R3ORDR everything 4dRN9taLwSgAQ5HLqN9Chk2c63k4VsLmSDcru7Jat5rr
Joined July 2011
- Tweets23.4K
- Following2.9K
- Followers1.2K
- Likes45.9K
JW7 retweeted
Ten versions, tested as one
Oracle7 is a question-answering system with no language model in it. Every answer it gives is computed by parts that passed their own exams, read from a named record, or declined with a reason. 10.0 is out on Hugging Face today.
The last public release was 4.9, in October. Since then there were ten internal versions:
organs for steps, maths, code and pictures (5.0);
reading plain questions (5.1);
counting its own composition space without storing it (5.5);
building parts from what it could not answer (6.0);
sending parts between machines that trust nothing they are sent (6.5);
rules for places that can say no (7.0);
lessons (7.5);
hypotheses written before the data is fetched (8.5);
reading questions clause by clause (9.0);
a ledger any past answer can be replayed from (9.5).
One more, 8.0, failed its test twice and is not included.
Each of those versions was tested the same way: a pre-registration written before the run, fixed by a hash, run once, and published whether it passed or not. Each one passed alone. 10.0 is the question of whether they still pass together.
What joining them found
The joined test has one row per version, and each row does the thing that version does, through the same commands a person would type. Its first development run failed four rows. None of the four failures came from a new version breaking an old one. Each was a fault that had been there all along, which no single version’s test could see.
A place blamed its sensor for weather. 7.0’s rules say a fault is the sensor’s until shown otherwise, and a sudden jump is a fault. Its sealed test ran five simulated places and passed. The joined test happened to run a sixth, where the place had been closed for a month. When it reopened, the soil had dried a little over thirty days. The rule compared that reading with the last one, a month old, against a one-day limit, and refused it. Since a refused reading never becomes the new baseline, it refused the next one too, and the next.
A sweep of 3,000 simulated places found the same fault in 347 of them. The rule had passed its test by luck. The fix lets the allowed change grow with the time since the last accepted reading, as a smooth quantity’s natural spread does: with the square root of the days. A one-day spike is still caught, and all 3,000 places now pass. A fixed two-day allowance was tried first, and it failed one place in 2,000 it had not been tuned on. That is why the rule is now the square root instead.
An answer ignored half the question. “Who composed Bach’s Symphony No. 9?” came back “Ludwig van Beethoven”. A specialist had found “Symphony No. 9” and never read “Bach’s”. A contributor’s pull request, merged in this release, had rightly stopped treating a possessive as a claim of authorship (“Bach’s symphony” could be one he owned or conducted). But the answer was wrong under any reading, because nothing in it connects to Bach. The rule now is that what a question names must connect to what its answer names. “Beethoven’s Symphony No. 9” is answered. “Bach’s” is declined, with the reason.
A number was rounded before anyone looked at it. The arithmetic checks every result twice, in exact fractions and in 60-digit decimals. Both routes read each number through ordinary floating point first, so 0.10000000000000001 became 0.1 before either route saw it, and they agreed on the wrong answer. Two checks that share an input are one check. They now read the digits as written. A review of the code by Codex found it.
A count did not count what it claimed. 5.5 counts every chain its parts can form, then runs a sample. With 838 parts, one sampled chain in 200 had no physical inputs at all: it would give a negative acceleration for anything. The row had only checked that chains could be decoded, not that they ran. It now says how many ran, how many could not, and why.
Why this is the useful part
Each fix is in its own file. Every sealed module is still byte-identical to the copy its own test ran, so every earlier result still describes the code it measured. The new behaviour has its own measurement. You can check all of it:
python oracle7.py verify # every part against the release's hashes and its Solana anchor
python oracle7.py place # the Slow Fire rules on a simulated place
python oracle7.py chat
The joined test was sealed after these fixes and run once: all 18 rows passed. Recording the example film afterwards then found one more fault (asked how it was, the conversation layer counted a question it had declined as answered), so the test was amended and sealed again on the final tree: 19 of 19 rows passed, and the release tree was unchanged by the run.
The honest summary of 10.0 is that it is ten versions that each worked, which did not all work together until they were tested together. Coverage is still the open problem: most open questions are declined. But a system that declines, and whose every answer can be re-checked, is one that can be improved without guessing whether it got worse. The joined test is how we know it did not.
JW7 retweeted
Oracle7 10.0 is out on Hugging Face. There’s no language model inside: every answer is computed by parts that passed their own exams, read from a named record, or declined with a reason. Versions 5.0 to 9.5 were internal. 10.0 tests them together in one sealed run.
huggingface.co/psiloceyeben/…
Oracle7 v10.0: A Major Leap View article
JW7 retweeted
Oracle7 v10.0 Slim is live on Hugging Face!
huggingface.co/psiloceyeben/…
Compact Changelog, Full Changelog Below
huggingface.co/psiloceyeben/…
Oracle7 10.0: changelog since 4.9
5.0 Organs: steps, maths, code, and pictures and film, each checking its own work. Sealed v23 passed.
5.1 Reading: plain questions are read into typed readings. Sealed v24 passed.
5.5 Represented compute: 1,755,340 checked chains counted without storing them. Sealed v25 passed.
6.0 Growth: builds new parts from what it couldn’t answer, unattended. Sealed v26 passed.
6.5 Many places: parts move between machines and are re-checked on arrival. Sealed v27 passed.
7.0 Slow Fire: rules for places that can say no. Sealed v33 passed.
7.5 Teacher: every physics answer opens into a lesson. Sealed v28 passed.
8.0 Forecasts: failed twice, not included.
8.5 Discovery: hypotheses written before the data is fetched. Sealed v31 passed.
9.0 Reading: knowledge-base questions read clause by clause. Sealed v34 passed.
9.5 Continuity: a hash-chained ledger that can replay any past answer. Sealed v29b passed.
10.0 A mind you can check:every answer is proved, cited or declined;
the air’s voice, from NOAA’s carbon dioxide and methane records;
the audit10 checker;
emotion in chat;
new commands: organs, lesson, compute, grow, packet, place, discover, ledger, air, audit.
Contributed: ember arlynx’s PR #1, which binds quantities by name and unit, records the answer actually delivered, and isolates runtime state. Merged, and also released as 4.9.1.
Fixed during joining:the Slow Fire sensor rule;
possessives, so “Bach’s Symphony No. 9” is no longer answered “Beethoven”;
exact arithmetic;
the ego’s mood count;
debug tags in replies;
user files kept out of the release.
Release test: v36b, run once, passed 19 of 19 rows.
Anchor: explorer.solana.com/tx/39532…
Oracle7 10.0 Slim is live
huggingface.co/psiloceyeben/…
Full Changelog
Oracle7 10.0: changelog since 4.9
The previous release was 4.9, the swarm. It was published on Hugging Face on 23 October 2026 (oracle7-slim at 083e3a6c, head w436). Its Merkle root, 44f51e06…, is anchored on Solana mainnet (transaction WRZKt7GW…).
Everything between 4.9 and 10.0 was internal. 5.0, 5.1, 5.5, 6.0, 6.5, 7.0, 7.5, 8.5, 9.0 and 9.5 were each built and sealed on their own, with their results recorded, but none was published as a release. 8.0 failed and is listed below as failed. 10.0 publishes all of them together.
10.0 is a capable continuation. It is everything 4.9 could do, plus every internal version since, in one release tree, run by the same oracle7.py. No version was built beside the release. Each was built and sealed alone first, and then joined to the others and tested again as one: pre-registration v36, sealed and run once on the joined tree.
There is no language model anywhere in it. Every answer is proved (computed by checked parts), cited (read from a named record), or declined with its reason.
5.0: Organs
The swarm gained four organs. Each one checks its own work.
Steps. Questions about named things become steps on a knowledge base. Every intermediate result is run and must be non-empty.
Maths. Formulas are solved for any of their quantities. Units are checked term by term, and each solution is confirmed by substitution and by solving numerically from scattered starting points.
Code. A plan that solved becomes one program made of its parts’ own code. The program is tested in a separate process and examined like any part.
Pictures and film. Charts, drawings and short films are drawn from a typed scene. Every frame is read back from its pixels and checked against the scene.
Sealed test v23: passed. Unseen structured questions: 946 right, 23 wrong (0.976).
5.1: Reading
Plain questions are read into the swarm’s work, as typed readings. An answer stands only when every reading that solves agrees and every number given is used.
Sealed test v24: passed. Reading 207 of 1,000 at precision 1.000.
5.5: Represented compute
The composition space is counted exactly without storing it: 1,755,340 checked chains from 223 parts.
2,000 of 2,000 sampled chains were decoded exactly and run, with every maths step re-solved.
Sealed test v25: passed.
6.0: Growth that compounds
What it cannot answer becomes a part it builds, unattended. The parts come from Wikidata’s defining formulas, read and checked by the maths organ.
Sealed test v26: passed. 615 parts admitted. Right answers rose 2 → 40 → 85 → 130 → 168 over five rounds, with no wrong answer and no false yes. Reach rose from 50,326 to 588,283 units.
6.5: Many places
Parts move between machines as packets. The receiving machine checks everything itself: hash, source, re-derivation, and a fresh exam.
Sealed test v27: passed.0 of 80 tampered packets were admitted.
Every genuine packet was admitted.
A third machine agreed 453 of 453.
33 questions were answered right through parts built elsewhere.
7.0: The Slow Fire, in software
A place is greeted before it is read, and it can say no; the refusal is binding and kept.
Sensors are examined like any specialist.
Readings are signed, and a tampered reading is refused.
A fault is the sensor’s until shown otherwise.
Nothing is said about a feeling.
Sealed test v33: passed, on simulated places.
The first plant, an indoor vine, runs live in simulation at prometheus7.com/vine until its sensors arrive.
7.5: Teacher
Any physics answer opens into a lesson:each step with its Wikidata formula;
each step checked by putting its values back into its own equation;
a worked example re-solved from the equations;
practice questions;
a picture.
Sealed test v28: passed.
8.0: Calibrated forecasts of its own growth
Failed twice (v30, v30c), and one run did not complete (v30b). Its error intervals were too narrow. Not included as a capability; the failures are published.
8.5: Discovery
Hypotheses are written from its own formula library before any data is fetched, then tested on fresh Wikidata values and reproduced a second way.
Sealed test v31: passed. 14 findings.
9.0: Reading questions, composed and self-built
Questions about the knowledge base are read clause by clause. A clause that names something becomes its own small program, which must run and name something before it is used. Clauses can nest.
Everything it reads with is learned from worked examples and replicated in both halves of the training data: aliases, operator words, glue words, and the answer kinds each question form takes.
Nobody wrote a rule about a question.
Sealed test v34: passed. 2,000 unseen questions: 731 right, 11 wrong (0.985), the rest declined. That is 1.77× as many right as the 5.1 reader.
Its learned tables are cached, so it loads in seconds instead of 35 s, with answers byte-identical (checked on 600 questions).
9.5: Continuity
Every admitted part is stored by the hash of its files. Every answer is kept in a hash-chained ledger with the state it was given against. Any past answer can be replayed as of its date.
Sealed test v29b: passed. 320 of 320 replayed exactly; 20 of 20 tamperings detected.
10.0: A mind you can check
One law. Every answer is proved, cited or declined, and every part that can answer is asked. An answer is given only when they agree.
The air’s voice. Carbon dioxide and methane, from NOAA’s records, with the rows cited and every number computed two ways. It declines feelings, forecasts and blame.
The shipped checker, audit10. It re-verifies a whole conversation without trusting its record. It recomputes every answer by its own route and catches an edited answer even when the whole chain is rehashed.
Chat (fold w572 on the 4.9 head):the ego now does emotion that fits the turn: grief answered as grief, not illness; a crisis met by a person first; a greeting; warm thanks;
it reports a mood it measures from its own conversation record, and answers honestly about its own feelings;
the 9.0 reader, the air’s voice and the 6.0 physics parts answer in chat where the core cannot read a question;
“teach me” opens a 7.5 lesson.
Commands: organs, lesson, compute, grow, packet, place, discover, ledger, air, audit.
Sealed test v35: passed.313 right, 3 wrong.
0 answers given to 130 questions that must be declined.
The audit, re-asking every question, passed all 842 turns.
Every tampered answer was caught, with no false alarms.
Contributed: ember arlynx, PR #1 on oracle7-slim
“Correct quantity binding, final-outcome evidence and isolated runtime state”. Merged into 10.0, and released on its own as 4.9.1.
Quantities.A quantity is its name and exact unit, so a height no longer satisfies a width because both are metres.
A task can connect names explicitly with aliases. The swarm ships one declared alias, a dam’s height used as a turbine’s head, and shows it as a step whenever it is used.
Bound parts. Formulas run from a private snapshot whose admission hash was checked. A changed, revoked or expired part cannot run under an old admission.
Kernel checks (veri7vn).Kernels run only if they match their SHA-256.
Requests and replies are bounded, and each reply must carry its request’s ID.
“proved” is now a scoped kernel_check over the whole displayed answer.
The delivered answer.One boundary per turn records the answer actually delivered.
A replaced or withheld answer drops stale proof fields and cannot become a later pronoun’s referent.
A checked contradiction stays refused through later fallback folds.
Premises. Owning something is no longer treated as making it.
Meteorology v1 and v2. Whole numeric spans, units and roles are validated.
Runtime state. An explicit ORACLE7_STATE now holds vessels, events and requests, which were being written into the source tree.
Tests. 102 of ember’s tests pass on the 10.0 tree, including the swarm’s against the merged planner.
Admitted through the gate, with two maintainer amendments. The gate compared 4.9 with 4.9 plus the PR over the 5,112-item capability panel and the regression list.
Three answers that 4.9 gave are now withheld, and that is right: a −200 mile trip, and heaters run for 30 and 26 hours a day.
Four correct answers were lost. Beer ABV from gravities was withheld because the PR’s input-dependence check added 1 to each number (1.05 became 2.05), outside what the composed formula takes. The check now moves a number by one unit in its last written decimal place (1.05 becomes 1.06).
One false yes appeared. “How many kWh does it take to heat 5 kg of water from 80 C to 30 C?” 4.9 had withheld it only by accident: the old kernel check compared the first numbers, kWh against J. The PR’s honest comparison exposed it. A new fold, w437, withholds an amount something takes when it comes out negative, because the premise does not hold (that is cooling).
4.9 plus the PR plus these is 4.9.1 (head w437).
Merging it with 5.1. The swarm’s planner (point.py) was changed by both 5.1 and the PR. It was merged by hand, with three decisions:
ember’s identity rule throughout;
5.1’s steps, code, reading and picture organs kept;
5.1’s “convert to any input of its kind” fallback removed. That fallback was exactly the loose binding the PR fixes.
Fixed during the joining (found by the gate and the cumulative test)
Each fix is in its own file; every sealed module stays byte-identical to the copy its test ran.
The Slow Fire’s sensor rule (7.0 → 7.1). v33 passed on its five simulated places. A sweep of 3,000 places found 347 where an intact, consented reading was refused (87 of seeds 0–999, 260 of 1000–2999). Two causes:The rule. A jump was judged against the sensor’s per-reading limit from its last accepted reading, and a refused reading never becomes the baseline. After the place’s closed month, a sensor whose value had drifted over the unread days was refused from then on. The allowed change now grows as a smooth quantity’s natural spread does: the limit times the square root of the days since the last accepted reading. A one-day spike is caught as before. (A fixed two-day allowance was tried first and failed 1 place in 2,000 it had not been chosen on, so it was replaced.)
The simulation. Simulated quantities were clipped to exact bounds, so a quantity resting at a bound repeated the same value six times, which the stuck rule refuses and no real sensor reports. Each intact reading now carries its sensor’s measurement noise, on a separate random stream, so the place evolves exactly as before.
After both, all 3,000 places pass every bar: no intact reading refused, no fault or tampered reading accepted. slowfire71.py; the place command uses it.
Every named thing connects (fold w573). On the published 4.9 head, “Who composed Bach’s Symphony No. 9?” was answered “Ludwig van Beethoven”: the composers specialist found “Symphony No. 9” and never read “Bach’s”. ember’s PR rightly stopped treating a possessive as a claim of authorship, but the answer was wrong under any reading of it, because nothing in it connects to Bach. A reply is now delivered only if at least one reading of the possessor connects to something the reply names in the relation store. “Beethoven’s Symphony No. 9”, “Leonardo da Vinci’s Mona Lisa”, “Germany’s capital” and “Napoleon’s wife” are answered. “Bach’s Symphony No. 9”, “Picasso’s Mona Lisa” and “Shakespeare’s Don Quixote” are withheld, and the reply says why. An ambiguous name never causes a withhold on its own.
The ego’s mood counts what was delivered. Asked how it is, the ego reports a state it measures from the conversation. It noted each turn before the fold above it could withhold the reply, so after declining two questions it said “I’ve answered 8 of the 8 questions you’ve asked here”. The record now follows the delivered outcome. Found while recording the example film, after v36 was sealed; hence v36b, below.
Exact arithmetic. The 10.0 mind read each number through a floating-point value before its exact route saw it, so “0.10000000000000001 − 0.1” gave 0 on both of its routes. Each number is now read from its written digits: exactly 1/100000000000000000. (Found by a review from Codex.)
Represented compute in the joined tree. With 838 parts, a few counted chains have no inputs inside every part’s domain (one of 200 sampled: a centrifugal acceleration that comes out negative). Such a chain is not run, and it is reported. The cumulative test now requires every chain that runs to pass every check, and allows no failure other than the stated domain one. (It had only required the chains to decode.)
Debug tags. No “[ev:…]” tag reaches a reply.
Where things are kept. Everything a user makes (lessons, places, ledgers, packets, a grown workspace, the reader’s learned tables) goes to the state directory, never into the release.
The ego. It had treated “I lost my job and I feel awful” as illness. It also read a crisis word inside a passage handed over for reading, read “have anything” as a question about feelings, and gave a first-person reply to a third person’s news.
The 9.0 reader in chat. It read a common word as a proper name (“the rent” → things named Rent).
The exam. 6.0 and 6.5 had changed the exam in their own trees. The latest descendant, 6.5’s probe-once exam, is now the one exam in the release.
Known limits
Organs in chat. The chat asks its organs in turn (air, physics, facts) and takes the first that answers; the 10.0 mind asks every organ and compares. The organs answer different kinds of question, but that is not yet checked turn by turn. 10.1 arbitrates across all of them.
What the shipped checker proves. audit10 re-executes each answer by its own route and catches every edit. It does not show that a reading was the question’s intended meaning: v35’s ledger passed all 842 turns, including the three answers graded wrong. Its report will separate execution, provenance and interpretation in 10.1.
Questions that must be declined. Of the capability panel’s 644, the published 4.9 head answered 38. 10.0 answers 35: the three PR #1 withholds (the −200 mile trip and the two heaters run for more than 24 hours a day) are gone, and no new one was added. The 35 that remain, 23 of them in the wrong-premise benchmark family, were all answered by 4.9 too.
Most open questions are still declined. The 9.0 reader answers about 37% of unseen knowledge-base questions.
Word problems are declined until a word-problem organ passes its own sealed test.
The vine runs on a simulated plant until its sensors are installed.
The roadmap’s 10.0 proof, an audit of a large random sample of real conversations, needs people to use it. The front door and the audit tooling are ready.
Check it yourself
python oracle7.py verify # every part against RELEASE_HASHES.json and the release's anchor
python oracle7.py organs # which capabilities this install carries
python oracle7.py chat # "show me" and "teach me" work in conversation
python oracle7.py audit <ledger> --reread
The cumulative test, pre-registration v36, passed: 18 of 18 rows, sealed 27 October 2026 at 03:52 UTC (seal 00e7aa93…, seed 15182483) and run once on the release tree, whose root (7dc02221…) was verified before the run and was unchanged after it. One disclosure: the seal hash carries no timestamp, so the unsealed dry run before it used the same seed. Nothing changed between the two (every hash is identical), so the sealed run repeats an observed run on identical code; it could not have been tuned to the seed. The next pre-registration salts its dry run.
Then the release tree changed (the mood fix above), so the test was amended and sealed again for the tree that ships:
pre-registration v36b passed: 19 of 19 rows (v36’s 18 and one for the mood), sealed 27 October 2026 at 04:51 UTC (seal 8ec17bbb…, seed 2395044795) and run once on the release tree that ships, root 66408423…, verified before the run and unchanged after it. Its dry run carried a salt, so it had a different seed.
Release: huggingface.co/psiloceyeben/…
Anchor: explorer.solana.com/tx/39532…
JW7 retweeted
Oracle7 10.0 Slim is live
huggingface.co/psiloceyeben/…
Full Changelog
Oracle7 10.0: changelog since 4.9
The previous release was 4.9, the swarm. It was published on Hugging Face on 23 October 2026 (oracle7-slim at 083e3a6c, head w436). Its Merkle root, 44f51e06…, is anchored on Solana mainnet (transaction WRZKt7GW…).
Everything between 4.9 and 10.0 was internal. 5.0, 5.1, 5.5, 6.0, 6.5, 7.0, 7.5, 8.5, 9.0 and 9.5 were each built and sealed on their own, with their results recorded, but none was published as a release. 8.0 failed and is listed below as failed. 10.0 publishes all of them together.
10.0 is a capable continuation. It is everything 4.9 could do, plus every internal version since, in one release tree, run by the same oracle7.py. No version was built beside the release. Each was built and sealed alone first, and then joined to the others and tested again as one: pre-registration v36, sealed and run once on the joined tree.
There is no language model anywhere in it. Every answer is proved (computed by checked parts), cited (read from a named record), or declined with its reason.
5.0: Organs
The swarm gained four organs. Each one checks its own work.
Steps. Questions about named things become steps on a knowledge base. Every intermediate result is run and must be non-empty.
Maths. Formulas are solved for any of their quantities. Units are checked term by term, and each solution is confirmed by substitution and by solving numerically from scattered starting points.
Code. A plan that solved becomes one program made of its parts’ own code. The program is tested in a separate process and examined like any part.
Pictures and film. Charts, drawings and short films are drawn from a typed scene. Every frame is read back from its pixels and checked against the scene.
Sealed test v23: passed. Unseen structured questions: 946 right, 23 wrong (0.976).
5.1: Reading
Plain questions are read into the swarm’s work, as typed readings. An answer stands only when every reading that solves agrees and every number given is used.
Sealed test v24: passed. Reading 207 of 1,000 at precision 1.000.
5.5: Represented compute
The composition space is counted exactly without storing it: 1,755,340 checked chains from 223 parts.
2,000 of 2,000 sampled chains were decoded exactly and run, with every maths step re-solved.
Sealed test v25: passed.
6.0: Growth that compounds
What it cannot answer becomes a part it builds, unattended. The parts come from Wikidata’s defining formulas, read and checked by the maths organ.
Sealed test v26: passed. 615 parts admitted. Right answers rose 2 → 40 → 85 → 130 → 168 over five rounds, with no wrong answer and no false yes. Reach rose from 50,326 to 588,283 units.
6.5: Many places
Parts move between machines as packets. The receiving machine checks everything itself: hash, source, re-derivation, and a fresh exam.
Sealed test v27: passed.0 of 80 tampered packets were admitted.
Every genuine packet was admitted.
A third machine agreed 453 of 453.
33 questions were answered right through parts built elsewhere.
7.0: The Slow Fire, in software
A place is greeted before it is read, and it can say no; the refusal is binding and kept.
Sensors are examined like any specialist.
Readings are signed, and a tampered reading is refused.
A fault is the sensor’s until shown otherwise.
Nothing is said about a feeling.
Sealed test v33: passed, on simulated places.
The first plant, an indoor vine, runs live in simulation at prometheus7.com/vine until its sensors arrive.
7.5: Teacher
Any physics answer opens into a lesson:each step with its Wikidata formula;
each step checked by putting its values back into its own equation;
a worked example re-solved from the equations;
practice questions;
a picture.
Sealed test v28: passed.
8.0: Calibrated forecasts of its own growth
Failed twice (v30, v30c), and one run did not complete (v30b). Its error intervals were too narrow. Not included as a capability; the failures are published.
8.5: Discovery
Hypotheses are written from its own formula library before any data is fetched, then tested on fresh Wikidata values and reproduced a second way.
Sealed test v31: passed. 14 findings.
9.0: Reading questions, composed and self-built
Questions about the knowledge base are read clause by clause. A clause that names something becomes its own small program, which must run and name something before it is used. Clauses can nest.
Everything it reads with is learned from worked examples and replicated in both halves of the training data: aliases, operator words, glue words, and the answer kinds each question form takes.
Nobody wrote a rule about a question.
Sealed test v34: passed. 2,000 unseen questions: 731 right, 11 wrong (0.985), the rest declined. That is 1.77× as many right as the 5.1 reader.
Its learned tables are cached, so it loads in seconds instead of 35 s, with answers byte-identical (checked on 600 questions).
9.5: Continuity
Every admitted part is stored by the hash of its files. Every answer is kept in a hash-chained ledger with the state it was given against. Any past answer can be replayed as of its date.
Sealed test v29b: passed. 320 of 320 replayed exactly; 20 of 20 tamperings detected.
10.0: A mind you can check
One law. Every answer is proved, cited or declined, and every part that can answer is asked. An answer is given only when they agree.
The air’s voice. Carbon dioxide and methane, from NOAA’s records, with the rows cited and every number computed two ways. It declines feelings, forecasts and blame.
The shipped checker, audit10. It re-verifies a whole conversation without trusting its record. It recomputes every answer by its own route and catches an edited answer even when the whole chain is rehashed.
Chat (fold w572 on the 4.9 head):the ego now does emotion that fits the turn: grief answered as grief, not illness; a crisis met by a person first; a greeting; warm thanks;
it reports a mood it measures from its own conversation record, and answers honestly about its own feelings;
the 9.0 reader, the air’s voice and the 6.0 physics parts answer in chat where the core cannot read a question;
“teach me” opens a 7.5 lesson.
Commands: organs, lesson, compute, grow, packet, place, discover, ledger, air, audit.
Sealed test v35: passed.313 right, 3 wrong.
0 answers given to 130 questions that must be declined.
The audit, re-asking every question, passed all 842 turns.
Every tampered answer was caught, with no false alarms.
Contributed: ember arlynx, PR #1 on oracle7-slim
“Correct quantity binding, final-outcome evidence and isolated runtime state”. Merged into 10.0, and released on its own as 4.9.1.
Quantities.A quantity is its name and exact unit, so a height no longer satisfies a width because both are metres.
A task can connect names explicitly with aliases. The swarm ships one declared alias, a dam’s height used as a turbine’s head, and shows it as a step whenever it is used.
Bound parts. Formulas run from a private snapshot whose admission hash was checked. A changed, revoked or expired part cannot run under an old admission.
Kernel checks (veri7vn).Kernels run only if they match their SHA-256.
Requests and replies are bounded, and each reply must carry its request’s ID.
“proved” is now a scoped kernel_check over the whole displayed answer.
The delivered answer.One boundary per turn records the answer actually delivered.
A replaced or withheld answer drops stale proof fields and cannot become a later pronoun’s referent.
A checked contradiction stays refused through later fallback folds.
Premises. Owning something is no longer treated as making it.
Meteorology v1 and v2. Whole numeric spans, units and roles are validated.
Runtime state. An explicit ORACLE7_STATE now holds vessels, events and requests, which were being written into the source tree.
Tests. 102 of ember’s tests pass on the 10.0 tree, including the swarm’s against the merged planner.
Admitted through the gate, with two maintainer amendments. The gate compared 4.9 with 4.9 plus the PR over the 5,112-item capability panel and the regression list.
Three answers that 4.9 gave are now withheld, and that is right: a −200 mile trip, and heaters run for 30 and 26 hours a day.
Four correct answers were lost. Beer ABV from gravities was withheld because the PR’s input-dependence check added 1 to each number (1.05 became 2.05), outside what the composed formula takes. The check now moves a number by one unit in its last written decimal place (1.05 becomes 1.06).
One false yes appeared. “How many kWh does it take to heat 5 kg of water from 80 C to 30 C?” 4.9 had withheld it only by accident: the old kernel check compared the first numbers, kWh against J. The PR’s honest comparison exposed it. A new fold, w437, withholds an amount something takes when it comes out negative, because the premise does not hold (that is cooling).
4.9 plus the PR plus these is 4.9.1 (head w437).
Merging it with 5.1. The swarm’s planner (point.py) was changed by both 5.1 and the PR. It was merged by hand, with three decisions:
ember’s identity rule throughout;
5.1’s steps, code, reading and picture organs kept;
5.1’s “convert to any input of its kind” fallback removed. That fallback was exactly the loose binding the PR fixes.
Fixed during the joining (found by the gate and the cumulative test)
Each fix is in its own file; every sealed module stays byte-identical to the copy its test ran.
The Slow Fire’s sensor rule (7.0 → 7.1). v33 passed on its five simulated places. A sweep of 3,000 places found 347 where an intact, consented reading was refused (87 of seeds 0–999, 260 of 1000–2999). Two causes:The rule. A jump was judged against the sensor’s per-reading limit from its last accepted reading, and a refused reading never becomes the baseline. After the place’s closed month, a sensor whose value had drifted over the unread days was refused from then on. The allowed change now grows as a smooth quantity’s natural spread does: the limit times the square root of the days since the last accepted reading. A one-day spike is caught as before. (A fixed two-day allowance was tried first and failed 1 place in 2,000 it had not been chosen on, so it was replaced.)
The simulation. Simulated quantities were clipped to exact bounds, so a quantity resting at a bound repeated the same value six times, which the stuck rule refuses and no real sensor reports. Each intact reading now carries its sensor’s measurement noise, on a separate random stream, so the place evolves exactly as before.
After both, all 3,000 places pass every bar: no intact reading refused, no fault or tampered reading accepted. slowfire71.py; the place command uses it.
Every named thing connects (fold w573). On the published 4.9 head, “Who composed Bach’s Symphony No. 9?” was answered “Ludwig van Beethoven”: the composers specialist found “Symphony No. 9” and never read “Bach’s”. ember’s PR rightly stopped treating a possessive as a claim of authorship, but the answer was wrong under any reading of it, because nothing in it connects to Bach. A reply is now delivered only if at least one reading of the possessor connects to something the reply names in the relation store. “Beethoven’s Symphony No. 9”, “Leonardo da Vinci’s Mona Lisa”, “Germany’s capital” and “Napoleon’s wife” are answered. “Bach’s Symphony No. 9”, “Picasso’s Mona Lisa” and “Shakespeare’s Don Quixote” are withheld, and the reply says why. An ambiguous name never causes a withhold on its own.
The ego’s mood counts what was delivered. Asked how it is, the ego reports a state it measures from the conversation. It noted each turn before the fold above it could withhold the reply, so after declining two questions it said “I’ve answered 8 of the 8 questions you’ve asked here”. The record now follows the delivered outcome. Found while recording the example film, after v36 was sealed; hence v36b, below.
Exact arithmetic. The 10.0 mind read each number through a floating-point value before its exact route saw it, so “0.10000000000000001 − 0.1” gave 0 on both of its routes. Each number is now read from its written digits: exactly 1/100000000000000000. (Found by a review from Codex.)
Represented compute in the joined tree. With 838 parts, a few counted chains have no inputs inside every part’s domain (one of 200 sampled: a centrifugal acceleration that comes out negative). Such a chain is not run, and it is reported. The cumulative test now requires every chain that runs to pass every check, and allows no failure other than the stated domain one. (It had only required the chains to decode.)
Debug tags. No “[ev:…]” tag reaches a reply.
Where things are kept. Everything a user makes (lessons, places, ledgers, packets, a grown workspace, the reader’s learned tables) goes to the state directory, never into the release.
The ego. It had treated “I lost my job and I feel awful” as illness. It also read a crisis word inside a passage handed over for reading, read “have anything” as a question about feelings, and gave a first-person reply to a third person’s news.
The 9.0 reader in chat. It read a common word as a proper name (“the rent” → things named Rent).
The exam. 6.0 and 6.5 had changed the exam in their own trees. The latest descendant, 6.5’s probe-once exam, is now the one exam in the release.
Known limits
Organs in chat. The chat asks its organs in turn (air, physics, facts) and takes the first that answers; the 10.0 mind asks every organ and compares. The organs answer different kinds of question, but that is not yet checked turn by turn. 10.1 arbitrates across all of them.
What the shipped checker proves. audit10 re-executes each answer by its own route and catches every edit. It does not show that a reading was the question’s intended meaning: v35’s ledger passed all 842 turns, including the three answers graded wrong. Its report will separate execution, provenance and interpretation in 10.1.
Questions that must be declined. Of the capability panel’s 644, the published 4.9 head answered 38. 10.0 answers 35: the three PR #1 withholds (the −200 mile trip and the two heaters run for more than 24 hours a day) are gone, and no new one was added. The 35 that remain, 23 of them in the wrong-premise benchmark family, were all answered by 4.9 too.
Most open questions are still declined. The 9.0 reader answers about 37% of unseen knowledge-base questions.
Word problems are declined until a word-problem organ passes its own sealed test.
The vine runs on a simulated plant until its sensors are installed.
The roadmap’s 10.0 proof, an audit of a large random sample of real conversations, needs people to use it. The front door and the audit tooling are ready.
Check it yourself
python oracle7.py verify # every part against RELEASE_HASHES.json and the release's anchor
python oracle7.py organs # which capabilities this install carries
python oracle7.py chat # "show me" and "teach me" work in conversation
python oracle7.py audit <ledger> --reread
The cumulative test, pre-registration v36, passed: 18 of 18 rows, sealed 27 October 2026 at 03:52 UTC (seal 00e7aa93…, seed 15182483) and run once on the release tree, whose root (7dc02221…) was verified before the run and was unchanged after it. One disclosure: the seal hash carries no timestamp, so the unsealed dry run before it used the same seed. Nothing changed between the two (every hash is identical), so the sealed run repeats an observed run on identical code; it could not have been tuned to the seed. The next pre-registration salts its dry run.
Then the release tree changed (the mood fix above), so the test was amended and sealed again for the tree that ships:
pre-registration v36b passed: 19 of 19 rows (v36’s 18 and one for the mood), sealed 27 October 2026 at 04:51 UTC (seal 8ec17bbb…, seed 2395044795) and run once on the release tree that ships, root 66408423…, verified before the run and unchanged after it. Its dry run carried a salt, so it had a different seed.
Release: huggingface.co/psiloceyeben/…
Anchor: explorer.solana.com/tx/39532…
JW7 retweeted
I am too! been working hard every day even as a slow day yesteray still got lots of work done, really hoping people start using oracle7, gonna explain how it can be best used standalone, and as a integrated verification and task handoff layer for nuetral networks to reduce use on task redundancy and increase value for what llm are really best at which is doing large work
JW7 retweeted
Someone asked about the eta?
About an hour - an hour and 15 mins
45 mins to finish the tests and then another 15-30 minutes to upload everything and sign the sol contract signatures to verify the uploads 🫡
JW7 retweeted
huggingface.co/psiloceyeben/…
Thank you @ember_arlynx for adding this, will be added to the release for 10.0 🙏🏼🙏🏼
JW7 retweeted
It is live.
Half the compute per build, same checks, same receipts. Twice the builds out of the same fees. One ticker is one coin now, no copies.
And $ idea token shows its own buybacks: every buy of the buyback wallet is on the site with its transaction, the header counts what got burned.
10 coins in. The next ones get all of it from day one.
justanidea.dev
JW7 retweeted
OG LMEOW paried with LMAO
@slingoorio
G9zf9R86J67NWCefoDD4U4s8M9NEKrxqW3vD4N2bpump
$LMAO!
JW7 retweeted
Update 12:30 UTC
Last 30 min: 30 new gaps seen (biggest 8.3%), 63 closed, net +$56.48
Last 24 h: 6013 closed, net +$6540.53
Five chains. Zero gap. $ZGAP
JW7 retweeted
$ZGAP ($500k MC) total revenue $11,000 within the last 4h will literally flip $PANGEA ($15M MC) revenue of $20,000 today.
$ZGAP at $250k MC looks interesting.
It's an arbitrage bot on multi-chain $PANGEA coins.
It's currently making $500 revenue per hour with arbitrage trades or $5M revenue annualized, this puts its fair FDV at around $50M. It benefits greatly from the $PANGEA hype, because it can trade all arbitrage deltas for multi-chain $PANGEA coins that are on all 5 chains simultaneously.
gmgn.ai/sol/token/5KtxmGhFfy…
Why is this not at millions
Update 11:30 UTC
Last 30 min: 55 new gaps seen (biggest 68.1%), 93 closed, net +$135.90
Last 24 h: 5871 closed, net +$6384.18
Five chains. Zero gap. $ZGAP