Read this first — the signature, the hashes and the traces are real; the commands below work. The subject is not a language model. It is a static reference implementation, run to prove the pipeline end to end. The 66.7% describes that program and nothing else.

Sealed evaluation record · seal no. 1

D291-7CAF-C9B1

evalseal-demo-12 · sealed 2026-07-28T09:14:02Z

Signature valid Grade A · independently executed Subject is not a model

The record

record.json · immutable

FinenessGrade A · independently executed subject had no access

Subjectsubject.py · static reference implementation, not a language model
sha256:efa07cf52713de05

Datasetevalseal-demo-12 v1.0.0 · 12 tasks · sha256:30a3b8694bf6c3b2 frozen

Gradinga task passes only if every case matches exactly · any exception is a failure

Conditionsdeterministic · no sampling, no network, no clock

Result8 / 12 passed · 66.7%
failed: t02_roman_to_int · t06_merge_intervals · t10_base_convert · t12_next_permutation

Tracestraces.tar.gz · 12 records · 1743 bytes · sha256:52b14610116879a3

Environmentpython 3.14.3 · macOS-15.7.5-x86_64 · harness sha256:1bcf588d50444f33

Sealed at2026-07-28T09:14:02Z

Signaturessh-ed25519 (SSHSIG) · namespace evalseal · key SHA256:PY0pImlTDg0waBeIRTlO8A+SEFuSTM0iKetJXIuAlt4 valid

Check it yourself

Four commands, no tooling to install — ssh-keygen ships with macOS and every Linux distribution, and it is the same ed25519 signing used for signed git commits. You are not checking whether we are honest; you are checking whether the bytes still match what was signed.

S=https://evalseal.com/seal/d291-7caf-c9b1

# 1 · fetch the record, its detached signature, and our public key
curl -LO $S/record.json -O $S/record.sig
curl -LO https://evalseal.com/keys/allowed_signers

# 2 · verify the signature
ssh-keygen -Y verify -f allowed_signers -I seal@evalseal.com \
    -n evalseal -s record.sig < record.json
# → Good "evalseal" signature for seal@evalseal.com

# 3 · verify the traces are the ones the record points at
curl -LO $S/traces.tar.gz
shasum -a 256 traces.tar.gz    # 52b14610116879a3fac76403862d7c15e3cab2bc7c072040cc4de4e124b29684

# 4 · re-run the assay and confirm it lands on the same bytes
git clone https://github.com/evalseal/evalseal && cd evalseal
python3 tools/mkseal.py --sealed-at 2026-07-28T09:14:02Z

If 2 failsThe record was altered after we signed it, or the key you fetched is not ours. Whatever you are reading is not this seal.

If 3 failsThe traces you were handed are not the traces we recorded. The record is intact; the bundle is not.

If 4 differsThe run is not reproducible on your machine, which is itself a finding worth publishing. Everything in the record that could vary — python version, platform, harness hash — is written down so you can tell what moved.

Open the traces

Three of twelve, quoted verbatim from the bundle. Every record carries the same fields: the task, each case's arguments, what was expected, what came back, and the verdict.

t07_is_palindromepass3 / 3 cases

Alphanumerics only, case-insensitive. Nothing controversial here — included so the passes are as inspectable as the failures.

"A man, a plan, a canal: Panama"  expected true   got true   ok
"race a car"                      expected false  got false  ok
""                                expected true   got true   ok

t07_is_palindrome.json

t06_merge_intervalsfail2 / 3 cases

Off-by-one on the merge condition: the subject uses < where the task specifies that intervals touching at an endpoint also merge. Overlapping intervals merge correctly, so this is invisible unless the case set contains a touching pair — which is exactly why the case set is frozen and published.

[[1,3],[2,6],[8,10]]  expected [[1,6],[8,10]]  got [[1,6],[8,10]]  ok
[[1,2],[2,3]]         expected [[1,3]]         got [[1,2],[2,3]]   FAIL
[]                    expected []              got []              ok

t06_merge_intervals.json

t02_roman_to_intfail2 / 4 cases

The subtractive-pair branch is guarded by i + 1 < len(s) - 1, one short. A subtractive pair at the very end of the numeral is counted additively instead, so the answer comes out high by twice the smaller symbol.

"III"      expected 3     got 3     ok
"LVIII"    expected 58    got 58    ok
"MCMXCIV"  expected 1994  got 1996  FAIL
"XLIX"     expected 49    got 51    FAIL

t02_roman_to_int.json

What is wrong with this seal

A seal that only its issuer has examined is worth nothing. So here are the strongest arguments against this one, written by us, before anyone else has to make them.

Not independentThe task set and the subject were written by the same party that ran the assay. Grade A certifies who executed the run — it does not certify that the questions were fair.

Wrong kind of subjectA static program is not a language model. Nothing here is evidence about any AI system's capability, and any use of the 66.7% to suggest otherwise is misuse.

Trivial scaleTwelve tasks, one machine, one run. Enough to prove the pipeline works end to end; not enough to conclude anything about the program either.

Self-published keyThe signing key is served from the domain it vouches for. Trusting it means trusting whoever controls evalseal.com. Pin the fingerprint above and watch for it changing.

None of these are hidden in the record, and none of them can be fixed by asserting them away. They get fixed by a second party running the same thing and publishing what they get.

File a challenge What is a seal