How early do we publish?

Pre-registration

The rules were written down and committed before the calculation code was merged. Once on our main branch the file is frozen: each calculation reads it from the main branch and refuses to run if any commit ever changed, deleted or re-added it, and our checks refuse any edit. A change of rules can only be made by adding a new version, listed here with its date. A new version applies only to announcements published after it was registered; earlier rows keep the version that was in force.

• Version in force: v1, read the full rules.
• SHA-256: 65d454cbaaeafa1d041a2eff63bdb1d34c5635712c40412cbf42295efcd76dcc
• Registered: on our main branch on 2026-10-31.

Method in short

• Events. Every verified retail event that has ever been in our data, as first published: later edits do not change the calculation, removed lines stay counted, and event files can only be appended to. Rows are listed for brands in our public data; events about other brands are counted with the same rules and published as anonymous figures.
• Dates. "Entered our data" is the first day the event is on the main branch of our repository (Toronto time), taken from its version history, never from file dates. "On our site from" is the next day, and not before 2026-10-26.
• On-sale evidence. A verified report that the brand is on sale at that retailer in that country, or a listing in our data with a recorded check and a start month. A month counts from its first day. A check record without a start month only shows the latest check, so it can show the brand was already on sale but never makes an announcement fulfilled. Online-only evidence does not count for an in-store launch.
• Lead = evidence date minus the first day on our site, in days, for fulfilled announcements only. The headline is the median.
• Detection lag = the day the event entered our data minus the report date, for every event, always shown split by daily review and one-off batches. Median and 90th percentile (nearest rank).
• Withdrawn only when a correction voids the announcement itself (same source, or issued after we published it). A correction that a listing is online only is not a withdrawal.

Known biases and limits

• Backfilled rows are not an out-of-sample test: the rules were written while those events could already be seen.
• The initial reviewed set has the date 2026-10-19 by rule, not from history, and was first shown on the site on 2026-10-26.
• One-off batches, including the initial set, collect reports that are months or years old, so they dominate the detection lag at the start. The figures are always shown split by batch type for this reason.
• An unfulfilled announcement means only that our data has no on-sale evidence by the deadline. The brand may be on sale without a source we have found.
• The version history behind the dates is in our data repository, which is not public, so the dates cannot be checked independently from the repository. The data file is republished every day at the same address, so anyone can archive copies with a web archive and use them as outside timestamps.
• If a brand leaves our public data, its rows stop being listed by name but stay counted.

Recompute the figures yourself

Download the data file and the rules file, then run this Python 3 snippet:

import json, math, statistics, hashlib
from datetime import date
d = json.load(open("data.json"))            # /track-record/data.json
v = d["prereg"]["active"]
assert hashlib.sha256(open("prereg-v%d.md" % v["version"], "rb").read()).hexdigest() == v["sha256"]
day = date.fromisoformat
a = d["announcement_lead"]
for r in a["rows"]:
    if r["status"] in ("fulfilled", "already_listed"):
        assert r["lead_days"] == (day(r["evidence_date"]) - day(r["visible_on"])).days
every = a["rows"] + a["hidden"]["rows"]     # hidden announcements count too
lead = [r["lead_days"] for r in every if r["status"] == "fulfilled"]
print("A: n", len(every), "fulfilled", len(lead), "median lead", statistics.median(lead) if lead else None)
b = d["detection_lag"]
rows = [dict(r, lag_days=(day(r["published_at"]) - day(r["reported"])).days) for r in b["rows"]] + b["hidden"]["rows"]
for kind in ("daily review", "one-off batch"):
    lags = sorted(r["lag_days"] for r in rows if r["batch"] == kind)
    if lags:
        print("B:", kind, "n", len(lags), "median", statistics.median(lags), "p90", lags[math.ceil(0.9 * len(lags)) - 1])

Track record, audit, corrections and rankings

• The monthly audit measures whether our listings are right. The track record measures how early we publish announcements. They are separate checks.
• An announcement voided through the corrections process is shown as withdrawn. It stays in the table.
• The track record does not enter any ranking and does not change any score or rank. Rankings follow only the published formula.