Harman
One gate, ten engines
Harman is how we put hard questions to AI agents. A gate reads the question and sends it to the engine that fits. Each engine accepts one kind of input and answers with a verdict from a fixed list. When something stays unresolved, the engine names it, so the next step is a concrete build rather than “further work needed”.
The engines
| Engine | The question it answers | Main verdicts | Status |
|---|---|---|---|
| pure-dialectic | Is this “X or Y” a real opposition? | real conflict with an operating rule · symmetric pair · category error · bad infinity | witnessed |
| agentic-pure-dialectic-tree | When a decision branches, does each branch hold on its own? | the same verdicts, node by node | candidate |
| freqdialectic-audit | Do N agreeing witnesses really count as N? | independent · partly independent · collapsed into one source · bounded | witnessed |
| dialectic-lifecycle | Which stage of a long inquiry is open, and should a closed one reopen? | open residual · closed · reopen · blocked · abstain | candidate |
| loop-closing | Is a named open item really closed? | closed · shrunk · named falsification · redirected · bad infinity | witnessed |
| reflexive-consolidation | Which signal did the summary leave out? | gap · retained signal · justified omission · quiescent | candidate |
| begriff-scout | Are the terms sharp enough to tell experiments apart? | lexicon ready · determinate · raw label · dead | witnessed |
| marx-filter | Does a Hegelian resolution stay in thought, or does it name the practice where it would be tested? | passes · stays in thought · treats objectivity as alienation · double error · practice unnamed | candidate |
marx-filter is new; the gate calls it after an opposition has been turned into an operating rule, or when asked directly. Two more engines check delivery rather than method: storm-conductor asks whether one multi-agent run is ready to hand over, and grand-conductor asks whether several runs can be put in order. A helper, dialectic-iteration-loop, sits beside them and runs bounded experiments that end in keep, discard, retry or reframe.
Status says how far an engine has been checked by someone other than us. seed: written, not yet used in real work. candidate: used in our own work, not yet approved by an outside check. witnessed: approved on test questions written before the test by a judge from a different model family that had no part in building or running it, and only when every engine it relies on is witnessed too. storm-conductor and grand-conductor are at candidate status.
Audits on Hegel’s categories
Five audits sit beside the engines. Four of them work through one part of Hegel’s Logic; the fifth is a discipline for the moment before an irreversible step. They are newer than most of the engines: logik and mechanism-teleology-audit have been used in our own work, the other three not yet, and none has yet passed an independent test.
| Audit | Part of Hegel’s Logic | The question it asks | Status |
|---|---|---|---|
| modal-audit | Actuality, possibility, necessity (Encyclopaedia §§143–149) | Is this claim merely possible, actual, conditionally real, or necessary? | candidate |
| causality-reciprocity-audit | Causality and reciprocity (§§153–159) | Does A drive B, or is there feedback or a hidden common cause? | recipe over two lifecycle stages |
| mechanism-teleology-audit | Mechanism, chemism, teleology (§§195–212) | Is the agent’s purpose its own, given from outside, or absent? | candidate, shadow mode only |
| autonomous-counterline | The method of the Logic (§§239–243) | Which counter-line follows from a line of inquiry’s own commitments? | candidate |
| logik | Being, essence and concept as three working layers | Before an irreversible or one-shot step: fire, hold, or change the game? | not yet rated |
What it argues against
Harman does not use Hegel as an inspiration. It takes a side in an argument that is still open, and its engines are that side’s arguments in working form.
The other side is the logic analytic philosophy was built on. Bertrand Russell stated its starting point in 1914: in every proposition there is, “besides the particular subject-matter concerned, a certain form”, and “the knowledge of logical forms is something quite different from knowledge of existing things”. Logic keeps the form; the content is left to someone else. Two years earlier, in The Problems of Philosophy, he had drawn the same line through the world: we can know how physical objects are related, but the objects themselves “remain unknown in their intrinsic nature”. A footnote in the same book grants that Kant’s thing in itself is “identical in definition” with his physical object.
That is one cut made twice. Form is cut from content, the thing is cut from its relations, and whatever falls on the far side is declared out of reach. Hegel named the cut in the introduction to his Science of Logic: logic so far had rested on “the separation of the content of cognition from its form” (our translation). Joseph Dietzgen, a tanner by trade, answered it in The Nature of Human Brain Work (1869): a thing is anything “in itself” only because it is something for other things. In his Letters on Logic he compares the formal logician to a potter who says he shapes only the vessel and has nothing to do with the clay. Form and body, he writes, “can be separated by words only, but not by actions.”
Three of the engines carry that answer:
- Nothing is left in itself. The unknowable remainder has a modern name: “further work needed”. loop-closing does not accept it. An open item is named precisely enough to build against, or the line ends in a named falsification that states what it closed.
- Relations come first. The atomist picture treats each agreeing source as a separate fact. freqdialectic-audit counts how many separate sources the agreement actually rests on, and treats independence as something to show, not to assume.
- Content before form. Formal logic gives “X or Y” a truth value. pure-dialectic first asks what kind of opposition it is: a real conflict, a symmetric pair, or a category error. Dietzgen did not throw out the law of identity. He called it fine “for household use”, and sound for as long as it stays aware of its own narrowness.
We call the cut ideology, not only error, because it does work for someone. Dietzgen made the charge himself: the formal logicians hand the hard question, how thought is connected to the world, over to metaphysics and religion, where official wisdom can go on serving “the ruling classes”. His chapter on morality shows the same shape in ethics. Every ruling class states its own conditional rules as eternal truths: the Russian noble finds serfdom rational, the English bourgeois finds the “liberty” of his wage worker rational. A truth with its conditions removed can be neither tested nor taken apart. The marx-filter is where Harman checks for this. A resolution that names no practice in which it would be tested leaves its object where it was; that was Marx’s objection in 1844.
Harman’s own verdict lists are forms too, and the argument applies to them. The condition we hold them to is Dietzgen’s: a form is usable while it knows its scope. Each verdict says what it covers and names what it leaves open, and the status column above says how far each list has been checked.
End to end
Hegel-inspired prompting for language models has existed since 2025, and a few open-source skills now walk an agent through a contradiction to a synthesis. Those skills apply one sequence of reasoning to whatever question arrives. Harman covers what comes before and after it as well: it types the question before anyone resolves it, checks whether agreeing evidence is really independent, judges whether an open item is truly closed or names what is left, audits what a summary dropped, and checks a multi-agent delivery up to the hand-over. As of October 2026, in our scan of arXiv, Google Scholar, Crossref and GitHub, we found no other system that covers that whole arc in one set of fixed-verdict agent skills.
Where it stands
Harman runs in our own work. Four engines are now witnessed: pure-dialectic, loop-closing, freqdialectic-audit and begriff-scout. For each, an outside check approved its output on test questions written before the test. What that shows is narrow: the test sets were small and mostly easy, and freqdialectic-audit was tested on one of its faces only. dialectic-lifecycle also passed an outside check on a public benchmark, ProcessBench, where it scored the same as the same model working without it; it stays at candidate because it relies on reflexive-consolidation, which has not yet been checked from outside. No engine has yet been shown to beat a capable model that works without it. The other engines are at candidate status. Of the five audits, three are candidates, logik is not yet rated, and causality-reciprocity-audit is a recipe over two lifecycle stages whose standing as a separate audit is still untested. The engines are private and not published as code.
Sources for the section above: Russell, The Problems of Philosophy (1912), chapters 3 and 8; Russell, Our Knowledge of the External World (1914), lecture II; Hegel, Science of Logic, introduction; Dietzgen, The Nature of Human Brain Work (1869), chapters 2 and 5, and Letters on Logic and The Positive Outcome of Philosophy, in the 1906 English translation by Ernest Untermann.