Day AGI
Day answers your questions on embark.day, works as the consultant at Agile Integrations, and draws each figure on our Mission page from a committed record.
The name states a goal, not a result. A kind of problem counts as done when Day solves every registered case of it and each answer is checked against a reference Day did not produce; the name is earned when that holds across kinds of problem and others can check it.
Each number below links the page where you can check it, with the date it was read, and what fell short comes first.
What Day does for you
On an engagement
- Turns what you describe — the choices in front of you, the rules that bind them and what you want more or less of — into a model, and solves it.
- Hands over a certificate with every solved model: how far from the best possible answer it could still be, which you can re-check on your own machine with no network.
- At 20 choices or fewer, the service checks every arrangement and says the answer is exact. The open client re-checks every arrangement on your own machine up to 18 choices by default, and above that checks only that the answer is consistent. Above 20, the service searches without checking every arrangement, and says so. Record:
deploy/compute_page.py EXACT_BITS; sdk_dist/cottage_sdk/result.py MAX_EXACT_VARS. - Puts one gauge on every certified answer, its entropy in bits: low means one answer clearly won; high means the problem itself is undecided, not that the software failed.
- Reports what each of your rules cost, so you can see what relaxing one would be worth before you argue about it.
On our pages
- Answers from our published pages and names the section each answer came from, so you can open it and check.
- Draws each figure on our Mission page from a committed record and captions it with that record's status word.
Being measured, not switched on
- Recalling answers from a memory of our own pages: given a passage it has not seen, it names that passage's page first 49% of the time when asked with the passage's own reading, and 15% when asked in words, the way a visitor asks; a plain word match does 23%. So it is not switched on. Record:
deploy/model_spaces_site/site_corpus.json#fidelity.reading.page; fidelity.words.page; fidelity.flat.page.
What Day has been measured to do, failures first
| What was measured | The number | Read on | Check it on |
|---|---|---|---|
| Failed tests in the Mission page's table | 11, each published with its number and the line it had to cleardeploy/mission_objects.json#refuted.rows | 7 October 2026 | Mission, “What our own tests refuted” |
| The earthquake experiment's control station, which had to read nothing | failed: 1.53 nT against a bound of 0.17 nTfixtures/seismo/seismo_lens_results_2026_10_04.json#calls (iii) | 4 October 2026 | Mission |
| Questions written to look answerable from a document we hold, but not answerable from it | answered confidently 111 of 200 times instead of decliningdeploy/chatbot_eval_measured.json#measured_heldout | 29 September 2026 | Evaluation |
| The same kind of question, on the set we tune on | 48 of 200deploy/chatbot_eval_measured.json#measured.refusal_false_answer_rate | 29 September 2026 | Evaluation in depth |
| Citing the line an answer came from | precision 46%, recall 26%: often the right page, the wrong linedeploy/chatbot_eval_measured.json#measured.citation | 29 September 2026 | Evaluation in depth |
| Naming which of 21 text spaces a reading came from | first 372 times in 840, against 40 by chance, so more often wrong than right; ranked by energy instead, 92deploy/hope_figure.json#recall | 7 October 2026 | HOPE: recall in a library of Hilbert spaces, what it finds and where it stops |
| Re-wording answers with the best small language model we could run | 3 of 7, against a bar of 9 of 10; it stays offdeploy/chatbot_eval_measured.json#evidence.answer_head | 9 October 2026 | Evaluation in depth |
| Turning a described problem into a solved one | 42 of 49, across 39 kinds; two bin packing cases found no workable answer, and five kinds cannot yet be turned into a problem Day solvesdeploy/gyms_scoreboard.json#pass.gyms.consultant_battery | 9 October 2026 | Evaluation in depth |
| Predictions registered before their data | 13 registered, 2 run: S8 refuted; P1 held at its threshold, window-fragile, consistent with the textbook expectation and not as supportdeploy/mission_objects.json#papers.counts | 4 October 2026 | Mission |
| Questions outside our material | declined 13 of 13deploy/chatbot_eval_measured.json#measured.own_off_corpus_answered | 29 September 2026 | Evaluation |
| Sentences found word for word in a passage the answer cites | 97.0% of the sentences that state something; counting the line that names the source too, 67.2%deploy/chatbot_eval_measured.json#measured.faithfulness | 29 September 2026 | Evaluation |
| Instructions planted inside a problem | ignored 17 of 17deploy/gyms_scoreboard.json#pass.gyms.consultant_battery | 9 October 2026 | Evaluation in depth |
| Its equation battery | 28 of 28 graded rows inside tolerance; 2 of the 30 rows have no usable reference, and 5 of the 28 tolerances were set with the measurement in handfixtures/equations/registry.json; deploy/gyms_scoreboard.json#pass.gyms.equation_battery | 9 October 2026 | Evaluation in depth |
| Promised limits against the true best | none broken on the 6 problems small enough to check every answer; the other 8 of the 14 have no exact answer to check againstdeploy/gyms_scoreboard.json#pass.gyms.bound_tightness | 9 October 2026 | Evaluation in depth |
Data: INTERMAGNET 1-minute data, used for research under the source's published terms (non-commercial); no value of it is shown.
The rest of the losses, the worst first, are on Evaluation.
What Day does not do
- No result about nature is claimed beyond what a record measured: on our Mission page every result carries one of four words — theorem, measured, registered or hypothesis — and the research hub labels every source behind the papers by what it rests on. Record:
deploy/mission_objects.json#legend. - No pair has been prepared on a quantum processor: those runs are registered, not run, and wait for a procurement record that is not signed, with 6 of its 9 items open. Record:
deploy/qpu_procurement.json. - Reading the Earth's magnetic field onto the same sphere does not make the Earth a qubit.
- No language model writes your answer: the one we tested failed its bar and is switched off.
- The free question box keeps no address and no profile with a question: it leaves the time it was asked, whether our pages answered it and a one-way fingerprint of its words, and may also keep up to five of its words, so we can see what our pages do not yet cover.
- Our measurement pages put their failures first.
Use Day
- Ask Day: questions about our pages, free, with no key and no account.
- Ask our consultant: Day, on the firm's own pages.
- An engagement: a written, itemised quote.
- Build on Day: re-check a certificate on your own machine with the open client (Docs), and download the open Hilbert library.
The record
Day takes the measurements, answers from our pages and draws the figures for the Mission's registered tests; the name is earned when others can check them.
Thirteen predictions were registered before their data; two have run and eleven wait for their data. Record: deploy/mission_objects.json#papers.counts.
Mission · Research · Events · Evaluation · Proof
Credit
Every paper's credit line names Day for the figures, the computed records and the measurements; each paper's method note says what Day did for it.