The method

01 . The standard

A reading is a claim about a public, and every claim here carries the apparatus that would let someone else disagree with it precisely.

The front door puts the difference in two sentences: social listening, sold on, is closer to surveillance, and an insight carried back into the room it came from is the root of acquaintance. Here is what stands behind them. A population is read, the reading goes to whoever can pay for it, and the people counted never see the count — and nothing in the method requires that. The method does not change when the reading is published back instead of sold on; who ends up holding it does. So the corpus, the instruments and the daily sheet are open, at k ≥ 20, to anyone.

Four rules govern anything published on this host. They are not style preferences. Each one has a gate behind it, and one of them runs on every deploy.

• Every number carries its spec line

  The query, the window, the run, and the code path that produced it. A number without a provenance line is not publishable. This is the rule that makes the rest possible, because it is what turns a disagreement into a reproducible one.

• Heard and counted are different marks and never merge

  A number a human read off a surface is heard. A number the pipeline computed is counted. Most current dial numbers are heard, and the surfaces say so rather than rounding the distinction away. The rung a number files at follows from the mark: what a person read off a surface files as reported, what the pipeline computed files as thin or measured by the depth of its baseline.

• Failure renders like success

  Gaps, failed harvests and thin coverage appear at the same size and in the same voice as the wins. A source that returned nothing says so, in place. A page that only shows what worked is a page that has been edited into an argument.

• Copy reports the harvest

  Surfaces report counts, deltas, timestamps, dial values and resolutions. No analyst voice, no adjective doing the work of evidence. Specimen copy simulates the harvest, never the analyst. A lexical gate runs on every deploy and refuses the known offenders.

02 . Reproduction is the test

The claim a reading makes is not "trust this number". It is "run this and get the same number". Every published reading prints the query beside the figure so the second thing is possible.

The final step of reproducing a reading is a diff against what was published. Reproduction that stops before the comparison is a tutorial, not a proof.

This is also the contribution route that is open today. A disagreement about arithmetic is a factual dispute rather than an authored work, so it needs no licence from you and no settled terms from us. If your diff disagrees with ours, send it.

03 . Rule sets, not demographics

A public is described here by what it does and the codes it runs on: its lexicon, what it treats as trying too hard, which references mark membership, what it turns away. Age bands and income brackets describe the same people to somebody else, and a reading can carry both; the rule set is the part it is built on.

This follows from the gate in the mission rather than from taste. Demographics describe a group to somebody else, which is what makes them the natural unit for a market report. Rule sets describe a world to the people inside it, which is what makes a reading survivable when it is handed to its own subject.

The practical consequence is that a persona here is a rule set plus its evidence: every rule in it points at the posts it was read from, and a rule that cannot is filed as a hypothesis until it can.

04 . The worked records

The standard above is abstract. These are the places it is carried out, and they are the pages to read if you want to argue with the method rather than about it.

05 . What the method does not claim

• Not prediction. A reading describes a public as it currently runs. Where a forward statement is made it is a distribution with its assumptions printed, not a number with a date on it.
• Not representativeness. A public that posts is not the same population as a public that exists. Every reading inherits the bias of the places it was read, and the directory exists so that bias is inspectable rather than implied.
• Not a substitute for asking people. What a group says about why it does something is a different object from what it does. Both are evidence. Neither replaces the other, and a method that pretends otherwise is picking the one that is cheaper to collect.
• Not settled. Most current dial numbers are heard rather than counted. That is stated on the surfaces that carry them rather than fixed by relabelling.

06 . A segment read, and what it could not say

The section above says what the method does not claim. This is a run where that was the whole result. On 26 August 2026 the instrument was pointed at one segment, skincare actives, over one home room and 23 control rooms. One of the six Decode layers filled. The other five abstained, and each one says what it would take to fill it.

The file prints what the run does not claim, and the first item is the one that matters: it does not claim the segment has no values, no boundaries and no trajectory. It has all three. The instrument cannot see them yet, which is a different sentence and the only one the evidence supports.

The corpus behind this run is described on the mechanism page, which is where the two blockers are counted across all 105 rooms rather than in this one.

07 . A measurement that returned nothing

The sociolect question is whether a room has vocabulary that is dense in it and near-absent everywhere else. Run on 26 August 2026 over eight weekly windows, 24 rooms and 8,748 distinct terms, the answer was that no room did. The run is kept here because a measurement that returns nothing and says why is evidence about the instrument, and discarding it would leave only the runs that found something.

None of the nine is a coinage. They are common nouns, two ingredients, a brand, a decade and a country: words missing from one 234,456-entry spell-check list rather than words a community made. The run's own headline says the same thing, that no distinctive shared vocabulary was detected in any room.

Read that as the instrument's answer over this corpus rather than as a statement about language. The 24 rooms are general-interest Reddit rooms in advice, confession and explainer formats, and a question format is a different object from a subculture. Forty-two of the 83 candidates come from one room.

08 . Purpose, necessity, proportionality

The Berkeley Protocol on Digital Open Source Investigations asks anyone collecting at scale to state three things in writing: the purpose the collection serves, why it is necessary for that purpose, and what is not collected. It is explicit that the principle favours itemized, manual collection over bulk. This house is bulk collection across eighteen ingesters, so the burden is ours. This section is it.

Purpose. To make social rule sets visible, nameable and changeable by the people living inside them, which is the mission. Every published output of this house is a statement about a set. There is no page, product or deliverable whose subject is a person.

Necessity, which is the part usually left unsaid. A social rule is not observable in an individual. One person using a term tells you that one person used it, not that the term is a norm. A norm is a property of a population, and the smallest object that can carry one is a group. So the floor of twenty distinct accounts behind every published cell is doing two jobs, and this house has only ever claimed the first:

Three things follow, and each is why the line is drawn at what leaves rather than at what is read. You cannot count distinct accounts without counting accounts. You cannot detect a community without an account-level edge list, because a community is defined by who overlaps with whom. And you cannot know a cell is large enough to publish without holding the data the count is performed on. The alternative the Protocol nominally prefers, a hand-picked sample, produces a different object: one whose selection criteria are unstated and whose bias is therefore invisible.

Proportionality. The line is drawn at egress: read individuals, publish only crowds. What follows is enforced rather than intended.

• No individual egress. Handles are quarantined and the author column is empty across the database. Stated exactly, because the exactness is the point: this holds today by grant and by a single routing function every ingester passes through, with a database constraint written and not yet executed. It is not sealed, and calling it sealed would be the kind of sentence this page exists not to write.
• No inferred attribute at person level, at any tier. The corpus may record what an account publicly did. It may never record what an account is. The first is a corpus; the second is a dossier, and pseudonymising a dossier does not change which one it is.
• No cross-platform identity join. The salt is per platform and per window.
• No verbatim that is searchable back to its author. A term, a construction, a frequency. Not a sentence somebody wrote.
• Collection is scoped by room, never by person.

Two places where this is not yet satisfied, printed here rather than left for a reader to find.

• The retention term is open. Minimization argues for a shorter corpus and the mechanism argues for a longer one, because an instrument that cannot see further back than a social listening tool is a social listening tool. That question is tracked and unresolved, and it is the one place minimization has not been answered.
• A floor that holds on one release does not automatically hold across several. Two publications that each clear twenty accounts can be intersected into a cell that does not. The rule against re-identification covers this and no mechanism checks a new release against the ones already out. Recorded as an open finding.

The people this applies to have never interacted with this house. They posted in public rooms and do not know it exists, which is the reason the burden binds at all. The answer is that nothing is ever published about one of them, only about populations they are part of. Someone who does hand this house data has a stronger claim to a stated clock, not a weaker one, and that clock is ruled before the first seat is issued rather than after.