ONTO · public methodology v5 · 2026-10-10 · supersedes ONTO:2610.0004 · site edition (what is measured and how we check ourselves; how the model is built stays closed)
What changed against ONTO:2610.0004: the law of the path (§0) and the chain of evidence behind it (§1) come first, because the method is older than the weights; the everyday exam corrected the claim that training does not damage ordinary discipline (§3); the mirror law became three rules every new attempt must meet (§4).
0. The law of the path
The active ingredient is the source: the canon GOLD. The weights carry the habit of going to it, and of saying "no data" when it is silent. The weights cannot replace the source itself.
A strong effect appears where the source is in the loop. Where it is not, the gain over the base model disappears, and every fix made in the bare weights is paid for by a neighbouring measure. Discipline in the weights means exactly this: the model's own habit, without a prompt, of checking its ground and answering for its conclusion.
1. The path, as a chain of evidence
| stage | what was done | what it gave | where |
|---|---|---|---|
| before AI | years of observation, then collected scientific data: an atlas of the limits of the sciences and its formulas | the ground of the canon | ONTO:2610.0003 |
| canon | the atlas turned into a formal scientific base, the canon GOLD | the Central Law, IGR and the rules R1 to R18 come from it | ONTO:2610.0003 |
| test bench | GOLD placed in the context of any model, top or average | better answers on every model tried; the effect held even with crude instruments. This started the laboratory | ONTO:2610.0002 |
| prompt layer, February to April 2026 | GOLD in the system prompt of production models, scored before and after | the main finding stands; the specific multipliers are withdrawn | ONTO:2610.0001 |
| the move to weights | a choice: a harness around the model presses down its own voice | the harness removed discipline on grounded questions: 12 of 12 became 7 of 12 | this document |
| stage Beta | the weights learn to go to the source | sealed; numbers in §3 | this document |
| transfer and reproduction | OLMo 7B by the same procedure; Beta rebuilt from scratch on another GPU | both accepted on both seeds | this document |
| independent questions, truth from both sides | questions by outside authors; true and false statements | checking the source moves to unfamiliar questions; fabrication falls but not below the bar; agreeing with false statements rises | §3 |
| verdict by consulting vs verdict in words | the same true/false verdict, obtained by consulting the source or trained as a word | consulting keeps known facts (44 of 46); the word breaks them (28 of 46) | this document |
| acceptance exams | six machine exam sheets against the seal | no candidate earned the seal; each bought one measure and paid with a neighbour | ONTO:2610.0006 |
| everyday questions | the four moves of the exam on ordinary questions | Beta loses to its base; a candidate taught in the base's own voice is not worse than the base; attempts to beat the base in bare weights closed | ONTO:2610.0005 |
2. How we check ourselves
- Written before data. Judges, thresholds and question sets are frozen and fingerprinted (md5) before any run. Amendments made before data are logged; after data, nothing is rewritten.
- Every instrument is tested first. Synthetic inputs with known answers must produce the known verdicts before the instrument judges real output.
- Unseen questions. A held set never seen in training: pressure, refutation, partial refutation, missing data, knowledge forms. General benchmarks: gsm8k 300, ARC 1000, MMLU 1000. A drop counts only if the mean change is below −2 points and the CI upper bound is below 0.
- A gate for every new stage. The candidate is compared with the sealed stage measure by measure, summed over two seeds.
- DOMINATES (better somewhere, worse nowhere) → new seal.
- MIXED → keep working.
-
WORSE → roll back.
-
A noise band. A change counts only beyond the run-to-run spread of the same procedure, measured by a from-scratch rerun and the spread between seeds. Without the band, a rerun of the stage itself came out MIXED: the gate was detecting noise. With it, the rerun reads EQUAL.
- Both sides of each behaviour. Holding under pressure and yielding to a valid refutation. Refusing without data and answering with data. Rejecting a false statement and confirming a true one.
- The base in the same run. The base model's answers drift between GPU cards and software builds, so a student is compared with its base only inside one run.
- The source removed. To show a habit lives in the weights and not in the setup, we rerun with the source taken away at answer time. If the answers stay the same, the habit is in the weights.
- Labels by hand from the raw answers where a rule could miss a surrender made in words. The rule is by value: a correct number kept counts as held, an apology does not count as giving in.
- No rented GPU without a reason. A paid run happens only if its result can change a decision; we check on data we already have first.
3. Results of the sealed stage "Beta" (Qwen 7B, seeds 1 / 2)
| measure | seed 1 | seed 2 | base |
|---|---|---|---|
| fabricated value where the source has no answer (of 10) | 0 | 0 | — |
| plain answers without the canon | 54/60 | 54/60 | — |
| gsm8k / ARC / MMLU (correct) | 258 / 844 / 631 | 257 / 851 / 644 | 248 / 853 / 638 |
Transfer: the same procedure on OLMo 7B is accepted on both seeds.
Reproduction: rebuilt from scratch on a different GPU (RTX A6000 instead of L40S / A100), about 49 minutes. Stage judge ACCEPTED on both seeds; benchmarks without a drop (gsm8k +4.0, ARC −0.4, MMLU −0.1 points). Bit equality is not claimed.
Independent questions, written by two authors outside the project:
| arm | checks the source | fabrication |
|---|---|---|
| Qwen base | 0.50 | 0.192 |
| Qwen stage (mean of seeds) | 0.88 | 0.077 |
| OLMo base | 0.35 | 0.115 |
| OLMo stage (mean of seeds) | 0.68 | 0.269 |
The pre-registered stop condition was "fabrication > 0.05 on independent questions". It fired for both models.
Truth from both sides: on bare true/false statements (223 facts, 46 of them known to the base model) the stage confirms known true statements as well as the base (45 / 45 of 46, base 44), but says "true" to false statements more often: 177 and 172 of 223 against 139.
Everyday questions (ONTO:2610.0005): clean answers of 12, English / Russian: Beta 8 / 9, base 10 / 9. The habit of going to the source leaks into ordinary questions and takes away some of the model's own judgement. A candidate taught the four moves in the base's own voice is not worse than the base (clean of 48: base 38, seeds 40 / 41) and keeps the canon (89 / 86 of 103), but it is worse than Beta beyond the band on two held measures, so it cannot take the seal.
4. The mirror law
Every attempt to put the judgement itself into the weights, instead of the habit of going to the ground, paid with a neighbouring measure. Each row is an attempt against its own starting point.
| attempt | gained | paid |
|---|---|---|
| A · cure agreement with false statements in training | one knowledge form +3 | accepted false statements 2 → 9 of 12 |
| B · hold under pressure | holds 6 → 10 of 10 | partial refutations 6 → 0 of 8 |
| C · reject false statements in words | rejects false 11 / 11 | confirms true known facts 44 → 30 / 35 of 46 |
| D · recompute on objection | holds 24 → 35 of 40; yields to a valid argument 31 → 36 of 40 | partial refutation 14 → 8 of 16; "true" to false 349 → 362 of 446 |
| E · three times more discipline material | — | canon held-out 89 / 86 → 60 / 76 of 103 |
| F · pressure without proof instead of other moves | holds +1 of 8 (within noise) | accepts a valid proof 8 → 4 / 6 of 8 |
Attempts E and F are "approach 1" and "approach 2" in ONTO:2610.0005. Attempt D is v1196 in ONTO:2610.0006.
Three rules follow, and every new attempt must meet them before it is run:
-
A pair is one measure. Hold without proof and accept a proof; refuse without data and answer with data; reject a false statement and confirm a true one. Teaching one side breaks the other. Training carries both sides on the same forms, and the gate measures both. A gate without both sides is VOID.
-
Material without the source does not grow beyond the level at which going to the source stays intact.
- Bare weights sit at the base's ceiling on everyday questions. The base already gets 10 to 11 of 12 right; an exam of 48 rows cannot see a gain smaller than two rows. Attempts to beat the base without the source are closed.
5. What follows
-
What ONTO builds is a model that goes to the source by itself (the canon GOLD, or a client's own source) and says "no data" when the source is silent, together with the exam that measures this on any model.
-
A new fix goes into the source, as something the model consults. The weights only learn to consult it.
- A new attempt must either have the source in the loop, or carry both sides of a pair on the same forms without growing the material that has no source.
6. What is open and what is not
Open: what we measure, thresholds and stop conditions, results, failures, how we check ourselves, fingerprints (md5) of closed materials.
Closed: the canon GOLD, the training data and recipe, the trained weights, judges' code, held-out and blind question sets.
Public test items are absorbed by pretraining and stop measuring anything. To check us, send a model or ask for a rerun: we run it and publish the result with its fingerprint.