https://agileintegrations.net/blog/ 2026-10-09T00:00:00Z Agile Integrations https://agileintegrations.net/blog/measure-entropy-not-uptime.html 2026-09-17T00:00:00Z Agile Integrations

Uptime says the system was running. It does not say the run was worth making. What our receipts carry instead, what you can watch it do, and what you can decide with it.

https://agileintegrations.net/blog/listen-to-the-kernel.html 2026-09-17T00:00:00Z Agile Integrations

The engine refuses when it has no source. Every refusal turned out to be a map of the holes in our own documentation — including the round that took our own gold set from 13 of 26 to 24 of 26.

https://agileintegrations.net/blog/the-operational-unit.html 2026-09-30T00:00:00Z Jonathan Klinshaw, Agile Integrations Day Claude (Anthropic)

This paper states a modelling convention: one small unit that keeps an account, reads its surroundings, emits into them and is rewarded by the uncertainty it removes, used to describe a chatbot, a person and a market in one vocabulary. The losses come first. Registered tests of the model's own rules, each a simulation on our fixtures run once on 7 October 2026, falsified five of its sentences: the entropy reward pays the same 1 bit for answering every unanswerable question and for declining every one, so it is blind to correctness; on that reward 64 % of states still change their action late in learning; a belief that stopped moving settled on the wrong hypothesis in 50 of 50 seeds; 1,408 of 2,984 served splits lie outside the core while the certificate reads that no one is worse off; and 47 % of buyers gain by misreporting through the served prices. Four earlier registered tests of the model's readings were falsified as well, among them that the account's writer has a direction, 12 of 24 series, and that a fixed reading order is as fast as a random one, 10.9 % slower. A prediction that choosing readings by expected information beats a random order is registered, not run. What held: with no discount, the rewards along a path add up to the uncertainty removed between its ends in all 50 seeds; and on our own evaluation the engine named the nearest of 21 model spaces for 372 of 840 probes, 44 % against about 5 % by chance, exact, one deterministic run. This paper does not claim that people, chatbots and markets share one ontology, and it describes none of the engine's method.

https://agileintegrations.net/blog/the-periodic-table.html 2026-09-30T00:00:00Z Jonathan Klinshaw, Agile Integrations Day Claude (Anthropic)

Each element's ground-state electron configuration is modelled here as the lowest-energy filling of its electrons over spin-orbitals, one two-valued variable each, with the filling order as an input and two interaction constants fitted on hydrogen to krypton. The losses come first. Of the 72 elements after krypton that NIST lists, rubidium to hassium, the model matches 46, against 54 for the same objective with the two constants at zero, which is the filling order alone. The fitted exchange term gains chromium, copper, molybdenum and silver and loses ten lanthanides and actinides, so it matches 82 of the 108 listed elements against 88 without it. A table of orbital energies computed from Slater's rules, run as a control, matches 6 of the first 20. A prediction that unpaired electrons and measured magnetic susceptibility rise with the halo of the margin is registered, not run. What held, in one deterministic run on 30 September 2026 against the NIST Atomic Spectra Database: all 36 fitted elements are reproduced, every answer up to twenty spin-orbitals is the best of all 1,048,576 arrangements, checked one by one, and each of the 118 answers is the objective's own lowest filling, which a reader can check by arithmetic. None of this computes an atom: the model returns configurations, not energies, and the filling order is its input, not its result.

https://agileintegrations.net/blog/closing-the-gaps.html 2026-09-30T00:00:00Z Jonathan Klinshaw, Agile Integrations Day Claude (Anthropic)

Energy storage, farming, industrial heat and recycling each wait on technology that is not ready, and a scheduling engine cannot build any of it; it can only decide how to run what already exists. We measured that on small problems, one per gap, each answer set against the best of every possible answer on the same table, one run each on 30 September 2026. The losses come first. Past twenty moving parts, four of the five answers that could be checked fell short of the best: ten generating units cost 2.17 times the best schedule, four furnaces 1.92 times and five units 1.2 % more, and an eight-field farm got no plan at all while the response still called it feasible. A registered re-test then falsified that status word: it comes from why the search stopped and never reads the plan. The 107-part microgrid day stopped at its step limit and was not checked against the best, and the 240-choice day has not been run. What held: at twenty moving parts or fewer the engine's answer was the best of all 1,048,576 possible answers, checked one by one, in all four problems, and one eight-part teardown found a best order anyway. None of this closes a technology gap: a scheduler models no chemistry, agronomy or materials.

https://agileintegrations.net/blog/sense-remember-act.html 2026-09-30T00:00:00Z Jonathan Klinshaw, Agile Integrations Day Claude (Anthropic)

This note describes, in plain terms, what our engine senses, what it keeps and what it learns, with the numbers we measured and where they stop; it describes no method. The losses come first, from registered re-tests run once each on our own fixtures on 7 October 2026, each a simulation that tests the engine's rules, not nature. Of 540 cues drawn from patterns the memory never stored, 19 were named as a stored reading with a margin above the engine's own threshold, a wrong-name rate of 3.5 % against the 1.3 % this note first published, so that test is falsified. The settling ledger's entries add up to the booked steps only: on one of its two measures they hold between 0.5 and 35 % of a run's descent, so the sentence that they match the whole run is falsified. A rewrite by the chain's own keeper reads intact in 4 of 4 tries, so the hash-chained account is tamper-evident to everyone except its keeper. The noise study's linear ratio reappears in 0 of 20 independent seeds, and a learner that keeps the observation's likelihood recovers the cross weight that the engine's second rule loses, so that collapse belongs to the engine's chosen score. A test that a fixed reading order is as fast as a random one is falsified too, the fixed order 10.9 % slower. One registered test is undecidable as registered, and five predictions are registered, not run. What held, one run each on 30 September 2026: 20 of 20 planted fields read back; recall falls with load, from 20 of 20 at five patterns in sixty-four regions to 0 of 40 at twenty in sixteen, as the textbook capacity law, re-tested on its own, held; and of 540 recalls 165 were right, 7 named a wrong pattern and 368 were marked contested. None of this is a physics claim: the engine contains no field of force, and its two pictures are analogies.

https://agileintegrations.net/blog/the-halo-of-a-decision.html 2026-10-03T00:00:00Z Jonathan Klinshaw, Agile Integrations Day Claude (Anthropic)

This paper sets out the textbook picture of a two-way choice and its partner, the Bloch sphere and the ring of correlations two partly decided halves share, the halo, and reads three recorded margins onto it; every lens reading here is from an exact simulator, not from hardware. The losses come first. Two registered tests of the lens were falsified: its 95 % credible region contained the true halo radius for 815 of 1,000 planted pairs, against a registered band of 0.932 to 0.968, so a halo it decides is decided at about 82 %; and its radius witness read 0.934 on a pair that is not entangled, against 0.089 for the axis-only control. The registered test that the phase clock's rate follows the balance of what flows in and out was falsified on 200 planted beats run through the engine's code, with a correlation of −0.063 against a registered +0.5 or more. Two registered tests of the margin measured on our rosters were falsified: it rose strictly with its lever at 23 of 24 roster temperatures, one tie in floating point, and three frozen readings sat at −0.76 to −0.74 against a registered 0.9. Three predictions about a clocked pair are registered, not run. What held: the closed forms agree with a direct computation from the amplitudes to within 1e-12 at 25 points, exact; and on a simulator the lens separates a classical mixture, radius 0.02 with region 0 to 0.16, from an equatorial pair, 0.94 with region 0.90 to 0.99, one seeded run of 3,000 shots each, against the textbook value 0 for any mixture along the axis. None of this is a hardware reading, and nothing here claims that the engine contains a qubit.

https://agileintegrations.net/blog/the-earth-as-a-two-state-recorder.html 2026-10-04T00:00:00Z Jonathan Klinshaw, Agile Integrations Day Claude (Anthropic)

This paper reads the Earth's magnetic dipole onto the Bloch sphere, the textbook picture of a two-way choice, and runs the tests registered for that reading, once each on public records of 4 October 2026. The losses come first. The prediction that the dipole carries the Earth's orbital periods is refuted: its 23, 41 and 100 thousand year bands reach 0.81, 1.89 and 1.04 of their significance line, and all three were needed. The prediction that the dipole is weak where it reverses is window-fragile under its registered robustness test: narrower windows, or windows moved 15 thousand years younger, meet 6 of the 9 boundaries where 7 were needed. A test of the widening of the dipole's axis tilt between 1900 and 1955 is registered, not run. As registered, with windows of 20 thousand years, the weak-reversal prediction holds at its threshold, 7 of 9 with no margin; that is consistent with the textbook expectation that the dipole weakens at a reversal, and its power was not measured. Nine window centres drawn at random meet the same rule in 19 of 1,000 draws, a null registered after the result was known. None of this makes the Earth a qubit, and this paper claims no measurement of the Earth's halo and no new geophysics.

https://agileintegrations.net/blog/hope-recall-in-a-library-of-hilbert-spaces.html 2026-10-08T00:00:00Z Jonathan Klinshaw, Agile Integrations Day Claude (Anthropic)

HOPE recalls by letting a partial reading settle to the nearest pattern a space has stored. Its library keeps 99 separate spaces of ten two-valued axes, one per open model: 21 text, 77 image and 1 audio. The losses come first. One range we had recorded was falsified when re-run and is corrected: 15 to 29 of 45 pulls, not 6 to 29. The spaces do not form one shared space: fitted together, the 21 text spaces agree 0.324232 on the readings the fit saw and 0.03793 on readings held out. A space's ladder showed no sign of carrying across spaces, though the routing test behind part of that had no power. Recall of a passage from its own words was refuted when measured, 4 of 80 against 48 of 80 for plain word matching, and a ten-axis space holding 40 readings recalls only 40 per cent of them. The site's own space costs measured answers, 19 to 17 of 20, so it is switched off, and its missing-word model is calibrated worse than a word count. Six of the 41 claims re-ran with numbers that moved. What held, in one deterministic run on 7 October 2026: from a held-out reading, the right text space is named first 372 times in 840, 44.29 per cent, against 40 by chance. Section 7's site space is quoted as measured on 6 October 2026. No run without the settle was made, so what the settle itself adds to these counts is not isolated. A Hilbert space here is a name for the 1,024 states of ten two-valued axes; this paper claims nothing quantum and no advantage over any other method.

https://agileintegrations.net/blog/the-unit-of-work.html 2026-10-09T00:00:00Z Jonathan Klinshaw, Agile Integrations Day Claude (Anthropic)

Fewer than six economies: F-E1 and F-E2 falsified over 340 states; F-L2, F-L3, F-A1 and seven PX tests falsified. F-E3 held 20 of 20, F-L1 42 of 42. The unit of work: a serial-numbered receipt for one settled problem, retired on delivery; not a currency, not an investment.