The state of domain availability data

A frozen, hashed sample. A published harness. Dated runs kept rather than overwritten. And the findings against ourselves, first.

sample20 cases · 2026Q3.1

expected registered14

expected unregistered3

public runs0

latest public accuracy — no public run yet — dry runs stay owner-visible

Run it yourself

The sample and the harness are both published. Nothing here asks you to take our word for it.

git clone <this repo> && cd bench
python3 bench.py --sample sample.json    # sample 2026Q3.1
python3 report.py                        # renders the report

If a case looks wrong, say so and we will re-run it and publish the re-run. That is rule 7.

The sample

Frozen and hashed. Every case states why it is in there.

domain	expected	why it is in the sample	tags
stowed.co	registered	bootstrap-absent TLD; resolves, so registered. The original false-unregistered case.	bootstrap-absent regression
t.co	registered	bootstrap-absent TLD, 1-char label; GoDaddy .co endpoint 404s it despite registration.	bootstrap-absent short-label partial-coverage
google.co	registered	bootstrap-absent TLD, high-profile; same false-404 class.	bootstrap-absent partial-coverage
angel.co	registered	bootstrap-absent TLD; 404s on the rejected .co endpoint.	bootstrap-absent partial-coverage
vibecode.me	registered	bootstrap-absent TLD, young registration (2025) sourced via CT logs.	bootstrap-absent young-registration
vibe.me	registered	bootstrap-absent TLD, launch-era registration (2008).	bootstrap-absent launch-era
google.sh	registered	bootstrap-absent TLD, rarely-supported ccTLD.	bootstrap-absent rare-cctld
scopeweb.io	registered	bootstrap-absent TLD; registered 2026-08-20, same-day registration edge case.	bootstrap-absent same-day
example.com	registered	IANA reserved-for-documentation, registered 1995. Ancient control.	control ancient
ibm.com	registered	1986 registration; oldest-era control.	control ancient
zx.com	registered	2-char .com; premium-tier profile.	control short-label premium
zzqx7nvwplk39fake.com	unregistered	Gibberish control; must be unregistered on an authoritative endpoint.	control negative
zzqx7nvwplk39fake.io	unregistered	Gibberish on a bootstrap-absent TLD.	control negative bootstrap-absent
zzqx7nvwplk39fake.me	unregistered	Gibberish on a bootstrap-absent TLD.	control negative bootstrap-absent
xn--bcher-kva.com	unknown	Punycode (bucher.com with umlaut). IDN handling.	idn punycode
xn--fiqs8s.com	unknown	Punycode for a CJK label.	idn punycode
a.com	registered	1-char .com; registered 1993 but has no NS. Registered-yet-unresolvable.	short-label no-ns
thisisaverylongdomainlabelusedtotestmaximumlengthhandlingxyz.com	unknown	60-char label; near the 63-char DNS limit.	long-label
start.co	registered	Resolves to a parking operator (sedoparking). Parked classification case.	parked bootstrap-absent
lootlens.ai	registered	Registered, does not serve. Dormant classification case.	dormant

The ledger

Every run, dated, kept. A dry run is owner-visible and never published as a score.

date	mode	sample	correct	excluded
2026-09-19	dry	2026Q3.1	18	2
2026-08-20	dry	2026Q3.1	18	2
2026-08-20	dry	2026Q3.1	18	2

The method

1. Never benchmark a strawman — measure the thing people actually use.
2. Publish the harness. A number nobody can reproduce is an advertisement.
3. Freeze and hash the sample. A moving sample measures nothing.
4. Keep dated runs. Overwriting a run destroys the only trend you have.
5. An abstention is not a wrong answer, and must not be scored as one.
6. Say what you could not measure, and why.
7. Re-run anything a reader disputes, and publish the re-run.
8. Stay polite about competitors. The data is the argument.
9. Benchmark yourself hardest — findings against us go in the report.
10. A test that passes for the wrong reason is worse than a failing test.
11. The same trap returns wearing a different helper.
12. A check must be able to fail.

Rules 10 to 12 are about our own harness rather than anyone else's data. They are here because they cost us the most.