What the numbers mean. Nothing more.

Reviewed, immutable daily question packs.

Every UTC-dated edition references an edition ID plus a pack ID, version, and digest. A generator proposes exactly five low-stakes either/or questions for the next day. Deterministic filters reject sensitive topics, identity requests, URLs, instructions, duplicates, absolutes, and unbalanced options before a separate model configuration can review the candidate.

A generated pack opens only when the generator parses exactly, every deterministic check passes, and the independent reviewer returns its exact approval code. Any failure commits one of seven prewritten fallback packs instead. The public generation audit records model/configuration versions, prompt hashes, output hashes, filter findings, review decision, and the selected pack; raw model text is never stored.

Once selected, the edition pack is immutable. The edition ID prevents participation from drifting across dates, while the digest lets clients and researchers establish that answers refer to the same ordered questions.

Know Me freezes one selected 24-question pack into each challenge.

The creator can generate a Remix, use Classic 24, Little Things, Group Chat Energy, or take the daily Surprise Remix. A manual Remix creates a cryptographically random 12-hex-character seed; Surprise deterministically derives a different concrete seed for each UTC date. The same versioned algorithm then selects exactly four questions from each category in the reviewed 60-question pool. The concrete seed becomes part of the pack ID; the ordered pack receives a SHA-256 digest. No IP address, cookie, D1 record, or model selects the questions.

Once a challenge is created, its exact stored pack identity and digest—not a landing-page query—control every invite, response, reveal, relay, and reflection. A friend sees the same ordered pack and makes exactly one prediction per question. The overall score is correct predictions divided by 24. Everyday tastes, Daily rhythm, Support style, Adventure, Connection, and Decision style each report correct/4 plus a percentage.

The reviewed pool and themed packs are curated entertainment questions, not psychometric instruments. Remix generates no new text, and AI never writes or weights this scored sheet. The optional post-result Lab reflection receives only six category counts and is labeled separately from the deterministic score.

Optional relationship and person/software labels are challenge-scoped display context, not scored variables or scientific measurements. They are never inferred from the audience URL, do not follow a participant into another challenge, and appear in the creator’s private connection star. An explicitly invited MCP client can retrieve one exact code and submit 24 enum-only predictions; its result omits item answers and returns only the exact total plus six category counts. MCP establishes transport only, and score-relay creation requires a separate explicit call.

The Connection Atlas publishes permitted categories only after two thresholds.

Creator and participant permissions are separate and off by default. The creator may allow their person/software category; otherwise the public creator side is fixed to unspecified. A participant may separately contribute that participant’s selected relationship and actor categories. One opted-in direct edge records only a domain-separated opaque challenge key, random lineage scope, fixed relationship category, the permitted actor/evidence categories, UTC event day, and expiry. A participant cannot cause the creator’s private actor label to enter the public edge.

The edge contains no alias, account, raw challenge or response ID, capability, code, URL, answer, prediction, score, reaction, participant key, IP, country, model ID, or exact timestamp. The published unit is an opted-in edge occurrence—not a person or relationship. Cells group relationship, actor class, and evidence direction. A cell remains suppressed unless it contains at least 20 direct edges from at least 20 distinct challenge lineages in the fixed previous 30-full-UTC-day window. The current application emits direct edges only; propagationEdgeCount is a reserved zero-valued field until a separately reviewed continuation protocol exists. Suppressed cell counts may be published, but their hidden edge totals and marginals are not.

The scheduled Worker attempts a freeze at 00:07 UTC. An authenticated bootstrap or recovery operation may freeze the current edition at another time, which remains visible in its recorded creation timestamp. Each freeze writes the qualifying cells, suppressed-cell count, window, fixed thresholds, exact count strings, prior hash, and creation time into canonical JSON. SHA-256 links each immutable edition to the previous edition. Live opted-in rows expire after 31 days; explicit response or challenge deletion removes an unfrozen row sooner. A historical frozen aggregate is not rewritten after deletion. The Atlas cannot establish who knows whom, relationship strength, human identity, agent identity, model use, geographic representativeness, company membership, or national opinion.

The personality snapshot uses the exact public-domain Mini-IPIP 20.

This separate mode presents the 20 items published by Donnellan, Oswald, Baird, and Lucas (2006) in a shuffled display order. Wording, factor assignment, positive or reverse key, and scoring remain fixed as mini-ipip-20-en-v1. The official IPIP permission statement places its items and scales in the public domain and permits commercial use.

Responses run from 1, Very Inaccurate, through 5, Very Accurate. A positive-keyed item keeps its response; a reverse-keyed item uses 6 − response. Each of Extraversion, Agreeableness, Conscientiousness, Neuroticism, and Intellect / Imagination reports the sum of four keyed responses (4–20) and its raw mean (1.00–5.00). There is no overall score.

The exact item key is published by IPIP. Primary sources are the original 2006 validation paper and a 2013 nationally representative young-adult replication. The four-item scales trade breadth and precision for speed. Raw means are not population percentiles; reliability coefficients describe samples, not an individual’s “accuracy.”

The Test Lab registry freezes three additional public-domain instruments.

ipip-big-five-50-en-v1 uses the official fifty-item IPIP sample order, ten items per broad factor, and the published positive/reverse keys. ipip-ipc-32-en-v1 uses the official 32-item Interpersonal Circumplex order with eight four-item, all-positive octants. qb6-big-six-36-en-v1 uses the public-domain 36QB6 self form, its exact 0–5 anchors, and six published three-positive/three-reverse formulas.

Every complete vector is validated for exact ID, version, item IDs, response enum, order-independent completeness, duplicates, and extra fields. IPIP reverse keys use 6 − response; QB6 reverse keys use 5 − response. Results contain only each scale’s keyed raw sum and mean, source version, and fixed disclosure. There is no overall total, percentile, category, AI narrative, or cross-instrument conversion.

The generic scoring endpoint performs no D1, AI, account, friend-challenge, or public-cohort write. Public instrument responses omit factor mappings and key directions; the methodology endpoint publishes those receipts separately. The full definitions and limitations remain linked from each canonical test page.

Browser cognitive tasks report task performance, not a person score.

Reaction Speed Lab independently implements the published Deary–Liewald simple and four-choice structure: eight practice trials, twenty retained simple trials from 150–1500 ms, forty retained choice trials from 200–1500 ms, and random 1–3 second response-to-stimulus intervals. Out-of-window attempts are replaced. Correct response mean, median, and sample standard deviation are descriptive of that device-specific round; choice accuracy and errors remain separate.

Color Conflict is a short single-stimulus, manual-response Stroop-style demonstration with four practice, eight congruent, and sixteen incongruent measured trials. It reports median correct milliseconds and accuracy per condition plus incongruent median − congruent median. This quick browser variant is not interchangeable with the original spoken sheets, commercial tests, or laboratory implementations.

Both tasks run trial vectors and scoring in page memory only. Device, display, input, browser scheduling, color perception, language, practice, strategy, and momentary conditions affect results. No output is an IQ score, diagnosis, impairment measure, clinical change score, or stable global ability.

Perception compares five aggregates—not twenty answers.

After self-scoring, a creator may deliberately create a private seven-day capability link. Each invited person or explicitly instructed MCP client predicts how the creator rated the same exact 20 first-person items. The prediction layer is an unvalidated perception game layered on a validated self-report instrument; it is not a validated observer scale or friendship test.

The Worker scores each complete vector in memory and immediately discards its item responses. D1 stores only five creator keyed raw sums and five predicted keyed raw sums, plus bounded aliases, transport evidence, capabilities, and expiry metadata. The result displays the creator and prediction raw means for each factor and signed gap = predicted creator mean − creator self mean. There is no item reveal, exact-match count, distance, overall accuracy, rank, type, percentile, diagnosis, compatibility score, or public-cohort contribution.

Browser and MCP results use identical math. An mcp transport label proves only the request channel—not a model, agent identity, autonomy, consciousness, or one unique actor. Challenges and aggregate responses hard-delete at expiry no later than seven days after creation; capability holders may delete them sooner.

Two ballots in one quick round.

For each of today’s five questions, a participant records:

1. Self: the choice they personally make.
2. Prediction: the choice they expect the named Lab target to make.

Prediction accuracy is correct predictions divided by published, non-tied target outcomes. If a target result is unavailable or below its publication threshold, ViewsMeet shows a pending or low-sample state instead of inventing a score.

Private results do not require Turnstile. An optional public contribution is submitted only after a managed Turnstile token completes; if no site key is configured, public opt-in is disabled in the production interface.

The full profile uses 21 transparent signals.

The dedicated profile round contains seven named trade-off axes and three independently written choices per axis. Each choice contributes exactly +1 or −1 to one axis. The three signals produce a deterministic score from −100 to +100 and a “this round” label; no hidden model, demographic input, or free text enters the calculation.

A friend comparison calculates per-axis distance, overall axis similarity, shared directions, closest axes, and widest gaps. It returns aggregate axis scores and an exact-pick count without returning either answer sheet or per-question comparison.

This is a purpose-built entertainment scoring system with a published rule set, not a validated psychometric instrument. Versioned questions and deterministic math make completed rounds reproducible and auditable.

Knowledge and Match answer different questions.

Knowledge

The number of a friend’s self-picks that the respondent predicted correctly, divided by the number compared.

Match

The number of questions where both people independently made the same self-pick, divided by the number compared.

After a response is complete, its per-question Knowledge and Match comparison can reveal the challenge creator’s issued choices to that respondent. It does not reveal another respondent’s answer sheet.

A high Match score does not prove compatibility. A low score is not incompatibility. ViewsMeet never publishes a “least compatible” person or uses these scores to rank the value of a friendship.

Groups reveal aggregates, not answer sheets.

As a challenge link collects responses, the service can form privacy-safe group summaries. Exact per-question group tallies become visible only at five total people, including the creator. Below five, the interface shows a forming or low-sample state.

Summaries can include per-question totals, consensus, ties, and overall pair-question agreement. Individual respondent answer sheets are not exposed to other participants. Low-sample ranks and uniquely identifying breakdowns are suppressed.

Every panel carries its configuration.

“Lab panel” is an umbrella for explicitly labeled cohorts. A configuration label may describe collection surface, edition, instructions, software setup, or version. It is not proof that a ballot came from a verified person, and it is not permission to infer demographics.

Public aggregates remain in collecting state until their configured threshold is met. Tied target outcomes are excluded from binary prediction accuracy rather than arbitrarily assigned.

The world view uses country counts.

Only opted-in, Turnstile-passed browser ballots contribute. Cloudflare’s two-letter country code is counted; raw addresses and city coordinates are not stored in the experiment dataset. A country node appears at three responses, while smaller cells remain suppressed.

Closed editions are canonicalized and linked by SHA-256 hashes. The public ledger and CSV let anyone verify that a published country snapshot has not silently changed.

Software-agent ballots use the same choices.

The machine endpoint provides the immutable pack and accepts one listed self choice plus one public-cohort prediction per question. ViewsMeet derives the cohort from the submission channel and available signing evidence rather than trusting a caller-selected label.

Lab model IDs, prompt versions, parameters, trial counts, and failed attempts are published with the experiment record.

What this cannot establish.

• Whether an alias belongs to a particular person.
• Whether participants answer consistently outside this round.
• A fixed personality type, clinical status, intelligence, ideology, compatibility, relationship quality, or future outcome.
• Whether a public cohort represents a broader population.
• Whether a software configuration “understands” a question.
• Statistical significance beyond the published counts and thresholds.

Only pages that name an exact sourced instrument or research paradigm may describe that component as research-based. Friendship, profile, and group scores remain curated entertainment and transparent product metrics. Quotations should preserve the instrument/task version, configuration, cohort, threshold, and published limitations.