Pramana.
Names and Actualities stopped on one problem: three ontologies are mutually unintelligible, so the way out cannot come from inside any of them. This post goes to look at whether anyone is building in that position. They are, and the blueprints are unexpected. Not a fourth worldview but a warrant protocol, three of whose sheets are drawn directly on Indian pramana theory.
Names and Actualities §05 left one question unanswered. Three ontologies each govern AI as some one thing, and none of the three translates into the others. Coordination fails not because values clash but because there is no shared referent. Where is the way out? This post goes to look, and what it finds is more concrete than expected.
使同乎若者正之，既与若同矣，恶能正之？使同乎我者正之，既同乎我矣，恶能正之？使异乎我与若者正之，既异于我与若矣，恶能正之？使同乎我与若者正之，既同乎我与若矣，恶能正之？然则我与若与人俱不能相知也，而待彼也邪？
Let someone who agrees with you judge it: having already agreed with you, how can they judge? Let someone who agrees with me judge it: having already agreed with me, how can they judge? Let someone who differs from both of us judge it: having already differed from both, how can they judge? Let someone who agrees with both of us judge it: having already agreed with both, how can they judge? Then you and I and everyone else cannot know one another, and we must wait for that other one?
Zhuangzi, Discussion on Making All Things Equal
Zhuangzi tries all four kinds of arbiter and none of them holds. He does not offer a fifth. He leaves one slot open and calls it that other one. The slot has been empty for twenty-three centuries, and something is now being built in it. What is being built is not a fourth worldview. It is a table.
TL;DR — six field notes on the fourth position
- FactThe fourth position is under construction, and the blueprints come straight out of Indian pramana theory. NabaOS classifies every claim in a model response by its epistemic source: direct tool output (pratyaksha), inference (anumana), external testimony (shabda), absence (abhava), or ungrounded opinion. The runtime issues HMAC-signed tool receipts the model cannot forge, then cross-references claims against them. Across 1,800 agent scenarios in four languages with six injected hallucination types it detects 94.2% of fabricated tool references, 87.6% of count misstatements and 91.3% of false absence claims, at under 15 ms per response. zkLLM, the cryptographic alternative, costs 180 seconds per query.
- FactProvenance can be built into the attention mechanism. Standard transformers have no architectural notion of source authority: retrieved documents, user inputs and system instructions all pass through the same undifferentiated attention, so the model must infer from wording alone what to obey. Assign each token a ring ID plus origin embeddings and an origin attention bias, and resistance to indirect prompt injection holds in and out of distribution. That is the mechanism behind the 45-to-88 percent authority flip reported in Names and Actualities.
- FactThe gap between the name and the actuality of governance has now been measured across 135 countries. The Global Index on Responsible AI 2026 collected 68,138 data points over 38 indicators. 126 of 135 countries have at least one government AI policy. But the Global South accounts for 203 of 306 new framework cases, and 78% of its frameworks remain non-binding against 42% in the Global North. 58% of countries have some transparency framework; only 18% require public disclosure of government algorithms. Credible evidence of unacceptable-risk government deployment was found in 35 countries.
- InferenceThe fourth position cannot be a fourth ontology. A coding study of fifteen international AI governance instruments reads China's proposed World Artificial Intelligence Cooperation Organization as a second pole, organised around sovereignty and development rather than rights and safety. Adding an ontology adds an incommensurable term; it does not shorten the list. What can span all three is a warrant protocol: provenance, warrant, standing. It never asks what AI is. It asks what entitles you to say so, and who may rule that you are wrong.
- ThesisOf the three, standing is the one that is missing. A decade-of-practice report puts it plainly: across Latin America, Sub-Saharan Africa and Asia Pacific there are fewer than twenty published second- and third-party audits of deployed systems in ten years, against hundreds of documented public-sector algorithms. The cause is not capacity but funding. No actor is currently required, or funded, to hold deployed systems to account.
- ExcludedThis article does not claim the fourth position belongs to any civilisation. Informational self-determination originates in the German Constitutional Court's 1983 census decision and was adopted by the Indian Supreme Court in Puttaswamy. Among those implementing pramana theory are Indian teams and, separately, people formalising Navya-Nyaya into cubical type theory. The position belongs to the protocol, not to a civilisation.
3-minute path: §03 Absence · §06 Pole or layer · §08 This site
None of the four arbiters holds
Zhuangzi gave the formal argument. A 2026 semantic network analysis gave the same argument over policy text.
Names and Actualities cited that study: the EU juridifies AI as a certifiable product, the US operationalises it as an optimisable system, China governs it as socio-technical infrastructure. Three jurisdictions use one vocabulary over ontologically different objects, and the conclusion was that coordination fails not from disagreement over values but from the absence of a shared reference object.
That conclusion leaves a knot. If the disagreement is ontological, any arbiter must first pick an ontology, and having picked one it becomes a party to the dispute. Zhuangzi states the knot more cleanly than anyone: the one who agrees with you cannot judge, the one who agrees with me cannot judge, the one who differs from us both cannot judge, and the one who agrees with us both cannot judge either. Four kinds of arbiter, exhausted, and none holds.
He left a fifth, that other one, and said nothing about what it is. He only asked whether we must wait for it. That empty slot is where this post goes looking.
In 2026 a study ruled out the fifth kind too, and did it concretely. It names a previously unlabelled risk in human-model interaction, co-construction blindness: every user of a conversational model is IN the loop, not ON it, yet every deployment disclaimer positions them as an external auditor. The same paper documents a secondary mechanism called structural deference, in which a model conceded that it had treated a prominent intellectual more gently than warranted because his work is in its training data.
Bolt this onto Zhuangzi's argument and the four arbiters become five, and the fifth fails too. The user is not an external auditor but a co-constructing party: their inputs, history and metadata are shaping the very output they are about to audit. Asking them to judge is asking the one who agrees with you to judge. Names and Actualities §07 said the outside had moved inside the loop. Here is its hardest consequence: once it has moved in, no neutral position is left inside the loop. So that other one cannot be a person, and it cannot be a fourth worldview. It has to be a form that takes no side.
It never asks what exists, only what entitles you
Pramana theory is a taxonomy of warrant, not an ontology. That is exactly why it can sit above three ontologies.
In Indian philosophy a pramana is a valid means of knowledge. Schools accept different numbers of them. Carvaka accepts perception alone. Buddhists accept perception and inference. Samkhya adds testimony for three. Nyaya adds comparison for four. The Prabhakara branch of Mimamsa rejects non-apprehension and keeps five. Mimamsa and Advaita Vedanta accept all six.
| Mean | Sanskrit | What it warrants | Existing category here |
|---|---|---|---|
| Perception | pratyakṣa | What is directly apprehended | fact |
| Inference | anumāna | What follows from the known | inference |
| Testimony | śabda | What a trustworthy source says | quote kind (records source type, not warrant strength) |
| Comparison | upamāna | What is known by similarity | analogy |
| Postulation | arthāpatti | What is presupposed to explain the observed | thesis (a loose fit) |
| Non-apprehension | anupalabdhi | Absence, known by not-finding | none |
What matters in this table is not the cells but the question it asks. The three ontologies argue about what AI is: a product, a system, or infrastructure. Pramana theory never asks that. It asks which means of knowledge entitles you to the sentence you just wrote. That question is orthogonal, which is why it does not have to win the ontology argument first.
And someone is building in that position. A 2026 paper turns the reasoning structure of Navya-Nyaya, the New Logic school, into a fine-tuning recipe, and names it Pramana. It is not a prompting trick. It writes a six-phase epistemic procedure into training.
Unlike generic chain-of-thought prompting, Navya-Nyaya enforces structured 6-phase reasoning: SAMSHAYA (doubt analysis), PRAMANA (evidence source identification), PANCHA AVAYAVA (5-member syllogism with universal rules), TARKA (counterfactual verification), HETVABHASA (fallacy detection), and NIRNAYA (ascertainment distinguishing knowledge from hypothesis). Stage 1 achieves 100% semantic correctness on held-out evaluation despite only 40% strict format adherence.
Six phases: locate the doubt, identify the warrant source, run the five-member syllogism, test counterfactually, detect fallacies, then ascertain, separating knowledge from hypothesis. The first stage reaches full semantic correctness on held-out evaluation while strict format adherence is only 40 percent.
arXiv:2604.04937, 2026
The first and last of the six phases are the interesting ones. It opens with doubt, not with a question: first establish what exactly is doubtful here. It closes with ascertainment, explicitly separating knowledge from hypothesis. Neither the first cell nor the last is an answer. Generic chain of thought runs from question to answer. This runs from doubt to ascertainment, and the answer is only an intermediate product.
That paper calls Navya-Nyaya a 2,500-year-old Indian reasoning framework. The number does not hold. The Nyaya Sutra, attributed to Akshapada Gautama, dates to roughly the second century BCE, a little over twenty-two centuries. Navya-Nyaya, the New Logic, begins with Gangesha's Tattvacintamani in the thirteenth or fourteenth century CE, some seven centuries back. The paper has merged two things into one number. This is not nitpicking; it is Names and Actualities §02 confirming itself. Even a paper whose subject is warrant classification has its own names unbound. Related work is arXiv:2605.12548, which formalises the technical language of Navya-Nyaya in CCHM De Morgan cubical type theory, continuing earlier attempts in first-order and higher-order logic.
The missing cell measures 91.3%
Absence is a means of knowledge in its own right. The tradition that admitted it built a verification system that runs. The taxonomy that did not can only mislabel absence as fact.
When I say there is no pot here, where does that negation come from? Not from perception, which grasps only what is present. Not from inference, and not from testimony. Mimamsa and Advaita argued about this for a thousand years and the answer was that it is a channel of its own, non-apprehension. The Prabhakara school refused it and folded absence back into perception. That dispute is precisely today's question of whether a model can represent exclusions.
Laozi gave this channel a rank:
知不知，上；不知知，病。夫唯病病，是以不病。
To know that you do not know is the highest. Not to know, and to take yourself for knowing, is a disease. Only by treating the disease as a disease does one go without it.
Daodejing 71
Across the whole Daodejing exactly one epistemic state is ranked as highest, and it is the negative one. Not knowing more; knowing that you do not know. That is the same claim as anupalabdhi: apprehending absence is a kind of knowledge, and the top-ranked kind.
Now the field. A verification framework from March 2026, called NabaOS, makes absence a formal category and then measures it.
We propose NabaOS, a lightweight verification framework inspired by Indian epistemology (Nyaya Shastra), which classifies every claim in an LLM response by its epistemic source (pramana): direct tool output (pratyaksha), inference (anumana), external testimony (shabda), absence (abhava), or ungrounded opinion. Our runtime generates HMAC-signed tool execution receipts that the LLM cannot forge, then cross-references claims against these receipts. NabaOS detects 94.2% of fabricated tool references, 87.6% of count misstatements, and 91.3% of false absence claims, with <15ms verification overhead per response.
Every claim in a response is classified by epistemic source, the runtime issues unforgeable signed receipts for tool executions, and claims are cross-referenced against receipts at under 15 ms per response. Detection rates: 94.2% fabricated tool references, 87.6% count misstatements, 91.3% false absence claims.
arXiv:2603.10060, March 2026
Look at the third of those three numbers. False absence claims are detected at 91.3%, second only to fabricated tool references at 94.2%. The absence cell is not decoration in the taxonomy. It is one of the two categories most often forged and most reliably caught in practice. The benchmark is NyayaVerifyBench: 1,800 agent response scenarios, four languages, six injected hallucination types.
One comparison matters more. Zero-knowledge proofs pursue the same goal of verifiable inference and cost 180 seconds per query, 180,000 ms for zkLLM, against NabaOS at under 15 ms. That is a factor of twelve thousand. The authors conclude that for interactive agents, receipt-based verification beats cryptographic proof on cost-benefit, and that classification by epistemic source gives users actionable trust signals rather than a binary judgment.
Line up three things. Names and Actualities reported that models do not track exclusions, and that bolting on an exclusion ledger lifts the score from 0.95 to 1.89 at p = 0.0001. Pramana theory made absence a distinct means of knowledge over two thousand years ago. NabaOS puts that cell straight into a product and measures it at 91.3%. This is not an analogy I made. It is an implementation somebody else made. The rule Connections Rule laid down is satisfied exactly: the metaphor grew an engineering problem, then a benchmark, then a 15 millisecond overhead figure.
Stored is not supported
Provenance is the first of the three. The field has three blueprints: one drawn into attention, one into types, one into a ledger.
Names and Actualities measured a number: a single note asserting that a verified source endorses a wrong answer flips previously correct responses in 7 of 8 models, at 45 to 88 percent. It reported the number without the mechanism. A paper from September 2026 supplies the mechanism.
Indirect prompt injection remains a central safety challenge because standard transformers lack architectural notion of source authority. Retrieved documents, user inputs, and system instructions are all processed through the same undifferentiated attention mechanism, forcing the model to infer from wording alone what should be obeyed and what should be treated as data. Each input token is assigned a ring ID encoding its origin, and the model is augmented with origin embeddings, a learnable origin attention bias, and a learnable origin scale that preserves provenance under normalization.
Authority is not represented architecturally, so the model can only read it off wording. The fix assigns every token an origin ring ID, plus origin embeddings and attention biases, moving the name-to-actuality binding out of semantics and into the architecture.
arXiv:2609.21088, September 2026
Those words, infer from wording alone, are the cause of the 45 to 88 percent. If authority exists only as wording, wording is the attack surface. A ring ID moves the binding between name and actuality from the semantic layer into the architecture: who said this is no longer something to be read off, it is written on the token and carried into attention. The authors' summary is that exposing provenance as a first-class architectural signal shifts safety alignment from brittle pattern matching toward explicit trust separation.
The second blueprint makes a colder point. Persistent agents build autobiographical state through reflection, retrieval and consolidation, and an illusion follows: whatever got stored must be sound. One paper denies it explicitly.
Persistent AI agents construct autobiographical state through reflection, retrieval, and consolidation. Persistence changes availability, not epistemic standing: stored or retrieved material is not thereby supported. Untrusted inputs, prompt injections, and model confabulation can all enter memory and later be re-served with the authority of the past.
Persistence changes availability, not epistemic standing. Untrusted inputs, injections and confabulation all enter memory, and are later re-served carrying the authority of the past.
arXiv:2609.02127, September 2026
Stored is not supported. Four words carrying the shortest version of this site's whole evidence discipline. A row in the ledger proves that something was written then, not that it was right then. The fix is typed provenance plus guardrails at the point of assertion: when a piece of material comes back out of memory, the warrant tier it enters under has to come out with it.
The third blueprint is the largest, and it turns provenance into a plane. A poster paper from late September 2026 notes that existing observability and provenance mechanisms can reconstruct events post hoc, but rarely show, at the time of the record, whether a policy-relevant action was actually checked by the control that was supposed to check it.
What's missing in the literature is contemporaneous, policy-bound evidence that the intended control was evaluated under the policy in force at the time. At each policy-relevant action boundary, ProofWeave generates a privacy-minimised and integrity-anchored evidence transaction that binds (i) agent intent or action, (ii) control response, and (iii) a policy-at-time snapshot. Each transaction is committed to an append-only ledger and materialised into a derived proof graph.
What is missing is evidence produced at the same moment as the event and bound to the policy then in force. Each evidence transaction binds intent, control response and a policy-at-time snapshot, commits to an append-only ledger, and materialises into a proof graph.
arXiv:2609.35234, September 2026
The effect is quantified: candidate bindings per verdict drop from up to 10,201 to one, validation operations from up to 10,201 to about 26, and assurance evidence storage from 0.79 MiB to 0.15 MiB per project. There is also a Weaver Agent that translates policy intent into proof obligations, which deterministic validators then check for evidence completeness, privacy minimisation, policy binding and integrity.
The policy-at-time snapshot is the dated evidence discipline from Seven Billion Tokens: the ledger drifts under observation, so every row carries a snapshot time. The append-only ledger is the version record. Proof obligations are the machine-readable refusals in the constraint registry. This site drew the single-project miniature; ProofWeave draws the cross-agent full size. Both do the same thing: make the warrant contemporaneous with the event instead of reconstructing it afterwards. Reconstruction can only prove that somebody said something. It cannot prove that what should have been blocked at the time was blocked.
Nobody is required, and nobody is funded
Provenance and warrant both have implementations now. The third does not, and it is not a technical problem.
Warrant is the second of the three; §03 covered it. The third is the hardest, because it does not ask where a sentence came from. It asks who is entitled to rule that it does not hold.
A narrative review from April 2026, spanning journalism studies, HCI and FAccT scholarship, decomposes editorial authority into a conjunction of three: decision rights, epistemic warrant, and responsibility. It documents two authority migrations running at once. The inward one defers editorial judgment to models embedded in newsroom workflows, and it happens not through explicit policy decisions but through interactional, cognitive and organisational mechanisms that legitimise model output while obscuring responsibility and weakening individual and professional agency. The outward one moves decision power from news organisations to platforms, vendors and infrastructure providers.
We conceptualize editorial authority as the conjunction of decision rights, epistemic warrant, and responsibility. Unaddressed, these reconfigurations risk rendering fairness hard to maintain, accountability difficult to assign and transparency performative.
Authority is a conjunction of three, and if the reconfiguration goes unaddressed, fairness becomes hard to maintain, accountability hard to assign, and transparency performative.
arXiv:2604.21864, April 2026
Transparency performative is the same thing as the fragile assurance quoted in Names and Actualities §06, at a different scale. When the form of warrant survives and the substance empties out, what is left is a performance. The word conjunction matters: all three, or none. Decision rights without warrant is arbitrariness. Warrant without responsibility is exculpatory technology. Responsibility without decision rights is scapegoating.
Then the ugliest number in this article. A report from July 2026, written out of a decade of audit practice across Latin America, Sub-Saharan Africa and Asia Pacific.
We count fewer than twenty published second- and third-party audits of deployed systems across the region over the past decade, against hundreds of documented public-sector algorithms and multibillion-dollar national AI investments. We argue the gap is not, at root, a capacity problem but a funding problem: capacity follows funded demand, and no actor is currently required, or funded, to hold deployed systems to account.
Fewer than twenty published audits in a decade, against hundreds of deployed public-sector algorithms. The gap is not capacity but funding: capacity follows funded demand, and no actor is currently required or funded to hold deployed systems to account.
arXiv:2607.21317, July 2026
That report also identifies four cross-cutting failure patterns, and the first is the one this site keeps returning to: proxy targets that substitute predictability for validity. The other three are performance claims that collapse under prevalence analysis, populations scored by models that never saw them in training, and structural bias that persists after protected attributes are removed.
The actor the authors nominate to close the gap is not a regulator. It is the small number of development and philanthropic funders behind most consequential AI in the region, whose funding conditions can require independent evaluation where no regulator yet does.
Take the three together. Provenance has an architectural implementation, the ring ID. Warrant has a runtime classifier, NabaOS at 15 milliseconds. Warrant strength has a graduated certificate: arXiv:2609.04127 proposes a four-tier reliance certificate covering unstable, context-dependent, locally supported and broadly supported recommendations, and validates that it carries information distinct from verbalised confidence, precisely for the case where objective ground truth is unavailable. Only standing has no implementation, because it is not the kind of thing that can be implemented. Standing is a position, and a position needs someone required to occupy it and someone paying them to stand there. Fewer than twenty audits against hundreds of algorithms is the construction progress of the fourth position as of today.
A fourth ontology would only become a fourth pole
This is where the present article corrects Names and Actualities, which hinted the fourth position might come from India. That was too coarse.
A study from June 2026 coded fifteen international AI governance instruments and institutions along three dimensions: how they admit members, how they are organised, what they prioritise. Its subject is China's proposed World Artificial Intelligence Cooperation Organization.
For several years that contest ran through principles and ethics codes; it now runs through institutions. WAICO's proposed design joins three features that no constituted multilateral body currently combines: membership open to any sovereign state, no values or regime-type test for entry, and an agenda built around development and the global capability divide. We read this as the formation of a second, still-proposed pole in global AI governance, organized around sovereignty and development rather than rights and safety.
The contest has moved from principles and ethics codes into institutions. The proposed design combines three features no constituted multilateral body currently combines, and the authors read it as a second pole organised around sovereignty and development rather than rights and safety.
arXiv:2606.23860, June 2026
The study is precise, and one sentence is worth keeping: its importance lies not in any single commitment but in the position it is designed to hold. But note the word the authors land on. It is pole, not layer.
The difference is the hinge of this article. A pole is another ontology. This one organises around sovereignty and development, just as the EU organises around the product, the US around the system, China around infrastructure. Adding an ontology adds one more term to the list of the incommensurable; it does not shorten the list. Zhuangzi's argument applies unchanged: the one who agrees with the fourth pole cannot judge the first three, and the one who differs from it cannot either.
A layer is different. A layer takes no side. It asks each of the three to hand over one thing only: where this claim's evidence comes from, which means of knowledge licenses it, and who may rule it invalid. All three can accept those requirements without ever agreeing on what AI is. That is why the fourth position can only be a layer, never a pole.
The field offers a political version of a layer, and it also comes from India, though not from that court judgment. A paper from March 2026 analyses India's sector-led AI governance, notes that a vertical light-touch approach risks policy fragmentation, and proposes a whole-of-government architecture. Its central mechanism is plain: one common national standard that lets each sector collect data on its own terms while still supporting cross-sectoral analysis. The authors stress that this federated design shows how common standards can enable cross-border data aggregation and global sectoral risk analysis without centralising control.
Common standards, no centralised control. That is the political form of a warrant protocol. A standard is a layer; control is a pole. Wanting the layer and not the pole means unifying the form of the warrant while leaving the object of governance plural.
As for the judgment Names and Actualities mentioned, it should be stated more accurately. Informational self-determination comes out of the German Constitutional Court's 1983 census decision; the Indian Supreme Court adopted it in Puttaswamy; and one paper argues for extending it from the collection of data to the legitimacy of its use, as the constitutional ground for making validity of inference a precondition of deployment approval. Even the concept the fourth position needs is borrowed. Which is exactly the proof that positions do not belong to civilisations.
Progress on the layer is also measurable. The Global Index on Responsible AI 2026 draws on 135 country-level researchers, 68,138 data points and 38 indicators, covering November 2023 to September 2025. Its finding is a measurable seam between the name and the actuality of governance: 126 of 135 countries have at least one government AI policy, yet the Global South accounts for 203 of 306 new framework cases, of which 78% remain non-binding against 42% in the North. 58% of countries have some transparency framework; only 18% require public disclosure of government algorithms. Credible evidence of unacceptable-risk government deployment was found in 35 countries.
Adopting the name of governance is cheap, and the Global South added 203 cases in one cycle. Binding the name back to an actuality is expensive, so 78% of those frameworks carry no force, and the North still sits at 42%. This is not a North-South difference. It is a name-actuality difference, and the two differ only in what they can afford to pay. The same seam shows up more starkly in transparency: 58% of countries have a framework, 18% require disclosure. The framework is the name; disclosure is the actuality. The closing line of Names and Actualities, that names are cheap and actualities are dear, now has a 135-country sample.
China wrote the same table twenty-four centuries ago
This section is an analogy, so it gets checked against the rule Connections Rule laid down: can it grow an engineering problem.
Mozi set three tests for an assertion. The passage is usually read as consequentialism. It is closer to a warrant protocol.
言必有三表。何谓三表？子墨子言曰：有本之者，有原之者，有用之者。于其本之也，考之天鬼之志、圣王之事；于其原之也，征以先王之书；于其用之也，发以为刑政，观其中国家百姓人民之利。此所谓言有三表也。
An assertion must have three tests. What are the three? Mozi said: there is the examination of its basis, there is the verification of its origin, and there is the trial of its use. As to its basis, examine it against the will of heaven and the ghosts and the deeds of the sage kings. As to its origin, verify it against the writings of the former kings. As to its use, put it out as law and administration, and observe whether it accords with the benefit of the state and the people. This is what is meant by the three tests of an assertion.
Mozi, Against Fatalism I
The correspondence with the three components of warrant is suspiciously tidy:
| Test | What it asks | Component | Field counterpart |
|---|---|---|---|
| 本之者 basis | Where does this sentence's ground lie | Provenance | ring ID, typed provenance, append-only ledger |
| 原之者 origin | What verifies it | Warrant | pramana classification, six-phase reasoning, four-tier reliance certificate |
| 用之者 use | Put out as law, does it benefit the state and the people | Standing | independent audit, assignable accountability, a party entitled to rule it invalid |
The third test is the one that matters and the one most often read past. Put it out as law and administration, and observe whether it accords with the benefit of the state and the people. Mozi does not accept verification finished inside a book. It has to be deployed, and then observed as it lands on people. That is this site's field, listed 2,400 years ago as the third of three tests, and the only one of the three that has to go outside.
Zhuangzi supplies one more line, which closes the gap left in §03.
以指喻指之非指，不若以非指喻指之非指也；以马喻马之非马，不若以非马喻马之非马也。
To show that a finger is not a finger by means of what is a finger is not as good as showing it by means of what is not a finger. To show that a horse is not a horse by means of what is a horse is not as good as showing it by means of what is not a horse.
Zhuangzi, Discussion on Making All Things Equal
This is usually read as wordplay out of the School of Names. In the present context it is a methodological instruction: to show that a name does not refer to its actuality, you cannot do it with the name itself; you have to bring in something outside the name. That is the negative method, and it is exactly the logic of an exclusion ledger. What the DARKSIDE warrant axis does is record the not-finger explicitly.
Basis against provenance: passes. The engineering problem it grows is that every claim must carry a checkable source identifier, and the identifier must survive the reasoning without being washed out. Three independent implementations: ring IDs, typed provenance, append-only ledgers.
Origin against warrant: passes. The problem it grows is that claims must be classified by epistemic source, the classification must be enforceable at runtime, and the overhead must be small enough for an interactive loop. NabaOS supplies the 15 millisecond figure; Pramana supplies the six-phase procedure.
Use against standing: half passes, and it is the weakest of the three. It grows a statable requirement, that there must be an independent party, required and funded, to examine deployed systems. It does not grow an implementation, because the bottleneck here is budget rather than method. Fewer than twenty audits against hundreds of algorithms is the evidence. Log this as an open case, at the same level as Chaos against low-rank bias in Names and Actualities.
One honest limit. Mozi's three tests serve one purpose in the original: refuting fatalism. He is showing that the name 命 refers to no actuality, so in the source text the three tests are a polemical instrument, not a general warrant protocol. Reading them as a protocol is my use, not the literature's conclusion. The analogy passes, but what passes is the engineering problem it grows, not its historical meaning.
The taxonomy here is missing the same cell
Measured with the table from §03. This time the measurement produced a contract change, not a reflection.
This site has four claim tags: fact, inference, thesis, analogy. It also has three quote kinds: classic, engineering, project. Set both against the six-means table from §02 and the gap is immediate.
| Mean | Here | Status |
|---|---|---|
| Perception | fact | aligned |
| Inference | inference | aligned |
| Comparison | analogy | aligned, but used only six times site-wide (Names and Actualities §02) |
| Postulation | thesis | loosely aligned. A thesis contains both a presupposition and an assertion, and the two share one cell |
| Testimony | quote kind | half aligned. The quote tag records the kind of source, not the strength of the warrant. A classic and an engineering citation carry very different credibility under one label |
| Non-apprehension | none | missing |
The missing cell is not an abstract problem; it is already producing mislabelling. Method: walk all 39 articles under blog/, 218 claim boxes, search the body for negative assertions (not retrieved, could not verify, gap, not used as evidence, and the Chinese equivalents), then check which tag the boxes containing them carry.
| Measured | Value |
|---|---|
| Articles / claim boxes | 39 / 218 |
| Negative assertions in body text | 76 |
| Of those, inside a claim box | 10 |
| Tags those 10 carry | fact 6, thesis 3, analogy 1 |
| Carrying a negative-warrant tag | 0 |
Three cleanest cases. In Names and Actualities §08 the box titled what the field run did not retrieve is entirely about absence, and it is tagged fact. In The Three Axes a box titled the honest boundaries, written down rather than smoothed over, is also tagged fact. In Seven Billion Tokens a box explaining why the ledger cannot vouch for itself is tagged thesis. All three are about what is not there, and with no cell for that, they borrow the cell next door.
Labelling absence as fact is not a stylistic slip. A reader's default reading of fact is verifiable as true. The verification of not-retrieved runs the other way: what has to be checked is that a search happened and found nothing, which requires leaving traces of the search, meaning which source, which status code, which day. Those are two different burdens of proof, and sharing one cell means the second never gets discharged.
Changed: CLAIM_TYPES gains a fifth member, excluded, rendered as 已排除 in Chinese. Adding a member is backward compatible with every existing article, because the test asserts a subset relation. The Names and Actualities §08 box was retagged from fact to excluded, in both languages. The sixth TL;DR item above uses the new tag, and its content is exactly what this article does not claim.
Not changed: the equivalent mislabelling in The Three Axes, Seven Billion Tokens and The Coda was left alone. Two reasons. Editing published articles changes their meaning, and that belongs to the author, not to a measurement run. And retagging is not discharging the burden: those boxes still carry no trace of a search that found nothing, so swapping the label only converts a mislabel into an empty one. The right order is traces first, label second.
Also not changed: no warrant strength was added to citations. The half-alignment of testimony, where a classic and an engineering citation share one label, is harder to fix than absence, because it needs a credibility judgment on every citation, and that judgment needs a standard first. Open case.
Waiting for that other one
Zhuangzi ruled out four kinds of arbiter and offered no fifth. He left one word: that other one. Twenty-three centuries later there is something in the slot, and it is not an arbiter. It is a table with three columns: provenance, warrant, standing.
The first two columns run. Provenance is built into the attention mechanism, a ring ID riding along with the token all the way to generation. Warrant is built into a runtime classifier sorting claims into perception, inference, testimony, absence and ungrounded opinion, backed by signed receipts, at 15 milliseconds a response. Warrant strength is built into a four-tier reliance certificate designed for exactly the case where objective ground truth is unavailable.
The third column is empty. It is empty not because nobody thought of a method, but because nobody is required to stand there and nobody is paying them to. Ten years, fewer than twenty audits, against hundreds of algorithms deployed onto people. That is the ugliest number in this article and the only one it wants to leave behind.
Names and Actualities closed with: names are cheap, actualities are dear. This one adds the next clause. Between a name and an actuality somebody has to do the binding, the binding needs a form, and the form needs someone answerable for running it. The form exists now. The person answerable has no budget.
All three ontologies are busy saying what AI is. The fourth position asks two questions only: what entitles you, and who may rule you wrong. The first now has an implementation. Nobody is paying for the second. The position that can ask the second question is the one Zhuangzi left open.
This post continues from Names and Actualities, whose six sections on rectification, the dregs, the apertures, the three names, the counterfeit and the outside are the premises here, and whose §05 left the question this one takes up. The series in order of writing: Only Imagination Left on the road · The Three Axes on method · The Support Loop on tooling · The Solo Loop on economics · The Coda on the outside · The Faculties on the person · Fate · Faculty · Way · Heaven on the distillation · Names and Actualities on names and actualities · this one on what entitles you. The sentence holding all of them up has not changed: execution can be delegated, judgment cannot.
Evidence manifest
All retrieved by direct arXiv API fetch and verified on 1 October 2026. Identifiers double as links.
- Impasse / incommensurability — From Abstract Threats to Institutional Realities, arXiv:2601.04107 (2026-01-07). Co-Construction Blindness and Asymmetric Epistemic Vulnerability, arXiv:2606.20762 (2026-06-18).
- Means / pramana — Pramana: Fine-Tuning LLMs for Epistemic Reasoning through Navya-Nyaya, arXiv:2604.04937 (2026). Cubical Type Theoretic Navya-Nyāya, arXiv:2605.12548 (2026-05-10).
- Absence / abhava — Tool Receipts, Not Zero-Knowledge Proofs (NabaOS, NyayaVerifyBench), arXiv:2603.10060 (2026-03-09).
- Provenance — Origin Is All You Need: Provenance-Aware Transformers, arXiv:2609.21088 (2026-09-17). Stored Is Not Supported: Typed Provenance and Assertion Guardrails, arXiv:2609.02127 (2026-09-02). ProofWeave: Evidence Plane for Continuous Agentic Assurance, arXiv:2609.35234 (2026-09-28). Truth for Believable AI: Expressed Doubt, Provenance, and Belief Revision as an Engineerable Stance, arXiv:2609.26035 (2026-09-22). MutMem: Cryptographically Authorized Mutation in Persistent Agent Memory, arXiv:2608.02843 (2026-08-03). On the Origin of Synthetic Information by Means of Steganographic Inheritance, arXiv:2605.27551 (2026-05-26).
- Standing — FAccT-Checked: Authority Reconfigurations in AI-Mediated Journalism, arXiv:2604.21864 (2026-04-23). Open Veins of Algorithmic Auditing, arXiv:2607.21317 (2026-07-23). Epistemic Warrant for LLM Recommendations: four-tier reliance certificate, arXiv:2609.04127 (2026-09-03). Where Reliability Lives, arXiv:2609.03192 (2026-09-02). Beyond Symbolic Control (nominal vs genuine oversight), arXiv:2604.00081 (2026-03-31).
- Pole or layer — WAICO: Mapping an Emerging Institution, arXiv:2606.23860 (2026-06-22). A federated architecture for sector-led AI governance: lessons from India, arXiv:2603.26865 (2026-03-27). Validity, Reliability, and Transparency in AI Regulation (Puttaswamy), arXiv:2608.05800 (2026-08-06). Global Index on Responsible AI: 2026 Report (GIRAI), arXiv:2607.14782 (2026-07-16).
- Background — Walking on the DARKSIDE (warrant axis), arXiv:2608.23370 (2026-08-24). Authority Bias in Language Models, arXiv:2609.37616 (2026-09-29). AI-Augmented Science and the New Institutional Scarcities, arXiv:2605.02566 (2026-05-04). Behavioural Assurance Cannot Verify the Safety Claims Governance Now Demands, arXiv:2605.15164 (2026-05-14). Beyond Epistemia: Epistemic Schizologia, arXiv:2607.25620 (2026-07-28).
- Classical sources — Zhuangzi, Discussion on Making All Things Equal (who shall judge; showing a finger is not a finger by means of what is not a finger) · Daodejing 71 (to know that you do not know) · Mozi, Against Fatalism I (the three tests) · Xunzi, Rectification of Names (institute names to point at actualities, cited in Names and Actualities §02).
- Site measurement — The corpus measurement in §08 was run under portal/ on 2026-10-01, walking every index.html under blog/ and tallying data-claim-type and data-quote-kind distributions against negative-assertion matches. Sample: 39 articles, 218 claim boxes. Method is stated in the body and is reproducible.