# Pre-registration — `runs/2026-11-16-viallo-shared-album-storage`

```
run_id:      2026-11-16-viallo-shared-album-storage
product:     viallo
tier:        micro
written_at:  2026-11-16T17:30:00Z
revision:    1 — adversary pass precedes the freeze
```

## 0. Exploratory disclosure, before anything else

A first pass over the same stored pages ran earlier on 16 November 2026, before this file existed.
It produced the numbers that the draft `ARTICLE.md` v1 carries (52 on-topic pages, 16 stating the
rule, 7 stating the old one). Those numbers are therefore **exploratory**, and this run is not a
prediction of them; it is a verification of them under rules fixed in advance.

Two things follow, and both are binding on the article. The expectation below is marked
`post-hoc`. The confirmatory step is not the first pass repeated by the same reader: it is two
independent codings by coders working from section 3 of this file, at least one of which never saw
the first pass, and their agreement is reported as a measured number. Every share in the article
that depends on the coding is stated with its raw counts and marked exploratory.

## 1. Question

Of the web pages that answer whether shared albums count toward cloud storage, what share states
the rule Apple applies to albums created in iOS 27 or later, and what share still states the
pre-2027 rule as current?

The subject that can be counted is a page: one URL, one row.

## 2. Method

### 2.1 Inputs, fixed before measuring

Five responses from `serp/google/organic/live/advanced`, location 2840, language `en`, depth 100,
taken on 16 November 2026 for five phrasings:

1. `do shared albums take up storage`
2. `do icloud shared albums take up storage`
3. `do shared albums use icloud storage`
4. `does google photos shared album take up storage`
5. `are shared albums stored in icloud`

The stored files are `data/101-serp-depth100.json` to `data/105-serp-depth100.json` in the run
workspace `/Users/jakubzavada/Repositories/.in-progress/viallo-age-article`. Every organic item is
deduplicated on site plus normalised title, which yields the page list. Each page was fetched once
with a plain GET, and the HTML is stored under `data/110-crawl/` with a SHA-256 in
`data/corpus.json`. A page is **readable** if the fetch returned HTTP 200 and the extracted text is
over 500 characters.

### 2.2 Inclusion

A readable page is **on topic** if its text contains "shared album" or "shared albums" and at least
one of: storage, stored, space, quota, count. This is a lexical rule applied by script, not a
judgement.

### 2.3 Coding, three classes

Two coders code every on-topic page. Coder A is the orchestrating session. Coder B is a session
that never saw the first pass, gives a class and a quoted sentence per page, and works from the
extracted text only.

- **NEW** — the page states, as the rule now in force, that shared albums count toward a storage
  quota, and either scopes that to albums created in version 27 or later (iOS 27, iPadOS 27,
  macOS 27, visionOS 27, "new albums") or states the counting rule without any statement that
  shared albums never count. A page that mentions iOS 27 in a navigation bar or an unrelated
  paragraph without stating the counting rule is **not** NEW.
- **OLD** — the page states, as the rule now in force, that shared albums do not count toward
  iCloud storage, and nowhere on the page is the version 27 change mentioned. A page that gives
  the old rule and also shows the change is **not** OLD; it is NEW.
- **OTHER** — everything else: a question without an answer, a page about Google Photos that
  states only the Drive rule, a page quoting both rules without saying which applies, a page whose
  storage sentence is about the phone rather than the cloud.

Every code carries a quote. A page whose coder cannot quote the sentence that carries the rule is
coded OTHER regardless of the impression it leaves. Disagreements are re-read by the orchestrator
against both quotes, and if the page does not settle it, the class is OTHER.

### 2.4 Analysis fixed before the number exists

Reported: the count of pages in each class, the on-topic denominator, the readable and fetched
denominators, and Cohen's kappa between the two coders over three classes. The share of NEW pages
among results that ranked in the top ten of any of the five queries is reported beside the share
among the rest, with counts, because the question a reader actually has is whether the answer is
visible where people look. Below twenty samples every percentage travels with its raw count. No
confidence interval is claimed for the population of pages in the world: the SERPs are a census of
one day, not a sample of a distribution.

### 2.5 The second measurement

Six vendor claims are checked by fetching the vendor page and testing for the exact sentence, in
`dataset/check.py`. Each is binary and carries URL, quote, and retrieval date. A claim whose
sentence is absent is reported absent; the script exits non-zero.

### 2.6 Replication

Re-execution: a child session re-runs the classification script from the stored corpus and must
reproduce every count. Its report lands in `replication/`. Independent derivation is out of scope
at the micro tier.

## 3. Expectation (post-hoc, from the exploratory pass)

At least five pages state the old rule as current, and fewer than half of the on-topic pages state
the rule that applies to albums created today.

## 4. Falsifier

The expectation is wrong if fewer than three pages are coded OLD, or if NEW is at least half of
the on-topic set. Either result is publishable and the article changes shape accordingly.

## 5. Abandonment

The run ends `NOT_MEASURABLE` if fewer than thirty readable on-topic pages remain after dedupe, or
`METHOD_BROKEN` if the two coders agree on fewer than three quarters of pages (kappa below 0.6),
because a coding that two careful readers cannot reproduce is not a measurement.

## 6. Estimate

**What this needs from Jakub:** nothing. The inputs are already downloaded and stored.

Money: 0.086 USD, already spent on the five SERPs. No further paid calls are planned. Machine time:
about two hours, mostly the two coding passes and the figures.

## 7. Adversary

`ADVERSARY.md` is written by a session that received the question, the method and the inputs, and
no result. Its objections and their dispositions are recorded as an amendment below before the
file is frozen.

## 8. Amendments

(to be appended before `freeze`)

## 8. Amendments after the adversarial pass, before the freeze

Twelve objections, dispositioned. The adversary ran on the session's default model, not Opus, because
this profile routes subagents to DeepSeek; the deviation is recorded in `RECORD.md`.

1. **NEW had an asymmetric fallback.** Changed: NEW splits into **NEW-SCOPED** (names the version, the
   date or "new albums") and **NEW-UNSCOPED** (states the counting rule, no scope, no statement that
   albums never count). Both are reported; the headline uses the sum and the table shows both columns.
2. **The denominator is Google's ranking, not the web.** Accepted as a stated limitation and the
   question is narrowed: every share is "among the readable pages in these five top-100 results on
   16 November 2026". The article says so in the same sentence as the share.
3. **The query set is mixed-platform.** Changed: a platform field (apple, google, dropbox, other) is
   assigned from the domain before the coding pass, and the headline share is reported for
   Apple-scoped pages separately from the pooled set.
4. **A quotable sentence excludes tables and lists.** Changed: a quote may be a table row or a list
   item as extracted; the coder quotes the span, not a grammatical sentence. Pages whose rule sits in
   non-prose structure are counted and reported.
5. **A plain GET, no browser, no retry.** Changed: every non-200 and every page under 500 characters
   is re-fetched once with a browser user agent and a retry, and the outcome is recorded per URL. Pages
   still unreadable are listed with their status. Accepted limitation: a logged-in or paywalled page
   stays unreadable.
6. **Deduplication on title.** Changed: pages are also deduplicated on final URL (canonical), so one
   page returned under two rewritten titles counts once. A template syndicated across domains still
   counts per domain; stated limitation.
7. **Agreement measures reliability, not truth.** Changed: coder B receives a calibration set of three
   pages with a known answer before the sample, coded but excluded from the analysis, and both coders
   are blind to rank query and prior coding. Per-class agreement is reported beside kappa.
8. **Version 27 is hardcoded as the marker.** Changed: the marker becomes a field, `scope_marker`
   (version, date, "new albums", none), coded separately; NEW requires the counting rule, and the scope
   is derived after coding rather than gating the class.
9. **Top-ten versus rest is confounded by type and query.** Changed: page type and per-query rank are
   recorded; the top-ten set is "ranked in the top ten for at least one of the five" and is reported
   per query as well as pooled.
10. **The rule itself is asserted, and the vendor count was wrong.** Changed: the normative rule is
    quoted verbatim with URL, locale and retrieval time in section 2.5, and the vendor set is seven
    rows, not six. A row whose sentence is absent is reported absent, never as a false vendor claim.
11. **No page timestamps.** Changed: any on-page updated or published date is extracted per page and
    the OLD share is reported by age band where a date exists. Accepted limitation where none exists.
12. **The AI answer surface is outside the sample.** Changed: the AI overview returned with the first
    SERP is recorded as a separate, clearly labelled population with its cited sources, and is not
    pooled with organic results.
