay is waiting for is mostly this system catching up after an interruption of its own rather than either platform being slow — both were measured, and both finish delivering an hour within about a minute of it closing.

Adding up hours never crosses a change of model. Where the classifier changed part-way through a day, the hours before and after are assembled into separate figures rather than averaged into one, because two classifiers' readings are not comparable.

The day up to now is compared against calendar days rather than against other sliding days. Sliding periods overlap each other, so a history built from them understates how much the tone normally varies and would inflate every comparison measured against it.

What this does not say

• A day's figure exists only where the hours under it were measured, and only where they were measured recently enough for the routine run to reach them. Assembling the days over older hours is a deliberate operation somebody runs, not something a release does on its own, so a subject can have hourly figures reaching further back than its daily ones.
• The number of distinct people is not carried up from the hours. Author identifiers are discarded with the hour that produced them, so adding them would count one person posting each hour as twenty-four people. A day says plainly that it does not have that number rather than reporting a wrong one.
• The delay below rests on how long each source is held for after a period closes. Both of those delays have now been measured on the platforms themselves, and the longer of the two is what sets it — so shortening the other one would not bring a day's figure forward. It keeps a margin beyond what either measurement needs, because a period assembled early is missing part of itself for good while one assembled late is only late.

• single_community
• unscored_languages
• high_noise
• engagement_capped
• uncalibrated_classifier

The marks that carry from an hour onto the day assembled from it. Each is a fact about how the evidence was produced rather than about the period, so a day that dropped them would be claiming its reading had never been trimmed and had been scored by a settled model. Every other mark is worked out afresh at the wider period, from that period's own numbers — which is the only reading under which what it says is true of the row it is on.

Fixed by this product's own specification.

How much of the discussion was read

What share of the items fetched could actually be scored, and in which languages.

Every figure publishes how many items were fetched, how many were scored, and the share of the sample each language accounted for. Below the share stated here, the figure is marked as covering only part of what was said.

That mark is set on most figures today and is meant to be: the classifier reads a limited set of languages and the platforms are not written entirely in them, so a substantial part of most periods is counted and not scored. It is a limitation of this build, published as one rather than hidden by counting unread items as neutral.

The emotional grouping, where it is published, covers less than the tone figure does — it is available in one language only — so the two coverage figures are published separately rather than as one number.

What this does not say

• Coverage is the share of what was *fetched* that could be read. It says nothing about what share of the world's discussion of a subject these platforms are.
• An unread item is not evidence of anything. It is not counted as neutral, and it is not assumed to resemble the items that were read.

Confidence

How well supported a period's reading is — never how strong, how negative or how interesting it is.

Four parts are scored from 0 to 1: how many items were scored, what share of the fetched items they were, how far apart the communities were, and how recent the reading is. Each is published individually with the share it carried.

They are combined as a weighted geometric mean rather than an average, so one part near zero pulls the score down instead of being averaged away by three healthy ones. A reading of six comments is not two-thirds trustworthy for being recent, well covered and unanimous.

A part that cannot be scored has its share spread across the parts that can. Where only one community was measured, agreement is not scored as one — nothing was found to agree, and scoring it as agreement would make the least informative case look like the best one.

What this does not say

• Confidence has no access to the reading it describes. A dramatic swing on a low-confidence period stays low confidence.
• It says nothing about whether the discussion was right, informed, or representative of anyone outside it.

Freshness

Whether a reading is current, out of date, or absent.

A figure describes a period that has closed, so it is never revised once written. What ages is the gap between the newest figure and now, measured against how often that source's periods arrive.

A figure is marked as possibly out of date after the number of missed collections below. One missed collection is ordinary operation, and a mark that flipped on it would mean nothing.

A wider period is judged against its own width rather than against the hour it was assembled from. Judging a day against an hourly schedule would mark every day out of date the moment it was written.

What this does not say

• A source that stopped reporting leaves its last figure saying what it said. There is no figure claiming the discussion went quiet, because nobody measured that.
• Out of date and absent are different states and are shown differently. A chart draws a gap where nothing was measured, and never a line across it.

Precision

How far each published figure is rounded, and why the rounding is part of the method.

Rounding is applied before a figure is stored rather than on the way to a screen, so the stored figure and the published figure are the same number and can be recomputed from the stored counts exactly.

The parts of a split are published to more places than the figures themselves, because they are apportioned to sum to exactly one and the remainder has to land somewhere.

What this does not say

• Precision is not accuracy. A tone figure published to four places is a weighted mean of readings that are themselves uncalibrated.