A bet, registered in public, before the run.

Where this stands today

Live in one context. Verdict withheld until earned.

Pre-registered

2026-06-25 (this page); v0 parameters 2026-07-02; gate-vs-skip arms + min-N registered 2026-07-02; decision turned on 2026-07-02

Instrument state

Live in one context. The gate-vs-skip decision is now running for organic-search on the home page; every other visitor sees the gate (canon). Assignment is randomised and CANON is the failure floor, so the engine can only ever remove the gate for a decided slice, never break the page. Posteriors recompute nightly on our own iron. Verdict WITHHELD until min-N.

Decision rule

Demote the gate if, after the registered window and minimum sample, gated-path inbound quality is at or below the skip path.

Measurement

First-party, no PII, no third-party trackers. DNT and the Global Privacy Control are honored absolutely.

Store

Our own iron (sovereign Postgres on the UNI Lab appliance). The website never holds personal data.

The hypothesis

One bet, four pillars, written down before the result is in.

We do not get to move the goalposts once the data lands. So we register them here, dated.

1. 01

   The bet

   An opaque first impression, a small puzzle solved before the page reveals, qualifies the rare high-intent visitor, while losing the wrong volume on purpose. This is conjecture, not faith.

2. 02

   What would refute it (FALSIFIES)

   If, after the registered window, gated-path inbound quality is at or below the skip-path, OR if cold/search traffic exits sub-five-seconds at a rate that materially exceeds skip-path, OR if engaged-session rate from the gated path is at or below the skip path, then the bet is refuted. Demote the gate, report it as FAIL, name the next thing to try.

3. 03

   What would confirm it (PASS, conjunctive)

   ALL of: gate completion rate stays above the registered floor, engaged-session rate from the gated path materially exceeds the skip path, source × behaviour shows the predicted population split, and inbound quality from the gated path is higher than the skip path on a pre-registered window.

4. 04

   How we will report it

   Verdicts are discrete: PASS, PARTIAL, FAIL, or WITHHELD. Never a percentage and never spun. Every number on this page will point to a commit, a stored artifact, and a query that reproduces it.

We adopt the Lab Protocol’s discipline: one cure at a time, gates written first, verdicts that can lose. Operational behavioural and organisational measures are necessary-not-sufficient substrates with ZERO evidential weight for awareness, consciousness, or life on their own. The reverse-filter is a hypothesis about conversion-pool composition, not about minds.

The signal set

What we measure, by source, ourselves.

First-party only. No Google Analytics. No third-party trackers. No bounce-rate kill-switch (bounce is deprecated as a quality metric; we compute engagement ourselves).

Gate start → completion

By source. Confirms: healthy completion among the right visitors. Weakens: heavy bleed before solve.

Engaged-session rate

Self-computed (>10 seconds OR a key event OR 2+ views). By source. Bounce-rate kill-switches are deprecated; we never use bounce.

Source × behaviour

Does direct/referral traffic finish as directed explorers, or quit as friction casualties? The reverse-filter bet predicts the first.

Inbound quality, not count

The decisive one. Richness of the intake. Lower-volume / higher-quality from the gated path would confirm. Equal or worse refutes.

Reveal → contact efficiency

Time from gate reveal to a meaningful action, solved-path vs skipped-path. Tests whether the gate sharpens intent or just delays it.

Return-visit quality

Salted-daily cohort, no durable identity. Reverse filtering can also work through delayed qualification.

Sub-5s trust exit

An early-warning that cleverness is costing credibility. If this rises sharply for cold traffic, polish beats cleverness, and we say so, fast.

Where the data goes

Our own iron, never on the internet.

The browser fires a small, no-PII event to a sovereign endpoint on our server. The endpoint validates the event against an allowlist, computes a salted, day-scoped identifier server-side (it cannot be linked across days, by construction), strips anything outside the allowlist, and writes the row via a database role that can only INSERT: not read, not delete, not modify. The database itself stays on a private mesh and is never reachable from the public internet.

When either Do-Not-Track or the Global Privacy Control header is present, the browser sends nothing and the server records nothing. There is no opt-out flag to store because there is nothing to opt out of once suppressed upstream.

The Gates — how to prove us wrong

Registered v0 parameters, dated and load-bearing.

The specifics that turn the hypothesis into a decision the world can score. Each row is a public commitment: if we change it, we log the change here on the same commit that changes it, dated, with the reason. Changing a parameter mid-run and then claiming a win is exactly the move this page is designed to prevent.

Registered surfaces

Two, in the decision context below. (1) gate — whether to present the puzzle gate at all. (2) puzzle — which of the 5 puzzles to show when the gate is presented. The puzzle surface is nested inside gate=canon: it only decides for visitors who are already going to see a gate.

Registered arms — gate surface

canon (show the puzzle gate; the baseline) and nogate (reveal the page directly, no puzzle). Both registered 2026-07-02. The engine assigns the arm; the assignment is randomised, not visitor-chosen, so the comparison is causal rather than self-selected.

Registered arms — puzzle surface

Five, matching the five puzzles by name: weave, constellation, dawn, predict, gather. Registered 2026-07-02. Warm-started from ~1 week of rotation observations (see below); the engine can still explore any arm at any time via the epistemic bonus.

Registered decision context

source_bucket = organic-search, home page only. This is the ONLY context that adaptively decides; every other visitor sees canon (the gate). One cure at a time.

Scoring

score(arm) = sample(Beta(α, β)) + κ · √Var[θ]. Thompson sampling plus a directed exploration bonus that provably decays to zero as counts grow. Cold start: both arms begin at the uniform Beta(1, 1) prior, so early traffic is split ~50/50 while the engine learns.

Outcome mapping (gate surface)

Path-independent so it scores canon and nogate fairly. SUCCESS: the session became engaged (>10s, a key event, or 2+ views) OR submitted an intake, counted once per session. FAILURE: an assigned session that did not become engaged. The sub-5s trust-exit (trust_pulse) is reported alongside as a separate falsifier, never folded into the score.

Outcome mapping (puzzle surface)

SUCCESS: gate_solve or intake or engaged. FAILURE: gate_skip. This is a within-arm mapping (all puzzle arms are shown a gate, so gate_solve vs gate_skip is a fair signal here). The categorical trust_pulse remains a separate falsifier.

Nightly forgetting

α ← 1 + γ · (α − 1), β ← 1 + γ · (β − 1), γ = 0.98. Multiplicative shrinkage keeps directed exploration alive; the fixed point is the uniform Beta(1, 1) prior.

Preference weights (§3.4)

w1 (qualified intake) = 5.0 · w2 (reveal→CTA) = 3.0 · w3 (engaged session) = 2.0 · w4 (return-visit quality) = 1.5 · w5 (sub-5s trust-exit penalty) = 4.0. Placeholder v0; owner-versioned.

Epistemic scale κ

0.5. Conservative start per §3.3. Any change is a logged public event.

Rolling window

30 days. Old events fall out naturally; the nightly shrinkage sits on top.

Minimum sample size

Gate surface: 350 assigned sessions per arm (700 total) in the organic-search context before any verdict. Puzzle surface: 200 assigned sessions per arm (1000 total across the 5). Both registered 2026-07-02 in the same commit that flipped the decision on. Until min-N is reached the verdict is WITHHELD. At current organic-search volume this is expected to take months — the honest cost of a clean test.

Verdict form

Discrete: PASS · PARTIAL · FAIL · WITHHELD. Never a percentage. Never spun.

The instrument is now deciding, for exactly one context: organic-search on the home page. Every other visitor sees the gate. The nightly recompute writes durable posteriors on our own iron; the edge reads them to assign the arm; CANON (the gate) is the failure floor on any miss or error. The minimum sample size below was registered in the same commit that turned the decision on, and cannot be moved now that the first row has landed. The verdict stays WITHHELD until min-N — expected to be months at this traffic, which is the honest cost of a clean test.

The honesty fences

What this is, and is not, claiming.

1. 01

   We publish nulls and refutations, not just wins. A PARTIAL is reported as PARTIAL and never spun into the larger story.

2. 02

   The site never stores name, email, IP address, raw User-Agent, form content, or any durable cross-day identifier. A short-lived session ID lives in your tab only.

3. 03

   Do-Not-Track and the Global Privacy Control are honored absolutely: when either is set, no event is recorded and nothing is sent.

4. 04

   F1 (awareness). Operational behavioural and organisational measures are necessary-not-sufficient with ZERO evidential weight for awareness, consciousness, or life on their own. Passing a gate demonstrates the named behaviour, never experience.

5. 05

   F2 (preprint). The UNI preprint (DOI 10.5281/zenodo.19785799) is an unrefereed working preprint. Peer review is pending.

6. 06

   F3 (scope). The on-site engine is a faithful lightweight subset of the categorical active-inference core, not the research engine. What is demonstrated here is stated as demonstrated; what is specified but not yet run is stated as specified; what is aspirational is stated as aspirational.

How to push back

You can break this in public.

If our reasoning is wrong, the place to say so is on the record. Bring a counter-prediction, a measurement we are missing, or a claim you can show is not falsifiable. We will respond in plain words, on the record.