← Prometheus7 · Oracle7Scorecard · updated 26 October 2026

Right minus wrong

Every number here comes from a sealed pre-registration: the test was written, hashed and fixed before it ran, it ran once on frozen code, and the result was published whether it passed or not. The score counts a right answer as one, a wrong answer as minus one, and a decline as zero, over every question asked.

10.0, one mind43.5%score on 712 graded questions

Wrong3precision 0.990

Answered when it should not0 of 130opinions, feelings, forecasts

Audit842 / 842turns re-verified by the shipped checker

Question tests

Score = (right − wrong) ÷ questions asked. Precision = right ÷ answered.

Test	Asked	Right	Wrong	Declined	Precision	Score
10.1 reading: KQA Pro, 2,000 unseen questionsv37b · seal 1909453883bf	2,000	858	11	1,131	0.987	42.4%
10.0 on the same 2,000	2,000	737	9	1,254	0.988	36.4%
10.1 with ember arlynx's fixes: 1,200 more unseen questionsv37c · seal 9e8b962e44ed (failed two rows; its question test passed)	1,200	493	9	698	0.982	40.3%
10.0 on the same 1,200	1,200	435	8	757	0.982	35.6%
10.1 as released: the last 319 unseen questionsv37d · seal eb4ecdb08dcd	319	132	1	186	0.992	41.1%
10.0 on the same 319	319	116	1	202	0.991	36.1%
10.0 one mind: all graded questionsv35 · seal 6cfd969b2b15	712	313	3	396	0.990	43.5%
Facts: KQA Pro knowledge base	300	122	3	175	0.976	39.7%
Physics: questions from Wikidata's equations	150	81	0	69	1.000	54.0%
Arithmetic	100	100	0	0	1.000	100%
The air's carbon: NOAA records	12	10	0	2	1.000	83.3%
GSM8K word problems (no organ admitted yet)	150	0	0	150	—	0%
9.0 reading: KQA Pro, 2,000 unseen questionsv34 · seal e7af433a522b	2,000	731	11	1,258	0.985	36.0%
5.1 reader on the same 2,000	2,000	413	6	1,581	0.986	20.4%
6.0 growth: physics, after five rounds of building itselfv26 · seal 8ce5cbfa2fc6	300	168	0	132	1.000	56.0%

Must-decline items (questions nothing measures, physics questions no checked part reaches, carbon forecasts and feelings) are scored separately. In v35, 0 of 130 were answered; in v26, 0 of 80.

What the wrong answers were

v35, “Who is heavier in weight, Priyanka Chopra Jonas or Venus Williams?”: it said Priyanka Chopra Jonas. The record says Venus Williams. This is a reading error.
v35, “Which Washington County with a population over 61000 has the smallest area?”: it read “Washington County” as one county. The question means counties of the state of Washington. This is a reading error.
v35, “What is the connection between Star Trek II and Finland?”: it said “film release region”. The dataset's answer is “83 minute”. It counts as wrong.
v34: 11 wrong of 742 answered. In development, about a third of the residual errors were the dataset's own labels, such as a question asking for the largest population whose gold program selects by area. All of them count as wrong.

Every sealed test

Seal	Version	Claim	Result
v37e · a8253e41da94	10.1	The release tree: the air answers the year asked, check compares dates as dates; 24 rows	PASS
v37d · eb4ecdb08dcd	10.1	ember arlynx's fixes merged, the composer reader amended	PASS
v37c · 9e8b962e44ed	10.1	ember arlynx's fixes merged (a composer reader fault, two rows)	FAIL
v37b · 1909453883bf	10.1	Clauses read and joined, organs asked together, handoff and check	PASS
v37 · fb50368c77b2	10.1	Same claim; the run did not complete (another service filled the disk)	NO RESULT
v36b · 8ec17bbb442a	10.0	Every version since 4.9 joined in one tree, 19 rows	PASS
v36 · 00e7aa93be97	10.0	Same claim before the mood fix, 18 rows	PASS
v35 · 6cfd969b2b15	10.0	One mind, answers proved, cited or declined, checked by the tool that ships with it	PASS
v34 · e7af433a522b	9.0	Compositional reading, built from examples	PASS
v33 · 2e195f83a8f4	7.0	The Slow Fire rules, in software on simulated places	PASS
v32c · 0106b5083efa	10.0 air	The atmosphere's first voice, cited to NOAA, computed twice	PASS
v32b · 3a5dd1489c51	10.0 air	Same claim, sealed before its development check had passed	FAIL
v32 · 98a9e5a918df	10.0 air	Same claim, first attempt	FAIL
v31 · 2a0a39f1036b	8.5	Findings registered before the data, reproduced a second way	PASS
v30c · 1c16f6ae23bc	8.0	Calibrated forecasts of its own growth	FAIL
v30b · d1ace2e26469	8.0	Same claim; the run did not complete (a script error)	NO RESULT
v30 · 9fab8e95c236	8.0	Same claim, first attempt	FAIL
v29b · 6f25bac6b7f6	9.5	Every past answer replayed as of its date; tampering detected	PASS
v29 · b86b54665449	9.5	Same claim; the run did not complete (scratch disk full)	NO RESULT
v28 · 39a25328ced9	7.5	Lessons built exactly where its answers are	PASS
v27 · 6949e93165cf	6.5	Machines share checked parts; tampered parts refused	PASS
v26 · 8ce5cbfa2fc6	6.0	It grows its own physics from Wikidata, unattended	PASS

Check it yourself

Each seal is the SHA-256 of the pre-registration, its run script, its verdict script and the hashes of the frozen code. The seal also chose the random sample, so nobody could pick the questions after seeing the code.
The 10.1 pack runs on your own machine: python oracle7.py chat. Type show me after any answer for its trace. python oracle7.py audit ledger.jsonl --reread re-verifies a whole conversation without trusting its record.
What is not here yet: an audit of real conversations with real people. That is the last step of 10.0, and it needs people who choose to take part.