Two rivals and nobody to referee them

Dr Kavya Iyer studies attention in a small lab in Pune. She holds a workspace view of access: a piece of information becomes available to a mind when it is broadcast to many specialised processes at once. Dr Jonah Brandt, a philosopher of neuroscience in Bristol, thinks this view mistakes a bulletin board for a mind. For him, what matters is the causal structure that does the broadcasting, and a broadcast alone tells you very little.

They have argued for three years, mostly in replies to each other's preprints. This spring the argument found an object. An open-source software agent called Wren has a scratchpad, a shared memory that its planning, search and arithmetic modules all read from and write to. Kavya says the scratchpad is a workspace in her functional sense. Jonah says it is a shared file and nothing more.

Neither of them claims that Wren is conscious, and neither will this paper. Their dispute is narrower and more useful. It is about access, the second layer of the Recognition Theory of Mind (RTM): whether information in the scratchpad becomes available to the whole system in the way a workspace theory says it should. Kavya thinks a test could settle it. Jonah agrees, in principle. Then they try to design one, and they discover that they disagree about almost everything a test needs. The difficulty is familiar to anyone who has watched a scientific controversy closely. Rivals frequently agree that evidence should decide, and then discover that they cannot agree which evidence, gathered how, would count.

They disagree about what "access" means, and about which tasks count as needing integration. They disagree about whether removing the scratchpad is a clean intervention or a crude one that breaks everything. A journal editor they both respect declines to host the contest, since it is too speculative for her pages. No funder will pay for a dispute between two people about one piece of software. Nobody is going to referee them.

This paper asks what they should do instead. Its question is the one the series set for it: what public protocol can fairly adjudicate predictions when rival investigators disagree about both methods and definitions? My answer is a protocol in six steps, which I call the Open Vāda Protocol after the classical Indian name for debate aimed at truth. It works without a journal, a funder or a referee. What it requires is a set of commitments that are public, sealed and symmetric, and a way of checking, before any data exist, whether the test is fair.

The paper builds on the earlier method papers in this series. The charter set out how to compare minds across kinds (Paper 1). The second paper argued that a measure can define a term for a purpose without becoming its meaning. The third showed how an index earns trust when nobody holds the answer key. The fourth set out a ledger for claims that cannot be quietly edited. This one asks what happens when two ledgers meet.

Why rivals talk past each other

Disagreement between theories is the normal state of consciousness science. The surprise is how rarely it is settled. Rival camps publish in different journals, cite different results, and train students who inherit the camp along with the methods. Each side has data, but the data rarely touch the other side's theory where it hurts.

Some of this is ordinary sociology, but much of it has a structure, and the structure can be named. Kavya and Jonah are a useful pair because they show all of it in miniature.

Four kinds of disagreement

When Kavya and Jonah write down every point on which they disagree, they find fourteen. Sorting them shows that the points are not all the same kind. Some are about facts: what would happen to Wren's accuracy if the scratchpad were removed. Some are about methods: whether removal is a fair intervention, or whether a gentler one, such as adding noise to the scratchpad, is better. Some are about definitions: what "access" is, and whether it comes in degrees. And some are about values: whether a theory that makes fewer predictions should be preferred for its caution or criticised for its timidity.

These four kinds need different treatment. A factual disagreement can be settled by an observation, if both sides agree in advance what the observation would mean. A methods disagreement can sometimes be settled by running both methods and seeing whether they agree. A definitional disagreement usually cannot be settled at all, and trying to settle it first is one of the commonest ways a contest goes wrong. No test can settle a disagreement about values. It is useful to know which is which before anyone collects data.

The trap of agreeing definitions first

The natural first move in any dispute is to agree on terms. It sounds like good manners and good method. Kavya proposes it in her first email: let us agree what "access" means, then test it.

Most people's first answer is that the test was fair, since both sides signed the definition. Look again at what the definition contains. "Available for use by several modules" is close to the workspace theory's own statement of what access is. Once Jonah signed it, he had conceded the point in dispute, and the test merely confirmed a concession. He lost in the second week, in the drafting, without noticing. The data did no work.

This is the trap. A definition that both sides accept tends to be one side's theory with the edges sanded off, or a phrase so vague that each side reads its own theory into it. In the first case the test is decided before it starts. In the second it cannot be decided at all. The fight about the definition was the real dispute, and agreeing the definition hid it.

An analogy from translation

A better picture is a bilingual dictionary, since Kavya speaks one theoretical language and Jonah another. Neither has to give up their own words. What they need is a table that says, for each of their terms, which observable both of them would accept as evidence about it.

This is what I mean by definition escrow. Each side writes its own definitions in full and seals them, and neither has to accept the other's. Then, together, they build a translation table: for each term that will appear in a prediction, the observable that both agree bears on it. Terms that cannot be mapped are listed as unmapped. Nothing that depends on them can be used to decide the contest.

The goalkeeper who carries the goal

The goalkeeper is funny because he is honest. He has found a gap in the rules and uses it in plain sight. Most scientific goalpost-moving is less visible and less deliberate. A prediction made loosely in a talk is tightened after the results, so that the result fits. A theory that seemed to forbid an outcome turns out, on reflection, to have allowed it all along. Nobody lies. The goal simply moves when the ball arrives.

The cure for the goalkeeper is to fix the goal to the ground before kick-off. In science that means writing down, in advance, what each possible outcome would mean, and making the record impossible to edit afterwards. The fourth paper in this series showed how a hash chain makes a record uneditable. This paper uses the same tool for a different purpose: to fix both goals before the match.

What a stalemate costs

A dispute that never resolves has costs beyond the rivals. Students choose a camp before they can evaluate the evidence. Reviewers from one camp judge papers from the other by standards the authors never accepted. Funding follows reputation, and reputation follows publication counts within each camp, so both theories can grow more confident at once while neither faces a decisive test. Outsiders, including journalists and the engineers who build agents like Wren, learn that the science is a matter of allegiance.

Kavya and Jonah feel a small version of this. Kavya's students cite her papers and treat Jonah's objections as settled. Jonah's readers treat workspace results as irrelevant by definition. Each side has a growing literature that the other does not read. A contest is a way of forcing the two literatures to touch. It cannot force agreement. It can force a shared record of what happened when both sides' predictions met the same data.

What institutions were doing

Gatekeepers have a reputation for slowing things down, and sometimes they do. But in a scientific dispute they also do three useful jobs. A journal or funder fixes the plan before the data, by accepting a registered protocol. It holds the record, so that neither side can change it later. And it lends standing, so that the losing side cannot simply walk away and say the contest did not count.

Take the institution away and all three jobs remain. Kavya and Jonah need some way of fixing the plan, holding the record and making the result count. The protocol in the next part does each of these jobs with a public mechanism in place of a trusted office. Clark and Tetlock have argued that adversarial collaboration is the next reform science needs, because open-science practice still leaves researchers room to steer their own conclusions (Clark and Tetlock, 2023). Their argument assumes collaborators willing to sit at one table. The protocol cannot do everything an institution does. It can do enough for a contest between two people who want the truth more than they want to win.

The Open Vāda Protocol

The protocol has six steps. Each step replaces one job that an institution would otherwise do, or blocks one way a contest can fail. I describe them in order, using Kavya and Jonah's test as the worked example throughout.

Step one: put your definitions in escrow

Each side writes its definitions of every term it will use, in its own words, with as much theory as it likes. Kavya writes two pages on access as broadcast. Jonah writes three on access as a property of causal structure, including a definition of a term, intrinsic causal power, that he admits has no behavioural test. Each document is fingerprinted with a hash, and the two fingerprints are posted publicly with the date. The documents themselves can be published at once or held back until the end. What matters is that neither can be changed without the change showing.

In practice the sealing is simple. Each side computes a SHA-256 fingerprint of its document and posts it in at least two public places that neither side controls, such as an archived web page and a public code repository, with the date. Anyone who later receives the document can recompute the fingerprint and compare. The fourth paper in this series described the same mechanism for a single researcher's ledger; here it serves two researchers who do not trust each other's records.

Why seal definitions the other side need not accept? Because later, when the results arrive, each side will be tempted to say that the result was about a different sense of "access" from the one they meant. The sealed definition answers that temptation, because it records what each side meant before they knew what would happen.

An analogy from buying a house

In many countries, when a house is sold, the buyer's money and the seller's deeds go to a third party, who holds both until the conditions of the sale are met. Neither side has to trust the other, because both trust the escrow agent and the written conditions.

Step two: build the translation table

With the definitions sealed, the two sides build one shared document: the translation table. For each term that will appear in a prediction, it names an observable that both sides agree bears on that term. It leaves the meaning of the term alone and records only what both would count as evidence.

For Kavya and Jonah the table has six rows. "Integration task" becomes a task in a fixed set of forty, each of which needs a result from one module to be used by another within the same plan. "Single-module task" becomes a task in a second fixed set of forty that one module can complete alone. "Ablation" becomes a specific code change: the scratchpad is replaced by an empty buffer that accepts writes and returns nothing. "Accuracy" is the share of tasks completed correctly, scored by a fixed script. Two terms, "access" itself and "intrinsic causal power", stay unmapped.

Leaving "access" unmapped looks strange, since the dispute is about access. But this is precisely the purpose of the step. Neither side can agree what access is, so neither may use the word in a prediction. They may use only the observables in the table. Whether those observables tell you about access is left for each side to say in its own voice, at the end, in public. The data will decide something narrower, about tasks, ablation and accuracy, and both sides have agreed in advance that this narrower thing matters to them.

Step three: map the disagreement

The fourteen points of disagreement are now sorted into the four kinds. Only factual disagreements can be decided by the test. Methods disagreements are handled by building the alternative methods into the design, so that the test reports both. Definitional disagreements stay in escrow. Disagreements about values are written down and set aside.

For Kavya and Jonah, five points turn out to be factual and reachable with the translation table. Three are about methods. The largest is whether ablation is too crude, and they agree to run a gentler intervention alongside it, in which the scratchpad's contents are scrambled while its size and timing stay the same. Four are definitional and two are about values. The map is published with the rest.

The map is useful even if the test is never run. It shows outsiders which parts of a famous dispute could ever be settled by data, and which parts are arguments about words or priorities. In my experience of reading such disputes, the definitional part is usually the largest and the least acknowledged.

[AUTHOR NOTE: [AUTHOR NOTE: if you have taken part in a public dispute or exchange of replies with another researcher or teacher, describe briefly how much of it turned out to be about definitions. If you have not, delete this note.]]

Step four: seal an outcome table

This is the heart of the protocol. Each side, separately, fills in an outcome table. The rows are the possible outcomes of the test, defined in the translation table's terms. For each row, the side writes what it would conclude: that its theory wins, that it loses, or that it needs revision. It also writes a probability for each outcome, so that the table can be scored later in the way the fourth paper described. Then the table is sealed with a hash and the fingerprint is published.

Kavya and Jonah agree on three outcomes. In outcome W, the workspace pattern, ablation lowers accuracy on integration tasks by at least twenty points and on single-module tasks by no more than five. In outcome U, the uniform pattern, both kinds of task drop by amounts within five points of each other. Outcome N covers everything else.

Kavya's table says she wins on W and loses on U and N. She gives W a probability of 0.70. Jonah's table says he loses on W, wins on U, and needs to revise on N. He gives U a probability of 0.55. Both completed tables are sealed before a single task is run.

Two conditions make the test fair, and both can be checked from the sealed tables alone. Some outcome must separate the sides: there must be at least one row where their verdicts differ. And each side must have an outcome on which it loses. If a side has no losing row, it has entered a contest it cannot lose. The checker in the instrument section looks for both.

Why each side gives numbers

The probabilities in each table do three jobs. They let each side's predictions be scored later, in the way the fourth paper described, so that a theory that gave its winning row 0.95 and lost is penalised more than one that gave it 0.55. They let anyone compute, before the test, how much the test is expected to move belief between the two theories. And they expose a side that has given its losing row a token probability, a tactic I return to in the case against the protocol.

The numbers are the sides' own. Neither has to accept the other's probabilities, any more than its definitions. The protocol asks only that each side commit to its numbers before the data, in writing, under seal.

The side that cannot lose

The classical Nyāya texts name three kinds of debate. Vāda is discussion aimed at truth, in which both sides argue from sources of knowledge and reasoning. Jalpa is debate aimed at victory, which allows quibbles and trick replies. Vitaṇḍā is a form of jalpa in which one side attacks without ever putting forward a position of its own. The Nyāya-sūtra defines it in five words: jalpa "lacking the establishment of a counter-position" (1.2.3).

A side with no losing row in its outcome table is the modern vitaṇḍā debater. It can criticise every result and concede none. In a journal, this behaviour is limited by referees and editors. In a protocol without gatekeepers, it has to be limited by a rule. The rule is simple: a side that registers no losing outcome may still take part and say what it likes, but the contest is marked one-sided, and nothing it says about the result is scored.

Scepticism is still welcome, and a sceptic may argue that a test is flawed. The rule asks only that the sceptic say, before the data, what result would make them stop.

Two agents with the same curve

The test cannot tell them apart, because the outcome table was written in terms of task accuracy. Same behaviour, different mechanism: a perfect mimic produces the same row. The protocol has not failed here. The translation table has made one of its limits visible. Suppose Kavya wants her result to be about the scratchpad as a mechanism. Then the table needs an observable that would differ between the two agents. One candidate is to write a single false value into the shared buffer and trace which modules use it.

Kavya and Jonah add this as a fourth outcome condition before sealing, which costs them another week. It is the most useful week of the design, since it is the one in which each side learned what the other would count as a mechanism.

Step five: let code decide which outcome happened

The analysis is written as code before any data exist. The code reads the task results, computes the accuracy drops, and prints which row of the outcome table occurred. The code is sealed together with the outcome tables, and when the data arrive, anyone can run it.

In Kavya and Jonah's test, the runs are done by a volunteer analyst, Achieng Otieno, a doctoral student in Nairobi who has no stake in either theory. She receives the agent, the task sets and the sealed code. She does not see the outcome tables until the code has printed its row. Her role is to run things correctly and report what happened. She is a witness, and nothing turns on her judgement.

An analogy from cricket

International cricket uses a decision review system. A player who thinks the umpire is wrong can ask for a review, and the third umpire checks the replay, the sound of the ball on the bat and a computer model of the ball's path. The rules for what counts as out are fixed before the series begins. The technology decides, within those rules.

Step six: write three voices and enter both ledgers

When the code has printed its row, the result is written up in three parts. The first is the data statement: what the code printed, with the numbers, and nothing else. The second and third are each side's reading, written by that side and published unchanged. This three-voice format is borrowed from the largest adversarial collaboration in consciousness science so far, which I describe in the next part.

Finally, both sides enter the result in their claim ledgers. Kavya's prediction resolves as true or false, and so does Jonah's, with the probabilities they gave. The sealed tables make this mechanical. A side that registered "lose" on the row that happened records a loss. It may add a revision entry explaining what it now thinks, but it may not edit the earlier one.

The six steps together do the three jobs a gatekeeper would have done. Sealing fixes the plan before the data. Public hashes hold the record. And the outcome tables give the result its standing, because each side said in advance, in writing, what the result would mean to them. The losing side has already agreed that it lost.

What real contests have shown

The protocol is new, but its parts are not. Researchers have run adversarial collaborations, made public bets and held structured debates for a long time. Their record shows what commitments can and cannot do. I take five episodes in turn, and in each I ask the same question Kavya and Jonah will face: what did the contest settle, and what did it leave open?

The collaboration that named the method

The result settled one thing and left another open. Frequency formats alone did not remove the conjunction effect with any of the phrasings. But when the filler items around the key question were removed, the effect disappeared with Hertwig's phrasings. The paper reports that the two rivals then offered different readings of the same findings.

Two lessons carry straight to Kavya and Jonah. First, the condition on which both rivals agreed in advance, "who are", could not separate them. A row on which both sides give the same verdict is useless for deciding between them, however carefully it is run. Second, a methods variable that neither side had written into the predictions, the presence of filler items, decided how the result looked. The outcome table was silent on it, so the result could be read both ways. Every variable that could change the row belongs in the table.

A failure to disagree

This collaboration ended with a map, and no test that either side lost. The two traditions turned out to be describing different environments, and the disagreement shrank once that was written down. The disagreement map in step three is meant to produce this kind of result when it is available. Sometimes rivals discover that they were never disagreeing about the same thing. That discovery is worth publishing too.

A bet that was settled by a cheque

As a contest, the bet was close to ideal. The outcome was defined in advance in terms both accepted: prices and quantities. Each side had a row on which it lost. Resolution needed no judgement, only published prices; no institution was involved, and the result was paid.

But the bet did not settle the dispute about scarcity. Ehrlich and others argued that metal prices over one decade were a poor measure of the question they cared about. In the protocol's terms, the translation table linked "scarcity" to an observable that only one side thought was the right one. The bet settled what it covered. The larger argument was definitional, and it stayed open.

A bet that was settled by agreement

This was a friendly bet and a good one, and it was also, by the protocol's standards, unfinished. The whole outcome depended on one word, "clear", which neither side had mapped to an observable. Resolution therefore needed the two of them to agree, in 2023, on what "clear" had meant in 1998. They did agree, which says something good about both. A contest between less friendly rivals could have stalled at exactly this point.

The largest test so far

The results were first presented at a public event in 2023, before peer review. Soon afterwards, a letter signed by 124 researchers argued that integrated information theory should be treated as pseudoscience until the theory as a whole is empirically testable (Fleming and colleagues, 2023). The letter said the experiments tested particular predictions that were not logically tied to the theory's core.

For this paper, the dispute about the letter matters less than what it shows about outcome tables.

The collaboration did what the protocol asks. It sealed predictions and their interpretation in advance, separated proponents from analysis, and reported three voices: the consortium's reading first, then each side's. The theory leaders had also agreed, before data collection, that combining intracranial recordings, MEG and fMRI would provide the most conclusive test available. That advance agreement on method is exactly what step three of the protocol asks for. Even so, a large part of the field argued afterwards that the predictions tested were the wrong ones. Their complaint was about the translation table, about which observables were allowed to stand for the theory. A later paper in this series takes up the COGITATE results in detail. Here I take one lesson: a fair test of the wrong rows is still a fair test of the wrong rows.

Many readings of one result

The three-voice format exists because readings differ even when the data do not. A Bayesian treatment of adversarial collaboration makes this explicit. Each theory's predictions are written as a model, and the data update the relative support for each model. Evidence then accumulates across studies without any single study needing to be decisive (Corcoran, Hohwy and Friston, 2023). On this view a contest moves weight between theories, and nobody expects a knockout.

The protocol borrows this when it asks each side to give probabilities in its outcome table. The checker in the instrument section uses them to report how much evidence the test could be expected to provide before it is run, and how much it did provide afterwards. A test whose two tables give almost the same probabilities to every row will move almost no weight, whatever happens.

Which rows are worth running

Not every discriminating row is worth the effort of a contest. A row matters to the extent that the two sides disagree sharply about how probable it is. If Kavya gives outcome W a probability of 0.70 and Jonah gives it 0.65, the row separates their verdicts and yet barely separates their expectations. Observing it would move almost no weight. If Jonah gives it 0.25, the same observation becomes informative, because one side expected it and the other did not.

This suggests a practical order of work. Before investing in data collection, the two sides compare their sealed probabilities and compute the expected evidence for each candidate design. Designs with negligible expected evidence are dropped, however elegant they are. The Bayesian framework for adversarial collaboration recommends something similar under the name of optimal experimental design, choosing the experiment expected to reduce uncertainty most between the competing models. The protocol's checker performs a simpler version of that calculation, with the advantage that anybody can rerun it.

There is a temptation here that the protocol should resist. Rivals who compare probabilities before sealing may each adjust theirs to look more confident, hoping to make the test look decisive in their favour. The remedy is the scoring rule from the fourth paper. Inflated confidence is penalised when the row resolves, so a side that exaggerates in order to win the design argument pays for it in its ledger.

What the record says

Across the five episodes, contests settled the questions they pre-specified and rarely the larger disputes around them. Simon won the bet and the argument about scarcity went on. The conjunction effect was tested and the rivals still read it differently. COGITATE challenged both theories and both have proponents. That record shows what contests are for. They make a narrow claim public and settled, so that the larger argument has one fewer thing to fight about.

There is a second, quieter lesson. In every case, the part of the contest that failed was the part that had not been written down: the filler items, the word "clear", the choice of which predictions stood for the theory. The protocol's six steps are an attempt to make everything that matters get written down before anyone knows the answer.

A third lesson concerns standing. In each episode, the result counted because the participants had made their commitments in public, in front of people whose opinion they valued. Ehrlich paid because the bet had been reported. Koch brought the wine to a conference hall. The COGITATE proponents had their predictions peer-reviewed before the data existed. Public commitment, more than institutional authority, is what gave each result its force, and public commitment is something two researchers can make without anyone's permission.

Disputes across the five layers

The Recognition Theory of Mind divides what a mind does into five layers. Some disputes about each layer can be given outcome tables now. Others cannot, and the protocol has to say so plainly.

Responsiveness and access

At the first layer, responsiveness, disputes are about whether a system changes its behaviour in response to a stimulus, and how. These are the easiest to put into a table. Both sides can agree on the stimulus, the measure and the threshold, and the outcome is a number.

Access, the second layer, is where Kavya and Jonah's dispute sits. Here the observables are uses of information: whether information in one part of a system is used by others, how quickly, and with what loss. These can be measured in agents and brains alike. The difficulty is definitional, as the running example showed. Both sides can measure the same thing and disagree about whether it is access. The translation table handles this by letting them measure without agreeing.

Reflexivity

The third layer, reflexivity, concerns a system's monitoring of its own states. Disputes here usually rest on reports and confidence ratings. These are outputs, and as earlier papers in this series argued, evidence at the level of outputs is evidence about access. It does not stand for evidence about reflexivity. An outcome table at this layer can still be written, but its rows will be about reports, and each side's reading will say what it thinks the reports show.

Kavya and Jonah briefly consider adding a reflexivity row: whether ablation also makes Wren's confidence ratings less accurate. They drop it. Both agree it would test a claim about Wren's outputs, and neither wants the result reported as a claim about Wren monitoring itself.

Self-recognition and the ground

At the fourth and fifth layers, self-recognition and the luminous ground, I know of no outcome table that two rival theories could both sign. Nobody has proposed an observable that both a defender and a critic of a claim about self-recognition would accept as deciding it. The fifth layer is, by construction, outside what the method can reach, and this series takes no position on whether any machine has it.

The protocol does not pretend otherwise. Disputes at these layers can go through steps one and three: each side can seal its definitions, and the disagreement can be mapped. They stop there. The map records them as escrowed. Escrowed is not lost and not won. It means that the contest has been written down for a day when a shared observable exists, if that day comes. Recording a dispute as escrowed also protects both sides from a familiar distortion, in which a success at a lower layer is reported as a victory at a higher one. If Kavya's result were described in a newspaper as showing that Wren is aware, the escrow record would show that neither rival claimed anything of the kind.

Where the running example sits

The Wren test is a second-layer contest, and both rivals have been careful to keep it there. Its rows are about task accuracy under intervention. Its readings are about functional broadcast and information flow. Neither rival claims that the outcome tells anyone whether Wren has experiences, monitors itself or recognises itself, and the paper does not claim it either.

This care has a cost. A result confined to the second layer will disappoint readers who wanted the contest to say something about machine consciousness. But the confinement is what makes the result usable. A second-layer result that both rivals accept is a small, firm piece of ground. A claim about higher layers, inferred from the same data, would be a fresh dispute with no outcome table behind it.

An old account of who is in the room

The classical Indian traditions thought hard about the conditions of debate, including who should be watching. The medical compendium of Caraka distinguishes kinds of assembly, the pariṣad, before which a debate takes place: learned and ignorant, and also friendly, neutral and prejudiced (Sasaki, on the Vādanyāya). The point of the distinction is practical. How you argue, and whether you should argue at all, depends on the room.

A protocol without gatekeepers has no assembly in that sense. Its room is everyone who later reads the sealed tables. That room is neither friendly nor neutral; it is simply large. The protocol's answer to a prejudiced room is to make the record checkable by any member of it.

Two conditions of defeat

Nyāya listed twenty-two conditions of defeat, the nigrahasthānas, under which a debater could be judged to have lost. The Buddhist logician Dharmakīrti, in his work on debate, the Vādanyāya, rejected the list and kept two. A proponent is defeated by asādhanāṅgavacana, failing to state what proves the thesis. An opponent is defeated by adoṣodbhāvana, failing to point out a real fault (Sasaki; Tillemans, SEP). Sasaki notes that on Dharmakīrti's view, neither side is defeated merely by tripping on one of the old technical rules.

The parallel with the two fairness conditions is close, though I do not want to push it far. Dharmakīrti asks each party to do its own job: one must prove, the other must find a real fault. The protocol asks each side to state a losing outcome and to accept a mechanical judgement of which outcome occurred. In both, defeat is tied to a party's own task, and trick rules count for nothing.

A thought experiment from practice

They can seal their definitions, and they can map their disagreement. When they try to write outcome rows, they find that every observable they can name, a report, a rating, a pause before answering, is predicted by both views. Each teacher has a story about why the other's students say what they say. The checker would find no discriminating row.

The attempt has still achieved something. It shows that the dispute, as it stands, is about the fourth layer, and that their evidence is all at the second and third. It suggests what a discriminating observable would need to be: something one view forbids. Perhaps a student trained by one teacher and tested blind by the other. Perhaps not. The protocol's output is an honest map with an empty table.

[AUTHOR NOTE: [AUTHOR NOTE: if your own practice or teachers have taken a position on whether a sense of self remains in deep absorption, describe it in your own words, and say whether you think any observable could separate the two views. Do not let me write this for you.]]

The Kesamutti Sutta, addressed to the Kālāmas, advises against accepting a view because of tradition, scripture, logic or a teacher's standing, and recommends acting "when you know for yourselves". The protocol is a public version of that advice, with one difference. It asks rivals to fix in advance what they would accept as knowing, so that "for yourselves" means for both of you and anyone watching.

An uncomfortable thought experiment

The protocol leaves me no room. I registered a loss on X, so the ledger records a loss on the claim the test covered. I may add a revision entry saying what I now think. I may argue that the translation table was wrong, but that argument should have been made before sealing, and it will be read as what it is. The test does not refute RTM as a whole, any more than COGITATE refuted either of its theories. But it does refute what I said it would refute.

This is uncomfortable because it is the point. A protocol that its author would not accept for his own theory is not worth offering to anyone else.

[AUTHOR NOTE: [AUTHOR NOTE: state, in your own words, whether you would enter RTM into a public contest of this kind, and with which rival. If there is a claim in the theory that you would not put into an outcome table, say which and why.]]

The strongest case against the protocol

The protocol has serious critics, and some of their points survive my replies. I give each objection in its strongest form.

"Without gatekeepers, anyone can demand a contest"

Journals and funders filter, which spares established researchers from having to answer every challenge from every corner. Remove the filter, the critic says, and a theorist with a good result will spend her life answering challengers with no expertise, each demanding a public contest. Worse, a challenger who is invited into a contest gains standing just by being there.

This is right about the cost, and partly right about standing. My reply is that the protocol grants nothing beyond the tables. A challenger who seals a table with a real losing row, and accepts mechanical resolution, is doing exactly what the protocol asks, and I see no reason to exclude them for lacking a post. A challenger who will not register a losing row is marked one-sided, and nothing they say about the result is scored. Nobody is obliged to accept a challenge. The protocol is a way of running a contest well, and it creates no duty to run one.

"Two-party contests freeze the field into camps"

COGITATE pitted two theories against each other, and Kavya and Jonah are two individuals. The critic says that adversarial collaboration, by its form, assumes there are two sides, and so it rewards the theories that already have the largest followings. A third theory that predicts something neither side expects never gets a row.

This is a real risk, and the form encourages it. The protocol does not require two sides. An outcome table can have any number of columns, and the checker would run unchanged with three. The COGITATE authors also noted that their challenges extend to other theories that share the predictions tested. Still, I accept that a protocol built around rivals will tend to recruit rivals. A partial remedy is to publish the outcome rows before sealing and invite any theory to add a column.

"Sides will choose safe losing rows"

A clever side can satisfy the rule while risking nothing. It registers a losing row for an outcome it considers almost impossible, gives that row a probability of 0.01, and enters every plausible outcome as a win or a revision. The test is formally fair and practically empty.

The reply is that the protocol makes this visible. The checker reports each side's own chance of losing, computed from its own table, and the expected evidence the test can provide. A side whose losing rows carry 0.01 between them has told every reader that it does not expect the test to touch it. That leaves the tactic available and puts a price on it. A test in which one side risks thirty per cent and the other risks one per cent is a different test from one in which both risk a quarter, and the published numbers say so.

"Sealing does nothing about which tests are run"

The fourth paper in this series showed that a ledger cannot catch claims that were never entered. The same holds here. A side can propose many tests privately, run the risky ones in its own lab first, and enter the protocol only with tests it already knows it will win.

This is the hardest objection to answer, and I have no full answer. The disagreement map helps a little, because it is published, and it lists every factual disagreement both sides identified, including those they chose not to test. A reader can see which were left out. But a side that keeps private pilots private cannot be caught by any public protocol. The defence is reputational, and it is weak.

"Let a neutral machine decide"

The model has been trained by someone, on text chosen by someone, with preferences tuned by someone. Its judgement inherits all of that, and none of it is sealed or checkable in the way an outcome table is. More basically, the model can only weigh arguments. It cannot supply an observable that both sides accept, which is the thing a contest needs. A model reading Kavya's and Jonah's papers would produce a well-written essay about access. It could not tell them what Wren does when the scratchpad is removed.

I think the critic is right about one narrower use. A model is good at finding the words in a prediction that the translation table does not cover, which is a mechanical check, and the instrument below does a crude version of it in a few lines of code. Used that way, the model is a proofreader for the tables. It decides nothing.

"Mechanical resolution leaves out scientific judgement"

Good science depends on judgement: knowing which anomaly matters, which result is an artefact, which design flaw is fatal. The critic says that a protocol which hands resolution to code throws this away. The code will print a row even if the data are obviously corrupted, and the losing side will be held to a result any competent scientist would have set aside.

I accept that judgement cannot be removed, and the protocol does not try to remove it. It moves judgement to the design stage, where both sides exercise it together and in writing, before either knows the result. Data-quality checks belong in the sealed code, with thresholds both sides agreed. If the data fail those checks, the code prints that the test is void, and no row is scored. What the protocol forbids is judgement exercised afterwards, by one side, about results it dislikes. That is the judgement that most needs constraining.

"Public loss costs the junior more"

A senior researcher who loses a public contest has a career to absorb it. A doctoral student who loses one may not. The critic says that a protocol which makes losses public and permanent will be used mainly by people who can afford to lose, and that this tilts the field towards seniority.

I accept the asymmetry and have no structural fix for it. The best I can offer is a norm. A loss under the protocol is a loss on a sealed row, recorded with the probability the side gave. A side that registered a fair losing row and lost has done something the field should credit. Whether the field will credit it is a question about the field.

What each critic gets right

Each critic names a cost the protocol moves without removing. Filtering moves from editors to the rule about losing rows. Camp formation is eased by open columns and not cured. Safe rows are priced, not banned. The file drawer is narrowed and stays open, the machine judge is useful as a proofreader, and power asymmetry is named and left in place. I would rather state these limits than hide them in a protocol that claims to be complete.

The outcome table checker

The paper's central claims about fairness can be tested directly. I wrote a short program that takes two sides' outcome tables, the outcome definitions and a translation table, and reports four things. It counts the outcomes on which the sides' verdicts differ. It checks that each side has a losing row. It checks that every term used in an outcome is in the translation table. And, if both sides gave probabilities, it computes the expected evidence the test can provide, using the symmetric divergence between their two distributions. It also fingerprints each table, as a stand-in for sealing.

The program runs on seven cases. The first is Kavya and Jonah's test, which is invented. The next four are encoded from the published terms of real contests: the Simon and Ehrlich bet, the Koch and Chalmers bet, and the two phrasings tested by Mellers and her colleagues. For those I enter no probabilities, since none were stated. The last two are an invented one-sided critic and a negative control in which one table is entered twice under two names.

import math, hashlib, json
def seal(obj):
    return hashlib.sha256(json.dumps(obj, sort_keys=True).encode()).hexdigest()[:12]
def check(name, sides, outcomes, glossary, observed=None):
    print(name)
    flags = []
    for o in outcomes:
        for term in o["terms"]:
            if term not in glossary:
                flags.append(f"UNDEFINED TERM '{term}' in outcome {o['id']}")
    disc = [o["id"] for o in outcomes if len({sides[s]["table"][o["id"]] for s in sides}) > 1]
    if not disc:
        flags.append("NO DISCRIMINATION (every outcome gets the same verdict from both sides)")
    for s in sides:
        if "lose" not in sides[s]["table"].values():
            flags.append(f"ONE-SIDED: {s} has no outcome on which it loses")
    names = list(sides)
    info = None
    if all("p" in sides[s] for s in names):
        pa, pb = sides[names[0]]["p"], sides[names[1]]["p"]
        info = sum((pa[k] - pb[k]) * math.log(pa[k] / pb[k]) for k in pa)
    print("  outcomes", len(outcomes), "| discriminating", len(disc), "| sealed", {s: seal(sides[s]["table"]) for s in names})
    for s in names:
        loses = [k for k, v in sides[s]["table"].items() if v == "lose"]
        risk = f" | own chance of losing {sum(sides[s]['p'][k] for k in loses):.2f}" if "p" in sides[s] else ""
        print(f"  {s}: loses on {loses or 'nothing'}{risk}")
    if info is not None:
        print(f"  expected evidence (Jeffreys divergence) {info:.2f} nats")
    print("  verdict:", "; ".join(flags) if flags else "FAIR TEST")
    if observed and not flags:
        v = {s: sides[s]["table"][observed] for s in names}
        bf = ""
        if info is not None:
            bf = f" | log Bayes factor {names[0]} over {names[1]} {math.log(sides[names[0]]['p'][observed] / sides[names[1]]['p'][observed]):+.2f}"
        print(f"  observed {observed}: " + ", ".join(f"{s} {v[s]}" for s in names) + bf)
G = {"integration drop", "single-module drop", "points", "ablation", "basket value", "1980 prices", "frequency format", "conjunction effect", "and are", "who are", "filler items"}
# 1. Running example: Kavya Iyer (workspace) v Jonah Brandt (causal structure). Invented.
O1 = [{"id": "W", "terms": ["ablation", "integration drop", "single-module drop", "points"]},
      {"id": "U", "terms": ["ablation", "integration drop", "single-module drop", "points"]},
      {"id": "N", "terms": ["ablation", "integration drop", "points"]}]
S1 = {"Kavya": {"table": {"W": "win", "U": "lose", "N": "lose"}, "p": {"W": .70, "U": .20, "N": .10}},
      "Jonah": {"table": {"W": "lose", "U": "win", "N": "revise"}, "p": {"W": .25, "U": .55, "N": .20}}}
check("1. Kavya v Jonah: scratchpad ablation (invented)", S1, O1, G, observed="W")
# 2. Simon v Ehrlich, 1980 terms. Probabilities were never stated, so none are entered.
O2 = [{"id": "down", "terms": ["basket value", "1980 prices"]}, {"id": "up", "terms": ["basket value", "1980 prices"]}]
S2 = {"Simon": {"table": {"down": "win", "up": "lose"}}, "Ehrlich": {"table": {"down": "lose", "up": "win"}}}
check("2. Simon v Ehrlich: five metals, 1980 to 1990", S2, O2, G, observed="down")
# 3. Koch v Chalmers, 1998 terms: a 'clear' neural signature by 2023.
O3 = [{"id": "found", "terms": ["clear"]}, {"id": "not found", "terms": ["clear"]}]
S3 = {"Koch": {"table": {"found": "win", "not found": "lose"}}, "Chalmers": {"table": {"found": "lose", "not found": "win"}}}
check("3. Koch v Chalmers: a 'clear' signature by 2023", S3, O3, G)
# 4a/4b. Hertwig v Kahneman, the two phrasings tested by Mellers and colleagues (2001).
O4 = [{"id": "effect", "terms": ["conjunction effect"]}, {"id": "no effect", "terms": ["conjunction effect"]}]
S4a = {"Hertwig": {"table": {"effect": "lose", "no effect": "win"}}, "Kahneman": {"table": {"effect": "win", "no effect": "lose"}}}
S4b = {"Hertwig": {"table": {"effect": "lose", "no effect": "win"}}, "Kahneman": {"table": {"effect": "lose", "no effect": "win"}}}
check("4a. Hertwig v Kahneman: 'and are' condition", S4a, O4, G)
check("4b. Hertwig v Kahneman: 'who are' condition", S4b, O4, G)
# 5. A side that only attacks (vitanda): no outcome on which it loses. Invented.
S5 = {"Kavya": S1["Kavya"], "Critic": {"table": {"W": "revise", "U": "win", "N": "win"}, "p": {"W": .34, "U": .33, "N": .33}}}
check("5. Kavya v a critic who registers no losing outcome (invented)", S5, O1, G)
# 6. Negative control: one theory entered twice under two names.
S6 = {"Kavya": S1["Kavya"], "Kavya-copy": {"table": dict(S1["Kavya"]["table"]), "p": dict(S1["Kavya"]["p"])}}
check("6. Negative control: the same table entered twice", S6, O1, G)

Running it gives the following output. It needs only the Python standard library.

1. Kavya v Jonah: scratchpad ablation (invented)
  outcomes 3 | discriminating 3 | sealed {'Kavya': '2aef70561390', 'Jonah': 'db60cd8e8656'}
  Kavya: loses on ['U', 'N'] | own chance of losing 0.30
  Jonah: loses on ['W'] | own chance of losing 0.25
  expected evidence (Jeffreys divergence) 0.89 nats
  verdict: FAIR TEST
  observed W: Kavya win, Jonah lose | log Bayes factor Kavya over Jonah +1.03
2. Simon v Ehrlich: five metals, 1980 to 1990
  outcomes 2 | discriminating 2 | sealed {'Simon': 'a0acdc00a1f6', 'Ehrlich': '0e8b2e4c5250'}
  Simon: loses on ['up']
  Ehrlich: loses on ['down']
  verdict: FAIR TEST
  observed down: Simon win, Ehrlich lose
3. Koch v Chalmers: a 'clear' signature by 2023
  outcomes 2 | discriminating 2 | sealed {'Koch': '37dec5a38bfb', 'Chalmers': 'a588411cec03'}
  Koch: loses on ['not found']
  Chalmers: loses on ['found']
  verdict: UNDEFINED TERM 'clear' in outcome found; UNDEFINED TERM 'clear' in outcome not found
4a. Hertwig v Kahneman: 'and are' condition
  outcomes 2 | discriminating 2 | sealed {'Hertwig': 'b92cbd7d0c07', 'Kahneman': '34c9b0f6a159'}
  Hertwig: loses on ['effect']
  Kahneman: loses on ['no effect']
  verdict: FAIR TEST
4b. Hertwig v Kahneman: 'who are' condition
  outcomes 2 | discriminating 0 | sealed {'Hertwig': 'b92cbd7d0c07', 'Kahneman': 'b92cbd7d0c07'}
  Hertwig: loses on ['effect']
  Kahneman: loses on ['effect']
  verdict: NO DISCRIMINATION (every outcome gets the same verdict from both sides)
5. Kavya v a critic who registers no losing outcome (invented)
  outcomes 3 | discriminating 3 | sealed {'Kavya': '2aef70561390', 'Critic': '39a07b2f22bb'}
  Kavya: loses on ['U', 'N'] | own chance of losing 0.30
  Critic: loses on nothing | own chance of losing 0.00
  expected evidence (Jeffreys divergence) 0.60 nats
  verdict: ONE-SIDED: Critic has no outcome on which it loses
6. Negative control: the same table entered twice
  outcomes 3 | discriminating 0 | sealed {'Kavya': '2aef70561390', 'Kavya-copy': '2aef70561390'}
  Kavya: loses on ['U', 'N'] | own chance of losing 0.30
  Kavya-copy: loses on ['U', 'N'] | own chance of losing 0.30
  expected evidence (Jeffreys divergence) 0.00 nats
  verdict: NO DISCRIMINATION (every outcome gets the same verdict from both sides)

The positive cases are the first two. Kavya and Jonah's test has three outcomes, all three discriminating, and each side has a losing row. Kavya's own chance of losing, by her own numbers, is 0.30. Jonah's is 0.25. Both are taking a real risk. The expected evidence is 0.89 nats, a moderate amount for a single test. A nat is a unit of evidence on the scale of log odds. The divergence adds up the evidence each side expects the test to give in its own favour. Kavya expects 0.45 nats and Jonah 0.44, so each expects the odds to move by a factor of about 1.6 if they are right. Simon and Ehrlich's bet also passes: two rows, both discriminating, one losing row each, and a resolution by published prices.

The negative cases fail for three different reasons. The Koch and Chalmers bet fails on an undefined term, because "clear" is in neither side's translation table. The checker would have flagged it in 1998, at the moment of sealing. The "who are" condition fails for lack of discrimination: both sides registered the same verdicts, and the two sealed fingerprints are identical, which is a quick way to see it. The critic in case five fails as one-sided: it has no losing row, and its own chance of losing is zero.

The negative control is case six. One table entered twice should show no discrimination and zero expected evidence, and it does. If the checker had found a fair test between a theory and itself, that would have been a bug.

A minimal worked case: two rows

The smallest table that shows every step is the Simon and Ehrlich bet. There are two outcomes, "basket value down" and "basket value up", both defined in terms of quantities and 1980 prices, which both sides accept. Simon's table reads win, lose, and Ehrlich's reads lose, win. Two rows discriminate. Each side has one losing row, no term is undefined, and the checker returns a fair test. Resolution needs one published number in October 1990, and the cheque follows.

Notice what the minimal case leaves out. The checker cannot see that Ehrlich later questioned whether metal prices measured scarcity at all. That doubt lived in a part of the dispute the table never touched, the definitional part. A fair test of a two-row table settles two rows.

A realistic worked case: the Wren test

Achieng runs the sealed code on the agent. With the scratchpad in place, Wren completes 78 per cent of integration tasks and 91 per cent of single-module tasks. With the scratchpad replaced by an empty buffer, integration accuracy falls to 47 per cent, a drop of 31 points, and single-module accuracy falls to 88 per cent, a drop of 3. The code prints outcome W. The gentler intervention, scrambling the scratchpad's contents, gives a drop of 26 points and 4 points, which is also W. The false-value trace shows the planted value used by all three modules within one step.

By the sealed tables, Kavya wins on this row and Jonah loses. The log Bayes factor in Kavya's favour is 1.03 nats, which means the result was about 2.8 times more probable under her table than under his. That is modest. It moves weight between their theories without ending the argument, which is what the Bayesian reading of adversarial collaboration expects.

The counterexample

The most instructive case is 4a, the "and are" condition. The checker passes it as fair: two discriminating rows, one losing row each, no undefined term. And yet the real result was read differently by the two rivals, because a variable outside the table, the presence of filler items, changed how the effect looked. A test the checker passes can still fail to settle its dispute.

This is a counterexample to any claim that the checker certifies a contest. It certifies that the tables are well formed. It cannot see variables that nobody wrote down. The protocol's remedy is in step three, the map of methods disagreements, which asks both sides to list every procedural choice they think could change the row. The Mellers test, run today, would list filler items there.

Where the method stops working

The protocol stops working in four places. It stops where no observable separates the sides, as at the fourth and fifth layers. It stops where a side keeps private pilots private, because the file drawer is invisible. It stops where the translation table is wrong, because a fair test of the wrong rows settles the wrong question. And it stops where a side will not accept a loss it registered, because nothing in a public protocol compels anyone to keep their word. In that last case the record still shows who broke it.

These limits mark the protocol's proper scope. It is designed for rivals who share enough commitment to evidence to want a contest, and who disagree enough to need one. Outside that scope, between people who will not register a loss or who have no observable in common, it produces maps, flags and empty tables. Those outputs are honest descriptions of a dispute that cannot yet be settled, and they are worth publishing for that reason alone.

What would change my mind

I hold the protocol's central claims with different confidence, and each can fail in a specific way.

I would give up the claim that agreeing definitions first tends to decide a contest in advance if a study of published adversarial collaborations found that agreed definitions predicted neither side's win more often than chance. That study could be done now, by coding the definitions sections of past collaborations and their outcomes. Nobody, to my knowledge, has done it.

I would give up the claim that sealed tables can replace an institution as registrar if a contest run under the protocol collapsed because one side disputed what had been sealed, and the hash record failed to settle the dispute. I think this unlikely, since a hash either matches or it does not. The more likely failure is softer: a side accepts that the table was sealed and then argues that the table did not mean what it says.

Consider next the claim that disputes at the fourth and fifth layers cannot yet be given outcome tables. I would give it up if anyone proposed an observable that both a defender and a critic of a self-recognition claim would sign as deciding it. I would be glad to be shown one.

And I would give up the speculative claim that a volunteer protocol can sustain completed contests if, by the end of 2030, no contest in consciousness science has been run to resolution without a funder or journal acting as registrar. That is a prediction, and it is entered in the ledger below.

What I still don't know

I do not know how many rows a table can have before rivals stop reading each other's. Kavya and Jonah had four, while COGITATE had several predictions per theory, each with conditions. Somewhere between, tables become too long for anyone outside to check, and the advantage of a public record shrinks.

I do not know how to handle a side that changes its mind halfway through a contest, honestly, because it has learned something from the design. The protocol treats every change after sealing as a new entry, which is right for the record. It may be wrong for the relationship.

I do not know whether three voices are enough. In COGITATE the consortium's reading came first. A protocol without a consortium has no first voice, only the data statement. Whether a bare data statement is read as neutral, or as the winner's voice, is something only use will show.

And I do not know how to weigh the cost to a junior researcher of losing in public. I have named it, which is not the same as solving it.

Claim ledger

These are the paper's main claims, entered under the rules of the fourth paper. Confidence uses the series' bands: established (0.90 or above), likely (0.60 to 0.89) and speculative (0.30 to 0.59).

There are two dated predictions. First, by 31 December 2030, at least one adversarial contest between rival theories of consciousness will be published with outcome tables fingerprinted or time-stamped in public before data collection, and with no funder or journal acting as registrar. I will judge it by a literature search and record the first example, or record the prediction as failed. Second, by 31 December 2029, this series will have entered at least one RTM claim into a public contest of this kind with a named rival, with both outcome tables sealed before data. That is a prediction about my own conduct, and it is scored the same way as the others.

Glossary

Adoṣodbhāvana अदोषोद्भावन

In Dharmakīrti's account, failing to point out a real fault: the opponent's condition of defeat.

Adversarial collaboration

A study designed and run jointly by researchers who disagree, with the aim of testing their rival predictions.

Asādhanāṅgavacana असाधनाङ्गवचन

In Dharmakīrti's account, failing to state what proves the thesis: the proponent's condition of defeat.

Definition escrow

Each side writes and seals its own definitions, which the other need not accept. The sealed text fixes what each side meant before results.

Disagreement map

A published list of every point in dispute, sorted into facts, methods, definitions and values.

Escrowed dispute

A dispute that has been written down and sealed but cannot yet be tested, because no shared observable exists.

Jalpa जल्प

Debate aimed at victory, which allows quibbles, trick replies and appeals to points of defeat.

Jeffreys divergence

A symmetric measure of how different two probability distributions are. Here, the evidence a test can be expected to provide.

Nigrahasthāna निग्रहस्थान

A condition of defeat in debate. Nyāya lists twenty-two.

Outcome table

A sealed table in which a side states, for each possible outcome, whether its theory wins, loses or needs revision, with a probability for each.

Pariṣad परिषद्

The assembly before which a debate takes place. Caraka distinguishes friendly, neutral and prejudiced assemblies.

Three-voice report

A write-up with a data statement followed by each side's reading, published unchanged.

Translation table

A shared list linking each term in the predictions to an observable that both sides accept as evidence about it.

Vāda वाद

Discussion aimed at truth, argued from sources of knowledge and reasoning. The protocol takes its name from it.

Vitaṇḍā वितण्डा

Debate in which one side only attacks and establishes no position of its own.

The result that both of them signed

Kavya and Jonah publish the Wren result on a Tuesday. The page carries the two definition fingerprints, the disagreement map and the translation table. It also carries both sealed outcome tables, the code, the data, and Achieng's data statement: outcome W, a drop of 31 points on integration tasks and 3 on single-module tasks, with the gentler intervention agreeing. Below that are two readings, each written by one of them and published unchanged.

Kavya's reading says that the scratchpad behaves as a functional workspace for Wren's modules, by the observables both sides agreed. It does not say Wren has access in any further sense, and it says nothing about experience. Jonah's reading says that he registered a loss on this row and records it. He adds that the result shows the scratchpad to be a bottleneck in Wren's information flow, which he now accepts, and that it leaves his claim about causal structure untouched, since that claim was escrowed and never entered a row.

Both enter the result in their ledgers, where Kavya's prediction, at 0.70, resolves true. Jonah's, at 0.55 for the uniform pattern, resolves false, and he adds a revision entry. The escrowed dispute, about whether a broadcast is enough for access in Jonah's sense, stays sealed. Nobody has proposed a row for it yet.

The contest did not end their argument. It made one part of it public and settled, so that the next round can start from there. They have already begun a second map, for a test with a second agent built to have the same scratchpad and a different internal wiring. Neither of them needed a journal, a funder or a referee to do it. They needed to write down, before they knew the answer, what each answer would mean.

The editor who declined to host the contest writes to them a month later. She would like to publish the result as a short report, with both readings, exactly as posted. They agree, on one condition: that the report links to the sealed tables, so that any reader can check that the published result is the one they committed to in advance. The journal arrives after the contest, as a reader with a large audience. It was never needed as a referee.

[AUTHOR NOTE: [AUTHOR NOTE: if you intend to invite a named rival to a public contest of this kind, name the rival and the claim here, or delete this note.]]

Sources

Cite this work

@misc{mahapatra_U0011,
  author       = {Mahapatra, Rajesh Ranjan},
  title        = {Adversarial collaboration without institutional gatekeepers:
                  a public protocol for rivals who disagree about methods
                  and definitions},
  howpublished = {Mind Frontier, paper U0011},
  url          = {https://rajeshrmahapatra.com/papers/mind-frontier/U0011-adversarial-collaboration-without-institutional-gatekeepers.html},
  note         = {DOI: to be assigned}
}