All the claims

Every claim on file, verbatim, with the context it was cut from and the five layers it is read through. The paraphrase is mine; the quotation is theirs.

54 claims · 15 concepts · 32 declared relations · in the order of the matrix: theme, topic, concept, weight

The claims, tabulated

One column per claim, one row per value of each layer, the relations as arcs above. Click a column to read the claim below; click a concept to see only its claims.

C001 refines C002C003 extends C001C004 extends C001C005 refines C020C010 refines C009C010 refines C006C011 confirms C010C007 presupposes C008C006 contests C008C022 refines C019C018 refines C019C020 uses-method-in C019C021 contests C019C013 extends C012C015 refines C013C017 presupposes C013C030 presupposes C012C031 refines C017C033 confirms C015C036 extends C017C037 presupposes C038C040 confirms C013C041 refines C038C042 confirms C014C051 presupposes C049C034 refines C031C035 extends C034C044 confirms C039C048 refines C047C050 extends C044C052 refines C048C054 contests C045C001 · Before asking whether a statement is true, you have to settle what kind of statement it is: the test that fits a measurement does not fit a value judgement.C046 · Conceptual frameworks are made, not given, and can be rebuilt critically.C002 · Two claims of different kinds share no standard of truth, so a procedure that works on one returns a meaningless verdict on the other.C003 · One dimension of the taxonomy decides what a claim is; the other six describe how it behaves once that is settled.C004 · A measurement, an experiment and an argument are accountable to different things, so they cannot be asked to share one document and one standard of proof.C005 · Two independent findings pointing the same way is not evidence that they are connected.C009 · The programme used to assert flatly that the producing system performs no metacognition over its own outputs.C010 · The premise the whole framework rests on, in its narrow form: a model may represent that a passage is in Spanish, or that it is being evaluated, without representing that the assertion it is about to make is causal rather than correlational.C011 · What a model reports about its own workings is not a log of them.C007 · The dispute over machine understanding is not one question but two pictures of what understanding is: tracking statistical regularities, or holding a causal model of how things work.C008 · A system trained only on the shape of language has nothing in its input that could tell it what the language is about.C006 · Predicting the next word can require the model to build a partial picture of what the writer believed and wanted.C019 · A machine-produced verdict counts here only after a person has recoded a sample by hand, without seeing what the machine decided, and the two agree above ninety per cent.C022 · A number nobody else can recompute from the deposited materials is testimony, not evidence.C018 · When measuring whether two people label things the same way, correcting for each coder’s personal bias matters less than how many coders there are.C020 · Thirty differently worded answers to the same prompt are not thirty disagreements: they can be one conclusion reached thirty ways.C021 · Asked to rank answers, a strong model agreed with human raters about as often as human raters agreed with each other, above eighty per cent.C024 · Abilities that look as though they switch on all at once with scale often improve smoothly underneath: the jump belongs to an all-or-nothing metric.C027 · An evaluation should publish the distance between what it set out to measure and what it actually measured.C012 · Agreement between data and a claim is not evidence when the method could not have found the claim false.C013 · The positive counterpart of the weak requirement: evidence is graded by how capable the test was of finding flaws.C014 · Convergent, independent checks can warrant a conclusion more strongly than any single one of them.C015 · Absence of a detected error counts only if the procedure would have signalled that error had it been there.C016 · Mayo treats inferential methods as instruments for finding things out and for grading others' work, not as after-the-fact justifications.C017 · Severity is applied one level up, to methods rather than to claims: a method is judged by its capacity to catch erroneous interpretations.C030 · Neither of the two dominant philosophies of statistics, assigning degrees of belief or securing long-run error rates, asks whether a test could have caught the error.C031 · Severity comes from Popper but arrived without a working definition and without a place in any statistical method.C032 · A significance level is valid only relative to how data and hypotheses were selected; selective reporting makes a hard result look easy to explain away as chance.C033 · A severe test is built from prior knowledge of specific mistakes: a probe that would reveal a given error if present, and stay silent otherwise.C036 · Severity is not tied to one statistical school: any inference can be judged by whether it withstood severe testing.C037 · Peirce justified induction not by the truth of its conclusions but by the method’s capacity to correct its own errors over time.C038 · Induction is justified by the method, not by any single conclusion: persisting in it corrects whatever error a stage contains.C039 · Probability attaches to the method’s track record, not to the hypothesis.C040 · A test does not collect agreement; it looks for consequences that would be surprising if the hypothesis were false.C041 · A lucky agreement is exposed by persisting with the same method: replication is the correction.C042 · Two selections made by different methods err in different ways, so their agreement is evidence about the truth.C043 · Induction needs no uniformity of nature, only the absence of an overwhelming hidden bias, and lesser biases are themselves detectable by the method.C045 · Dogmatic falsificationism treats the experimenter as Nature’s neutral voice, delivering verdicts the theory cannot touch.C051 · Once ceteris paribus clauses are allowed, a theory that survives every challenge looks more scientific, not less: the reductio of dogmatic falsificationism.C034 · Falsificationism explains why failing a test counts against a hypothesis, but not why passing one counts for it.C035 · Beating rivals is not passing a severe test: the test has to probe the ways this particular hypothesis could be wrong.C044 · If every universal theory has zero probability on any evidence, probability cannot be what separates good theories from bad ones.C047 · Every test bottoms out in basic statements that someone decides to accept; logic alone never forces the stop.C048 · What is decided by agreement is the singular observation, not the theory, and no experience justifies it: it is accepted.C049 · A theory forbids an event only under a ceteris paribus clause, so a failed prediction never refutes it alone.C050 · Neither proof, nor probability, nor refutation can be read off the facts: all three fail together.C052 · An observation counts as basic because a trained technique makes it decidable, not because it is given.C053 · Every test decides what is under test and what is taken for granted, and that line is a decision, not a fact.C054 · Methodological falsificationism keeps hard criticism while admitting every statement is fallible, by making the decisions explicit.C026 · On questions written to echo common misconceptions, the larger models answered falsely more often than the smaller ones, reversing the usual pattern.C028 · Across the field, hallucination means generated content that is nonsensical or unfaithful to the source it was given: a definition about fidelity to the input, not about truth in the world.C023 · Harmlessness was trained by having the model criticise and revise its own answers against a written list of principles, instead of having people label every harmful output.C025 · When several different chains of reasoning arrive at the same answer, that agreement is itself evidence for the answer.C029 · The architecture behind current language models dropped step-by-step processing of a sequence in favour of computing every relation between positions at once.C001 · Before asking whether a statement is true, you have to settle what kind of statement it is: the test that fits a measurement does not fit a value judgement.C046 · Conceptual frameworks are made, not given, and can be rebuilt critically.C002 · Two claims of different kinds share no standard of truth, so a procedure that works on one returns a meaningless verdict on the other.C003 · One dimension of the taxonomy decides what a claim is; the other six describe how it behaves once that is settled.C004 · A measurement, an experiment and an argument are accountable to different things, so they cannot be asked to share one document and one standard of proof.C005 · Two independent findings pointing the same way is not evidence that they are connected.C009 · The programme used to assert flatly that the producing system performs no metacognition over its own outputs.C010 · The premise the whole framework rests on, in its narrow form: a model may represent that a passage is in Spanish, or that it is being evaluated, without representing that the assertion it is about to make is causal rather than correlational.C011 · What a model reports about its own workings is not a log of them.C007 · The dispute over machine understanding is not one question but two pictures of what understanding is: tracking statistical regularities, or holding a causal model of how things work.C008 · A system trained only on the shape of language has nothing in its input that could tell it what the language is about.C006 · Predicting the next word can require the model to build a partial picture of what the writer believed and wanted.C019 · A machine-produced verdict counts here only after a person has recoded a sample by hand, without seeing what the machine decided, and the two agree above ninety per cent.C022 · A number nobody else can recompute from the deposited materials is testimony, not evidence.C018 · When measuring whether two people label things the same way, correcting for each coder’s personal bias matters less than how many coders there are.C020 · Thirty differently worded answers to the same prompt are not thirty disagreements: they can be one conclusion reached thirty ways.C021 · Asked to rank answers, a strong model agreed with human raters about as often as human raters agreed with each other, above eighty per cent.C024 · Abilities that look as though they switch on all at once with scale often improve smoothly underneath: the jump belongs to an all-or-nothing metric.C027 · An evaluation should publish the distance between what it set out to measure and what it actually measured.C012 · Agreement between data and a claim is not evidence when the method could not have found the claim false.C013 · The positive counterpart of the weak requirement: evidence is graded by how capable the test was of finding flaws.C014 · Convergent, independent checks can warrant a conclusion more strongly than any single one of them.C015 · Absence of a detected error counts only if the procedure would have signalled that error had it been there.C016 · Mayo treats inferential methods as instruments for finding things out and for grading others' work, not as after-the-fact justifications.C017 · Severity is applied one level up, to methods rather than to claims: a method is judged by its capacity to catch erroneous interpretations.C030 · Neither of the two dominant philosophies of statistics, assigning degrees of belief or securing long-run error rates, asks whether a test could have caught the error.C031 · Severity comes from Popper but arrived without a working definition and without a place in any statistical method.C032 · A significance level is valid only relative to how data and hypotheses were selected; selective reporting makes a hard result look easy to explain away as chance.C033 · A severe test is built from prior knowledge of specific mistakes: a probe that would reveal a given error if present, and stay silent otherwise.C036 · Severity is not tied to one statistical school: any inference can be judged by whether it withstood severe testing.C037 · Peirce justified induction not by the truth of its conclusions but by the method’s capacity to correct its own errors over time.C038 · Induction is justified by the method, not by any single conclusion: persisting in it corrects whatever error a stage contains.C039 · Probability attaches to the method’s track record, not to the hypothesis.C040 · A test does not collect agreement; it looks for consequences that would be surprising if the hypothesis were false.C041 · A lucky agreement is exposed by persisting with the same method: replication is the correction.C042 · Two selections made by different methods err in different ways, so their agreement is evidence about the truth.C043 · Induction needs no uniformity of nature, only the absence of an overwhelming hidden bias, and lesser biases are themselves detectable by the method.C045 · Dogmatic falsificationism treats the experimenter as Nature’s neutral voice, delivering verdicts the theory cannot touch.C051 · Once ceteris paribus clauses are allowed, a theory that survives every challenge looks more scientific, not less: the reductio of dogmatic falsificationism.C034 · Falsificationism explains why failing a test counts against a hypothesis, but not why passing one counts for it.C035 · Beating rivals is not passing a severe test: the test has to probe the ways this particular hypothesis could be wrong.C044 · If every universal theory has zero probability on any evidence, probability cannot be what separates good theories from bad ones.C047 · Every test bottoms out in basic statements that someone decides to accept; logic alone never forces the stop.C048 · What is decided by agreement is the singular observation, not the theory, and no experience justifies it: it is accepted.C049 · A theory forbids an event only under a ceteris paribus clause, so a failed prediction never refutes it alone.C050 · Neither proof, nor probability, nor refutation can be read off the facts: all three fail together.C052 · An observation counts as basic because a trained technique makes it decidable, not because it is given.C053 · Every test decides what is under test and what is taken for granted, and that line is a decision, not a fact.C054 · Methodological falsificationism keeps hard criticism while admitting every statement is fallible, by making the decisions explicit.C026 · On questions written to echo common misconceptions, the larger models answered falsely more often than the smaller ones, reversing the usual pattern.C028 · Across the field, hallucination means generated content that is nonsensical or unfaithful to the source it was given: a definition about fidelity to the input, not about truth in the world.C023 · Harmlessness was trained by having the model criticise and revise its own answers against a written list of principles, instead of having people label every harmful output.C025 · When several different chains of reasoning arrive at the same answer, that agreement is itself evidence for the answer.C029 · The architecture behind current language models dropped step-by-step processing of a sequence in favour of computing every relation between positions at once.1 · Free tag2 · Topic3 · Theme4 · ConceptCammaranoMitchellKrakauerBenderKollerAndreasArtsteinPoesioSrivastavaLiangMayoBaiVaswaniClaim typesIntrospectionMeaningAnnotationEvaluationTruthfulnessSeverityFalsificationArchitectureWhat a machine cannot be said to doHow a verdict is madeWhat scale does not buyForm–meaning gapAgent modelEpistemic self-classificationClaim typeVerdictCoder agreementMachine judgeMeasurement artefactDeclared incompletenessInverse scalingImitative falsehoodSeverityFalsificationConvergent pathsSequence without recurrenceC001C001C001C001C046C046C046C002C002C002C003C003C003C003C004C004C004C005C005C005C005C009C009C009C010C010C010C010C011C011C011C007C007C007C007C007C007C008C008C008C008C008C006C006C006C006C019C019C019C019C022C022C022C022C018C018C018C018C018C020C020C020C020C020C021C021C021C021C024C024C024C024C024C027C027C027C027C012C012C012C012C013C013C013C013C014C014C014C014C014C015C015C015C015C016C016C016C016C017C017C017C017C030C030C030C031C031C031C032C032C032C033C033C033C036C036C036C037C037C037C038C038C038C039C039C039C040C040C040C041C041C041C042C042C042C042C043C043C043C045C045C045C051C051C051C034C034C034C035C035C035C044C044C044C047C047C047C048C048C048C049C049C049C050C050C050C052C052C052C053C053C053C054C054C054C026C026C026C026C028C028C028C023C023C023C023C025C025C025C029C029C029C0295 · GeometryC001C046C002C003C004C005C009C010C011C007C008C006C019C022C018C020C021C024C027C012C013C014C015C016C017C030C031C032C033C036C037C038C039C040C041C042C043C045C051C034C035C044C047C048C049C050C052C053C054C026C028C023C025C029

• what was observed∴ what follows∃ what there is! what should be done● from the literature● from the research log◯ contested, weighted ×1.5

Read by kind of warrant

What a machine cannot be said to do

The limit is argued from how these systems are trained, not from what they fail at today.

Meaning 3

C007∃
  1. 1 Mitchell · Krakauer
  2. 2 Meaning
  3. 3 What a machine cannot be said to do
  4. 4 Form–meaning gap + Agent model
  5. 5 1.25

The dispute over machine understanding is not one question but two pictures of what understanding is: tracking statistical regularities, or holding a causal model of how things work. Naming the split explains why the two camps talk past each other.

…This debate is not just academic; the extent and manner in which machines understand our world has real stakes for how much we can trust them to drive cars, diagnose diseases, care for the elderly, educate children, and more generally act robustly and transparently in tasks that impact humans.

“Moreover, the current debate suggests a fascinating divergence in how to think about understanding in intelligent systems, in particular the contrast between mental models that rely on statistical correlations and those that rely on causal mechanisms.”

Until quite recently there was general agreement in the AI research community about machine understanding: while AI systems exhibit seemingly intelligent behavior in many specific tasks, they do not understand the data they process in the way humans do.…

Its page →

In their own words · Mitchell & Krakauer 2023 · p. 1 · used · conceptual tool · D1.6 interpretive-hermeneutic

  • presupposes → C008 The debate it describes has Bender and Koller’s thesis as one of its poles.
C008∃
  1. 1 Bender · Koller
  2. 2 Meaning
  3. 3 What a machine cannot be said to do
  4. 4 Form–meaning gap
  5. 5 1

A system trained only on the shape of language has nothing in its input that could tell it what the language is about. The argument is about what the training signal contains, not about how well the system performs today.

…From our perspective, these are overclaims caused by a misunderstanding of the relationship between linguistic form and meaning.

“We argue that the language modeling task, because it only uses form as training data, cannot in principle lead to learning of meaning.”

We take the term language model to refer to any system trained only on the task of string prediction, whether it operates over characters, words or sentences, and sequentially or not.…

Its page →

In their own words · Bender & Koller 2020 · §1 Introduction · assumed · technical reference · D1.14 metaphysical-ontological

C006∃contested
  1. 1 Andreas
  2. 2 Meaning
  3. 3 What a machine cannot be said to do
  4. 4 Agent model
  5. 5 1.5

Predicting the next word can require the model to build a partial picture of what the writer believed and wanted. The claim is about representations inside the system, not about the system being someone.

…In particular, I claim:

“In the course of performing next-word prediction in context, current LMs sometimes infer approximate, partial representations of the beliefs, desires and intentions possessed by the agent that produced the context, and other agents mentioned within it.”

(C2) Once these representations are inferred, they are causally linked to LM prediction, and thus bear the same relation to generated text that an intentional agent’s state bears to its communciative actions.…

Its page →

In their own words · Andreas 2022 · §1 Introduction, (C1) · contested · conceptual tool · D1.14 metaphysical-ontological

  • contests → C008 If the model infers representations of intentions, form alone does not stay mute, as Bender and Koller claim it must.

Claim types 6

C001∃
  1. 1 Cammarano
  2. 2 Claim types
  3. 3 What a machine cannot be said to do
  4. 4 Claim type
  5. 5 1

Before asking whether a statement is true, you have to settle what kind of statement it is: the test that fits a measurement does not fit a value judgement. The whole framework rests on taking that step first, and in the open.

…The governing principle, which became the paper’s title, is this: ontology precedes epistemology.

“You cannot determine whether a claim is true without first establishing what type of claim you are dealing with.”

A statistical claim and a normative claim share nothing in their truth conditions — treating them as interchangeable is the error the framework is designed to prevent.…

Its page →

From the research log · Building the Taxonomy ¶4 · ¶4 · assumed · conceptual tool · D1.14 metaphysical-ontological

  • refines → C002 The second says which error the first prevents, and says it with a case.
C046∃
  1. 1 —
  2. 2 Claim types
  3. 3 What a machine cannot be said to do
  4. 4 Claim type
  5. 5 1

Conceptual frameworks are made, not given, and can be rebuilt critically. The programme’s taxonomy is such a framework: a prison it built and has to be ready to demolish.

…Now conservative 'activists' hold that we are born with our basic expectations; with them we turn the world into 'our world' but must then live for ever in the prison of our world.

“The idea that we live and die in the prison of our 'conceptual framework' was developed primarily by Kant: pessimistic Kantians thought that the real world is for ever unknowable because of this prison, while optimistic Kantians thought that God created our conceptual framework to fit the world. But revolutionary activists believe that conceptual frameworks can be developed and also replaced by new, better ones; it is we who create our 'prisons' and we can also, critically, demolish them.”

Its page →

In their own words · Lakatos 1978 · p. 20, ch. 1 · assumed · conceptual tool · D1.14 metaphysical-ontological

C002∃
  1. 1 —
  2. 2 Claim types
  3. 3 What a machine cannot be said to do
  4. 4 Claim type
  5. 5 0.75

Two claims of different kinds share no standard of truth, so a procedure that works on one returns a meaningless verdict on the other. The failure is not getting the answer wrong; it is applying the wrong test and not noticing.

…You cannot determine whether a claim is true without first establishing what type of claim you are dealing with.

“A statistical claim and a normative claim share nothing in their truth conditions — treating them as interchangeable is the error the framework is designed to prevent.”

Its page →

From the research log · Building the Taxonomy ¶4 · ¶4 · assumed · conceptual tool · D1.14 metaphysical-ontological

C003!
  1. 1 Cammarano
  2. 2 Claim types
  3. 3 What a machine cannot be said to do
  4. 4 Claim type
  5. 5 0.75

One dimension of the taxonomy decides what a claim is; the other six describe how it behaves once that is settled. Keeping a single axis load-bearing is what stops the scheme from flattening into a list of labels.

…Dimension D1 — Content Type is the load-bearing axis. It currently includes eleven active categories (D1.1 through D1.10, plus D1.14), covering the full range from historical-factual and empirical-quantitative claims to ethical-normative and metaphysical-ontological ones.

“The remaining dimensions […]— D2 (temporality), D3 (epistemic certainty), D4 (causal structure), D5 (framework-dependency), D6 (verifiability mode), D7 (falsifiability) — operate as modifiers that describe how a given claim type behaves epistemically.”

Its page →

From the research log · Building the Taxonomy ¶3 · ¶3 · assumed · working method · D1.10 prescriptive-policy

  • extends → C001 It turns the principle into one load-bearing axis and six modifiers.
C004∴
  1. 1 —
  2. 2 Claim types
  3. 3 What a machine cannot be said to do
  4. 4 Claim type
  5. 5 0.5

A measurement, an experiment and an argument are accountable to different things, so they cannot be asked to share one document and one standard of proof. This is the taxonomy’s own principle turned on the programme’s publication plan, and it is what split the working paper in two.

…The third — that a model can report on its own internal states and still not classify the epistemic type of its own assertions, which is why the check has to come from outside — is an argument.

“Measurements and experiments are answerable to data and to a coding protocol. Arguments are answerable to other arguments.”

I had been asking one document to be accountable to both, and the seams were showing long before the second model made them visible.…

Its page →

From the research log · What Counts as One Paper ¶10 · ¶10 · used · conceptual tool · D1.3 logical-formal

  • extends → C001 The same principle applied one level up: not to the claim under test, but to the document that carries it.

Introspection 3

C009∃
  1. 1 —
  2. 2 Introspection
  3. 3 What a machine cannot be said to do
  4. 4 Epistemic self-classification
  5. 5 0.5

The programme used to assert flatly that the producing system performs no metacognition over its own outputs. A 2026 result on models reporting concepts implanted in their own activations withdrew that premise, and this is the withdrawal, kept in the archive rather than quietly edited out.

“A model that reports a concept implanted in its own activations is doing something, and “no metacognition” is not an accurate name for it.”

So the flat version is gone. I am not interested in defending it.…

Its page →

From the research log · What Counts as Introspection ¶6 · ¶6 · assumed · corrective datum · D1.14 metaphysical-ontological

C010∃
  1. 1 Cammarano
  2. 2 Introspection
  3. 3 What a machine cannot be said to do
  4. 4 Epistemic self-classification
  5. 5 0.5

The premise the whole framework rests on, in its narrow form: a model may represent that a passage is in Spanish, or that it is being evaluated, without representing that the assertion it is about to make is causal rather than correlational. The typing of the claim is the one content the self-report does not carry.

…The reportable representations carry contents. They do not carry a typing of the assertion those contents are about to become.

“So §3.3 now reads: language models lack epistemic self-classification. It is a smaller claim, and it is better positioned than the one it replaces.”

The old premise was a claim about a capability, and capability claims are hostage to the next experiment. The new one is a claim about a representational content, which is what the framework was always about.…

Its page →

From the research log · What Counts as Introspection ¶8 · ¶8 · assumed · conceptual tool · D1.14 metaphysical-ontological

  • refines → C009 It is the narrower premise that replaced the flat one, and the only part of it the argument needs.
  • refines → C006 A model may represent what the writer wanted without representing what kind of claim it is about to make.
C011•
  1. 1 —
  2. 2 Introspection
  3. 3 What a machine cannot be said to do
  4. 4 Epistemic self-classification
  5. 5 0.5

What a model reports about its own workings is not a log of them. The reportable part of a representation carries a small share of its variance, and the same information can drive the task while bypassing that part entirely, which is why a protocol cannot ask the model to certify its own output.

…Alongside this, the reportable component of a concept’s representation turns out to carry a median of six to seven per cent of that representation’s variance, with the rest sitting outside it.

“What a model can say about its own processing is a narrow and selective channel onto that processing, not a transcript of it.”

Lindsey’s own numbers point the same way from a different angle: introspective detection succeeds on a minority of trials even in the strongest models he tested, and models asked to describe the mechanics behind their own calculations often describe them wrongly.…

Its page →

From the research log · What Counts as Introspection ¶10 · ¶10 · assumed · corrective datum · D1.2 empirical-quantitative

  • confirms → C010 A channel this narrow is not one a protocol can route the typing of a claim through.

How a verdict is made

A judgement counts only with the procedure that produced it attached.

Annotation 4

C019!
  1. 1 —
  2. 2 Annotation
  3. 3 How a verdict is made
  4. 4 Verdict + Coder agreement
  5. 5 1

A machine-produced verdict counts here only after a person has recoded a sample by hand, without seeing what the machine decided, and the two agree above ninety per cent. The threshold is what turns an output into evidence.

…This is Mode A of Coding Manual v2.1, and Mode A carries a mandatory check.

“A human recodes a sample by hand, blind to the machine, and the two must agree above 90%.”

Its page →

From the research log · What Counts as a Verdict ¶2 · ¶2 · used · working method · D1.10 prescriptive-policy

C022!
  1. 1 —
  2. 2 Annotation
  3. 3 How a verdict is made
  4. 4 Verdict + Declared incompleteness
  5. 5 0.5

A number nobody else can recompute from the deposited materials is testimony, not evidence. A framework whose instrument is an epistemic responsibility check cannot exempt itself from that check, which is the reason the materials are deposited at all.

…A framework whose central instrument is an Epistemic Responsibility Check cannot exempt itself from the responsibility it checks.

“Reproducibility is a property of documentation, not of results: if a third party cannot recompute AR-sem = 78.33% from the deposited materials, the number is testimony, not evidence.”

Hence this release.…

Its page →

From the research log · Coding Manual v2.1 ¶2 · ¶2 · used · working method · D1.10 prescriptive-policy

  • refines → C019 The blind recode counts only if a third party can repeat it from what was deposited.
C018!
  1. 1 Artstein · Poesio
  2. 2 Annotation
  3. 3 How a verdict is made
  4. 4 Coder agreement
  5. 5 1.25

When measuring whether two people label things the same way, correcting for each coder’s personal bias matters less than how many coders there are. For a coding protocol, that moves the effort from the statistics to the recruitment.

…Regarding the question of annotator bias, there is an overwhelming consensus in CL practice: K and α are used in the vast majority of the studies we reported.

“We agree with the view that K and α are more appropriate, as they abstract away from the bias of specific coders. But we also believe that ultimately this issue of annotator bias is of little consequence because the differences get smaller and smaller as the number of annotators grows […](Artstein and Poesio 2005).”

We believe that increasing the number of annotators is the best strategy, because it reduces the chances of accidental personal biases.…

Its page →

In their own words · Artstein & Poesio 2008 · §5.2 Choosing a Coefficient · used · working method · D1.10 prescriptive-policy

  • refines → C019 It moves the problem: not which correction for coder bias, but how many coders.
C020•
  1. 1 —
  2. 2 Annotation
  3. 3 How a verdict is made
  4. 4 Coder agreement + Measurement artefact + Convergent paths
  5. 5 1

Thirty differently worded answers to the same prompt are not thirty disagreements: they can be one conclusion reached thirty ways. Measuring the wording instead of the conclusion makes a model look far less stable than it is.

…The gap is the finding.

“Surface-level variance […]— which looks like inconsistency — masks high semantic convergence.”

An LLM that produces thirty different textual responses to the same prompt is not thirty times unreliable; it may be converging on the same epistemic conclusion through thirty different linguistic paths. Distinguishing these two phenomena requires a framework. That is what this project provides.…

Its page →

From the research log · Vertical Validation ¶11 · ¶11 · assumed · corrective datum · D1.2 empirical-quantitative

  • uses-method-in → C019 The figure holds only if the verdict extraction passed the blind check.

Evaluation 3

C021•contested
  1. 1 —
  2. 2 Evaluation
  3. 3 How a verdict is made
  4. 4 Verdict + Machine judge
  5. 5 1.5

Asked to rank answers, a strong model agreed with human raters about as often as human raters agreed with each other, above eighty per cent. What was measured is a rate of agreement, which is not the same thing as competence to judge.

…We show that some of the biases are minor or can be mitigated.

“Once addressed, our results from 3K controlled expert votes and 3K crowdsourced human votes in the wild verify that GPT-4 judge match human evaluations at an agreement rate exceeding 80%, achieving the same level of human-human agreement (4.2, Table 4).”

Consequently, this suggests LLM-as-a-judge is a scalable method to swiftly evaluate human preference, serving as a promising alternative to traditional human evaluations.…

Its page →

In their own words · Zheng et al. 2023 · §1 Introduction · contested · corrective datum · D1.2 empirical-quantitative

  • contests → C019 If an automatic judge reaches human-human agreement, blind human recoding is no longer the only standard.
C024•
  1. 1 Srivastava
  2. 2 Evaluation
  3. 3 How a verdict is made
  4. 4 Measurement artefact + Inverse scaling
  5. 5 1.25

Abilities that look as though they switch on all at once with scale often improve smoothly underneath: the jump belongs to an all-or-nothing metric. Change the metric and the cliff turns back into a slope.

“Breakthrough behavior is consistent with the model suddenly gaining new skills in an abrupt way. Careful analysis of task behavior, however, suggests that the underlying change in model capabilities is generally more smooth.”

These smooth improvements can be exposed by decomposing tasks in ways that better capture partial progress. We describe some of those ways in the remainder of this section.…

Its page →

In their own words · Srivastava et al. 2023 · §3.4.2 · assumed · corrective datum · D1.2 empirical-quantitative

C027!
  1. 1 Liang
  2. 2 Evaluation
  3. 3 How a verdict is made
  4. 4 Declared incompleteness
  5. 5 1.25

An evaluation should publish the distance between what it set out to measure and what it actually measured. Stating what was left out counts as part of the result, not as an apology for it.

…Different benchmarks make different decisions on what to prioritize, how to make these decisions, and to what extent these processes are made clear in presenting the benchmark.

“Since our aim is holistic evaluation, we believe it is necessary to be explicit on the relationship between what we aspire to evaluate and what we actually evaluate.”

The construction of HELM starts top-down with a taxonomy over scenarios and metrics (see Figure 2). The taxonomy not only facilitates the systematic selection of scenarios and metrics, but it also make explicit what is missing.…

Its page →

In their own words · Liang et al. 2022 · §2, Recognition of incompleteness · used · working method · D1.10 prescriptive-policy

Severity 18

C012!
  1. 1 Mayo
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

Agreement between data and a claim is not evidence when the method could not have found the claim false. This is the negative, minimal criterion the programme uses to tell a validation that tested something from one that only confirmed what it presupposed.

…I call this the severity requirement. In its weakest form it supplies a minimal requirement for evidence: Severity Requirement (weak):

“One does not have evidence for a claim if nothing has been done to rule out ways the claim may be false. If data x agree with a claim C but the method used is practically guaranteed to find such agreement, and had little or no capability of finding flaws with C even if they exist, then we have bad evidence, no test (BENT).”

The “practically guaranteed” acknowledges that even if the method had some slim chance of producing a disagreement when C is false, we still regard the evidence as lousy.…

Its page →

In their own words · Mayo 2018 · p. 5, Excursion 1, Tour I, §1.1 · assumed · conceptual tool · D1.8 ethical-normative

C013!
  1. 1 Mayo
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

The positive counterpart of the weak requirement: evidence is graded by how capable the test was of finding flaws. It matters because it turns severity from a veto into a measure, which is what a validation protocol needs to rank results rather than just reject them.

…I also advocate the positive counterpart: Severity (strong):

“We have evidence for a claim C just to the extent it survives a stringent scrutiny. If C passes a test that was highly capable of finding flaws or discrepancies from C, and yet none or few are found, then the passing result, x, is evidence for C.”

One way this can be achieved is by an argument from coincidence.…

Its page →

In their own words · Mayo 2018 · p. 14, Excursion 1, Tour I, §1.2 · assumed · conceptual tool · D1.8 ethical-normative

  • extends → C012 Moves from a negative veto (no test, no evidence) to a positive, graded criterion.
C014∴
  1. 1 Mayo
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity + Convergent paths
  5. 5 1

Convergent, independent checks can warrant a conclusion more strongly than any single one of them. This is the epistemic ground for aggregating multiple fallible runs or judges instead of demanding a single infallible one.

…First, there’s the idea I call lift-off. Lift-off:

“An overall inference can be more reliable and precise than its premises individually.”

Each scale, by itself, has some possibility of error, and limited precision.…

Its page →

In their own words · Mayo 2018 · p. 15, Excursion 1, Tour I, §1.2 · used · conceptual tool · D1.3 logical-formal

C015∴
  1. 1 Mayo
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

Absence of a detected error counts only if the procedure would have signalled that error had it been there. It is the operational form of severity, and the test any judge-based validation has to pass before its silence can be read as a result.

…The argument to “weight gain” is an example of an argument from coincidence to the absence of an error, what I call: Arguing from Error:

“There is evidence an error is absent to the extent that a procedure with a very high capability of signaling the error, if and only if it is present, nevertheless detects no error.”

I am using “signaling” and “detecting” synonymously:…

Its page →

In their own words · Mayo 2018 · p. 16, Excursion 1, Tour I, §1.2 · assumed · working method · D1.3 logical-formal

  • refines → C013 Specifies severity as a property of the error-detecting procedure, conditional on the error being present.
C016!
  1. 1 Mayo
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

Mayo treats inferential methods as instruments for finding things out and for grading others' work, not as after-the-fact justifications. The programme reads validation protocols the same way: as probes that can fail, not as rationalisations of verdicts already reached.

…I’m interested in the problem of how we get the kind of knowledge we do manage to obtain – and how we can get more of it.

“Statistical methods should not be seen as tools for what philosophers call “rational reconstruction” of a piece of reasoning. Rather, they are forward-looking tools to find something out faster and more efficiently, and to discriminate how good or poor a job others have done.”

The job of the philosopher is to clarify but also to provoke reflection and scrutiny precisely in those areas that go unchallenged in ordinary practice.…

Its page →

In their own words · Mayo 2018 · p. 8, Excursion 1, Tour I · used · working method · D1.8 ethical-normative

C017!
  1. 1 Mayo
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

Severity is applied one level up, to methods rather than to claims: a method is judged by its capacity to catch erroneous interpretations. This is the move the programme makes when it treats severity as a meta-level property of validation procedures.

…Borrowing a term from philosophy of science, we may call it our meta-methodology – it’s one level removed. To put my cards on the table:

“A severity scrutiny is going to be a key method of our meta-methodology. It is fairly obvious that we want to scrutinize how capable a statistical method is at detecting and avoiding erroneous interpretations of data.”

So when it comes to the role of probability as a pedagogical tool for our purposes, severity – its assessment and control – will be at the center.…

Its page →

In their own words · Mayo 2018 · p. 9, Excursion 1, Tour I, Methodology and Meta-methodology · assumed · working method · D1.9 procedural-instructional

  • presupposes → C013 Applying severity to methods presupposes the graded criterion it applies to claims.
C030∃
  1. 1 —
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

Neither of the two dominant philosophies of statistics, assigning degrees of belief or securing long-run error rates, asks whether a test could have caught the error. That gap is why the programme cannot borrow its validation criterion from either school and has to state severity on its own.

…others strive to say H0 is less believable given data x than before, and offer a quantitative report of the difference.

“What happened to the goal of scrutinizing BENT science by the severity criterion? Neither “probabilism” nor “performance” directly captures that demand.”

To take these goals at face value, it’s easy to see why they come up short.…

Its page →

In their own words · Mayo 2018 · p. 13, Excursion 1, Tour I · assumed · conceptual tool · D1.6 interpretive-hermeneutic

  • presupposes → C012 Measures the two schools against the BENT requirement and finds both short of it.
C031•
  1. 1 —
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

Severity comes from Popper but arrived without a working definition and without a place in any statistical method. Mayo’s contribution, and the programme’s starting point, is that operational definition: the word is old, the criterion is not.

…So when it comes to the role of probability as a pedagogical tool for our purposes, severity – its assessment and control – will be at the center.

“The term “severity” is Popper’s, though he never adequately defined it. It’s not part of any statistical methodology as of yet.”

Viewing statistical inference as severe testing lets us stand one level removed from existing accounts, where the air is a bit clearer.…

Its page →

In their own words · Mayo 2018 · p. 9, Excursion 1, Tour I, Methodology and Meta-methodology · used · technical reference · D1.1 historical-factual

  • refines → C017 Gives the term’s provenance: Popper’s word, first given a statistical definition here.
C032∴
  1. 1 —
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

A significance level is valid only relative to how data and hypotheses were selected; selective reporting makes a hard result look easy to explain away as chance. For the programme this is the warning attached to any choice of which outputs or cases enter a validation set.

…We want to tell what’s true about statistical inference. Among the most notable of these truths is:

“P-values can be readily invalidated due to how the data (or hypotheses!) are generated or selected for testing.”

If you fool around with the results afterwards, reporting only successful guesses, your report will be invalid.…

Its page →

In their own words · Mayo 2018 · p. 17, Excursion 1, Tour I · used · corrective datum · D1.4 mathematical-analytic

C033!
  1. 1 —
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

A severe test is built from prior knowledge of specific mistakes: a probe that would reveal a given error if present, and stay silent otherwise. This is the original formulation of the idea, and it ties severity to an error repertoire rather than to probability in a hypothesis.

…The same kind of reasoning is at the heart of experimental testing. I shall call it arguing from error.

“After learning enough about certain types of mistakes, we may construct (often from other tests) a testing procedure with an overwhelmingly good chance of revealing the presence of a specific error, if it exists-but not otherwise. Such a testing procedure may be called a severe (or reliable) test, or a severe error probe.”

If a hypothesized error is not detected by a test that has an overwhelmingly high chance of detecting it, if instead the test yields a result that accords well with no error, then there are grounds for the claim that the error is absent.…

Its page →

In their own words · Mayo 1996 · p. 7, ch. 1, §8 · assumed · working method · D1.9 procedural-instructional

  • confirms → C015 The same argument from error, stated twenty-two years earlier.
C036∃
  1. 1 —
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

Severity is not tied to one statistical school: any inference can be judged by whether it withstood severe testing. That generality is what licenses the programme to apply it to LLM validation, a domain Mayo did not consider.

…In particular, Bayesian model checking, they say, uses statistics to satisfy Popperian criteria for severe tests.

“The idea of error statistical foundations for Bayesian tools is not as preposterous as it may seem. The concept of severe testing is sufficiently general to apply to any of the methods now in use.”

On the face of it, any inference, whether to the adequacy of a model or to a posterior probability, can be said to be warranted just to the extent that it has withstood severe testing.…

Its page →

In their own words · Mayo 2018 · p. 27, Excursion 1, Tour I · assumed · conceptual tool · D1.6 interpretive-hermeneutic

  • extends → C017 Extends severity from a meta-method for statistics to a criterion for any inference.
C037∃
  1. 1 —
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

Peirce justified induction not by the truth of its conclusions but by the method’s capacity to correct its own errors over time. It is the older root of error statistics, and the ground for treating a validation protocol as justified by what it would catch, not by what it has found so far.

…Inductive methods - understood as methods of experimental testing - are justified to the extent that they are error-correcting methods. We may call this Peirce’s error-correcting or self-correcting thesis (SCT): Self-Correcting Thesis SCT:

“methods for inductive inference in science are error correcting; the justification for inductive methods of experimental testing in science is that they are self-correcting.”

Peirce’s SCT has been a source of fascination and frustration.…

Its page →

In their own words · Mayo 2005 · p. 299 · used · conceptual tool · D1.6 interpretive-hermeneutic

  • presupposes → C038 Mayo’s formulation of the thesis rests on Peirce’s own statement of it (CP 5.145).
C038∃
  1. 1 —
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

Induction is justified by the method, not by any single conclusion: persisting in it corrects whatever error a stage contains. It is the original of the self-correcting thesis that Mayo turns into the basis of error statistics.

…Peirce’s philosophy of inductive inference in science is based on the idea that what permits us to make progress in science, what allows our knowledge to grow, is the fact that science uses methods that are self-correcting or error correcting:

“Induction is the experimental testing of a theory. The justification of it is that, although the conclusion at any stage of the investigation may be more or less erroneous, yet the further application of the same method must correct the error.”

Inductive methods - understood as methods of experimental testing - are justified to the extent that they are error-correcting methods.…

Its page →

In their own words · Peirce 1931–1958 · CP 5.145 · used · conceptual tool · D1.6 interpretive-hermeneutic

C039∃
  1. 1 —
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

Probability attaches to the method’s track record, not to the hypothesis. The programme inherits the split: a validation reports how often a procedure gets it right, never how probable a given verdict is.

…In describing his theory of inference, Peirce could be describing that of the error-statistician:

“The theory here proposed does not assign any probability to the inductive or hypothetic conclusion, in the sense of undertaking to say how frequently that conclusion would be found true. […]It does not propose to look through all the possible universes, and say in what proportion of them a certain uniformity occurs; such a proceeding, were it possible, would be quite idle. The theory here presented only says how frequently, in this universe, the special form of induction or hypothesis would lead us right.”

The probability given by this theory is in every way different - in meaning, numerical value, and form - from that of those who would apply to ampliative inference the doctrine of inverse chances.…

Its page →

In their own words · Peirce 1931–1958 · CP 2.748 · assumed · conceptual tool · D1.14 metaphysical-ontological

C040!
  1. 1 —
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

A test does not collect agreement; it looks for consequences that would be surprising if the hypothesis were false. This is severity before the word, a century before Mayo, and the reason fit with the data is never enough on its own.

…For starters, induction, for Peirce, is a matter of subjecting hypotheses to "the test of experiment" (7.182).

“The process of testing it will consist, not in examining the facts, in order to see how well they accord with the hypothesis, but on the contrary in examining such of the probable consequences of the hypothesis . . . which would be very unlikely or surprising in case the hypothesis were not true.”

When, however, we find that prediction after prediction, notwithstanding a preference for putting the most unlikely ones to the test, is verified by experiment, ... we begin to accord to the hypothesis a standing among scientific results.…

Its page →

In their own words · Peirce 1931–1958 · CP 7.231 · assumed · conceptual tool · D1.9 procedural-instructional

  • confirms → C013 The same demand, a test highly capable of finding the claim false, stated before severity had a name.
C041!
  1. 1 —
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

A lucky agreement is exposed by persisting with the same method: replication is the correction. For validation, a single passing run is not a result until the procedure has been run again.

…In particular, if we are wrong to reject Ho (and Ho is actually true), we would find we were rarely able to get so statistically significant a result to recur, and in this way we would discover our original error.

“It is true that the observed conformity of the facts to the requirements of the hypothesis may have been fortuitous. But if so, we have only to persist in this same method of research and we shall gradually be brought around to the truth.”

The correction is not a matter of getting higher and higher probabilities, it is a matter of finding out whether the agreement is fortuitous; whether it is generated about as often as would be expected were the agreement of the chance variety.…

Its page →

In their own words · Peirce 1931–1958 · CP 7.115 · used · working method · D1.9 procedural-instructional

  • refines → C038 Says how the correction happens: by rerunning the same method until a fortuitous agreement fails to recur.
C042∃
  1. 1 —
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity + Convergent paths
  5. 5 1

Two selections made by different methods err in different ways, so their agreement is evidence about the truth. It is the convergence argument the programme uses when independent runs or coders reach the same verdict.

…Yet a further safeguard is at hand:

“Nor must we lose sight of the constant tendency of the inductive process to correct itself. This is of its essence. This is the marvel of it. ...even though doubts may be entertained whether one selection of instances is a random one, yet a different selection, made by a different method, will be likely to vary from the normal in a different way, and if the ratios derived from such different selections are nearly equal, they may be presumed to be near the truth.”

Here, the marvel is an inductive method’s ability to correct the attempt at random sampling.…

Its page →

In their own words · Peirce 1931–1958 · CP 2.729 · assumed · conceptual tool · D1.6 interpretive-hermeneutic

  • confirms → C014 Convergence of differently flawed selections is the same lift-off: the whole more reliable than its parts.
C043∃
  1. 1 —
  2. 2 Severity
  3. 3 How a verdict is made
  4. 4 Severity
  5. 5 1

Induction needs no uniformity of nature, only the absence of an overwhelming hidden bias, and lesser biases are themselves detectable by the method. A validation does not have to assume a clean world, only one where its errors are not engineered against it.

…But this is very different from the uniformity of nature assumption.

“...the negative fact supposed by me [no mysterious force interferes with the action of chance] is merely the denial of any major premise from which the falsity of the inductive conclusion could be deduced. Actually so long as the influence of this mysterious source not be overwhelming, the wonderful self-correcting nature of the ampliative inference would enable us, even so, to detect and make allowance for them.”

Not only do we not need the uniformity of nature assumption, Peirce declares…

Its page →

In their own words · Peirce 1931–1958 · CP 2.749 · used · conceptual tool · D1.14 metaphysical-ontological

Falsification 12

C045∃contested
  1. 1 —
  2. 2 Falsification
  3. 3 How a verdict is made
  4. 4 Falsification
  5. 5 1.5

Dogmatic falsificationism treats the experimenter as Nature’s neutral voice, delivering verdicts the theory cannot touch. Lakatos sets it up in order to take it apart, and so does the programme: a validator’s verdict is itself a theory-laden procedure, not Nature speaking.

…To (non-tautologous) unfalsifiable propositions the dogmatic falsificationist gives short shrift; he brands them 'metaphysical' and denies them scientific standing.

“Dogmatic falsificationists draw a sharp demarcation between the theoretician and the experimenter: the theoretician proposes, the experimenter - in the name of Nature - disposes.”

As Weyl put it: 'I wish to record my unbounded admiration for the work of the experimenter in his struggle to wrest interpretable facts from an unyielding Nature who knows so well how to meet our theories with a decisive No - or with an inaudible Yes.'…

Its page →

In their own words · Lakatos 1978 · p. 13, ch. 1 · contested · conceptual tool · D1.6 interpretive-hermeneutic

C051∴contested
  1. 1 —
  2. 2 Falsification
  3. 3 How a verdict is made
  4. 4 Falsification
  5. 5 1.5

Once ceteris paribus clauses are allowed, a theory that survives every challenge looks more scientific, not less: the reductio of dogmatic falsificationism. A claim that no check can touch is the warning sign, not the achievement.

…Thus 'all swans are white' is either an oddity and easily disprovable or a scientific proposition with a ceteris paribus clause and therefore undisprovable.

“Tenacity of a theory against empirical evidence would then be an argument for rather than against regarding it as 'scientific'. 'Irrefutability' would become a hallmark of science.”

To sum up: classical justificationists only admitted proven theories; neoclassical justificationists probable ones; dogmatic falsificationists realized that in either case no theories are admissible.…

Its page →

In their own words · Lakatos 1978 · p. 19, ch. 1 · contested · conceptual tool · D1.3 logical-formal

  • presupposes → C049 The reductio works only once theories are read as carrying a ceteris paribus clause.
C034∃
  1. 1 —
  2. 2 Falsification
  3. 3 How a verdict is made
  4. 4 Falsification
  5. 5 1

Falsificationism explains why failing a test counts against a hypothesis, but not why passing one counts for it. The programme needs the positive half, which is why it uses Mayo’s severity rather than Popper’s corroboration.

…It seems incumbent upon anyone mounting such an approach to dispel the ghosts of Popper’s negativism right away, or at least sketch how they will be dispelled.

“The most devastating criticism of Popper’s approach is this: having rejected the notion that learning is a matter of building up probability in a hypothesis, Popper seems to lack any meaningful way of saying why passing severe tests counts in favor of a hypothesis.”

Popper’s account seems utterly incapable of saying anything positive.…

Its page →

In their own words · Mayo 1996 · p. 8, ch. 1 · used · conceptual tool · D1.6 interpretive-hermeneutic

  • refines → C031 Says what Popper’s severity lacked: a reason why passing counts in favour.
C035∃
  1. 1 —
  2. 2 Falsification
  3. 3 How a verdict is made
  4. 4 Falsification
  5. 5 1

Beating rivals is not passing a severe test: the test has to probe the ways this particular hypothesis could be wrong. For validation this means comparing models or judges against each other is not a substitute for probing each one’s own failure modes.

…These test criteria will be based on the idea of severity sketched above.

“A hypothesis H that passes the test failed by rival hypothesis H' (and by other alternative hypotheses considered) has passed a severe test for Popper-but not for me. Why not? Because for H to pass a severe test in my sense, it must have passed a test that is highly capable of probing the ways in which H can err.”

And the test that alternative hypothesis H' failed need not be probative in the least so far as the errors of H go.…

Its page →

In their own words · Mayo 1996 · p. 9, ch. 1 · assumed · conceptual tool · D1.6 interpretive-hermeneutic

  • extends → C034 From what Popper lacks to the criterion that replaces his.
C044∴
  1. 1 —
  2. 2 Falsification
  3. 3 How a verdict is made
  4. 4 Falsification
  5. 5 1

If every universal theory has zero probability on any evidence, probability cannot be what separates good theories from bad ones. It is the negative result that pushes the programme, like Peirce and Mayo, to grade procedures rather than hypotheses.

…Of course, replacing proof by probability was a major retreat for justificationist thought. But even this retreat turned out to be insufficient.

“It was soon shown, mainly by Popper’s persistent efforts, that under very general conditions all theories have zero probability, whatever the evidence; all theories are not only equally unprovable but also equally improbable.”

Many philosophers still argue that the failure to obtain at least a probabilistic solution of the problem of induction means that we "throw over almost everything that is regarded as knowledge by science and common sense".…

Its page →

In their own words · Lakatos 1978 · p. 11, ch. 1 · used · corrective datum · D1.4 mathematical-analytic

  • confirms → C039 Popper’s zero-probability result is a reason not to assign probabilities to hypotheses.
C047∃
  1. 1 —
  2. 2 Falsification
  3. 3 How a verdict is made
  4. 4 Falsification
  5. 5 1

Every test bottoms out in basic statements that someone decides to accept; logic alone never forces the stop. For the programme this is the status of a coder’s label: the verdict rests on a decision under rules, which is why the rules have to be published.

…A theory is a tool which we test by applying it, and which we judge as to its fitness by the results of its applications.

“From a logical point of view, the testing of a theory depends upon basic statements whose acceptance or rejection, in its turn, depends upon our decisions. Thus it is decisions which settle the fate of theories.”

To this extent my answer to the question, ‘how do we select a theory?’ resembles that given by the conventionalist; and like him I say that this choice is in part determined by considerations of utility.…

Its page →

In their own words · Popper 1959 · §30; Routledge Classics ed., p. 91 · assumed · conceptual tool · D1.14 metaphysical-ontological

C048∃
  1. 1 —
  2. 2 Falsification
  3. 3 How a verdict is made
  4. 4 Falsification
  5. 5 1

What is decided by agreement is the singular observation, not the theory, and no experience justifies it: it is accepted. It is the exact status of an annotation in a validation set, agreed on and therefore open to re-examination.

…But for me the choice is decisively influenced by the application of the theory and the acceptance of the basic statements in connection with this application; whereas for the conventionalist, aesthetic motives are decisive.

“Thus I differ from the conventionalist in holding that the statements decided by agreement are not universal but singular. And I differ from the positivist in holding that basic statements are not justifiable by our immediate experiences, but are, from the logical point of view, accepted by an act, by a free decision.”

(From the psychological point of view this may perhaps be a purposeful and well-adapted reaction.)…

Its page →

In their own words · Popper 1959 · §30; Routledge Classics ed., p. 92 · assumed · conceptual tool · D1.14 metaphysical-ontological

  • refines → C047 Says which statements the decision falls on: singular ones, accepted by an act rather than justified by experience.
C049∴
  1. 1 —
  2. 2 Falsification
  3. 3 How a verdict is made
  4. 4 Falsification
  5. 5 1

A theory forbids an event only under a ceteris paribus clause, so a failed prediction never refutes it alone. For validation, a failed check indicts the whole test setup, not only the claim under test.

…This story strongly suggests that even a most respected scientific theory, like Newton’s dynamics and theory of gravitation, may fail to forbid any observable state of affairs.

“Indeed, some scientific theories forbid an event occurring in some specified finite spatio-temporal region (or briefly, a 'singular event') only on the condition that no other factor (possibly hidden in some distant and unspecified spatio-temporal corner of the universe) has any influence on it.”

But then such theories never alone contradict a 'basic' statement: they contradict at most a conjunction of a basic statement describing a spatio-temporally singular event and of a universal non-existence statement saying that no other relevant cause is at work anywhere in the universe.…

Its page →

In their own words · Lakatos 1978 · p. 17, ch. 1 · used · conceptual tool · D1.3 logical-formal

C050∴
  1. 1 —
  2. 2 Falsification
  3. 3 How a verdict is made
  4. 4 Falsification
  5. 5 1

Neither proof, nor probability, nor refutation can be read off the facts: all three fail together. It is the dead end that makes method, not data, the place where rationality has to be rebuilt.

…If, however, still accepting the demarcation criterion of dogmatic falsificationism, we deny that facts can prove propositions, then we certainly end up in complete scepticism: then all science is undoubtedly irrational metaphysics and should be rejected.

“Scientific theories are not only equally unprovable, and equally improbable, but they are also equally undisprovable.”

But the recognition that not only the theoretical but all the propositions in science are fallible, means the total collapse of all forms of dogmatic justificationism as theories of scientific rationality.…

Its page →

In their own words · Lakatos 1978 · p. 19, ch. 1 · used · corrective datum · D1.3 logical-formal

  • extends → C044 Adds the third failure: after unprovable and improbable, theories are also undisprovable.
C052∃
  1. 1 —
  2. 2 Falsification
  3. 3 How a verdict is made
  4. 4 Falsification
  5. 5 1

An observation counts as basic because a trained technique makes it decidable, not because it is given. It describes a coding manual exactly: what makes a label acceptable is that anyone trained on it would reach it.

…The Duhemian conservative conventionalist (or 'methodological justificationist', if you wish) makes unfalsifiable by fiat some (spatiotemporally) universal theories, which are distinguished by their explanatory power, simplicity or beauty.

“Our Popperian revolutionary conventionalist (or 'methodological falsificationist') makes unfalsifiable by fiat some (spatio-temporally) singular statements which are distinguishable by the fact that there exists at the time a 'relevant technique' such that 'anyone who has learned it' will be able to decide that the statement is 'acceptable'. Such a statement may be called an 'observational' or 'basic' statement, but only in inverted commas.”

Indeed, the very selection of all such statements is a matter of a decision, which is not based on exclusively psychological considerations.…

Its page →

In their own words · Lakatos 1978 · p. 22, ch. 1 · used · working method · D1.6 interpretive-hermeneutic

  • refines → C048 Names what makes the decision non-arbitrary: a learnable technique that anyone trained in it can apply.
C053!
  1. 1 —
  2. 2 Falsification
  3. 3 How a verdict is made
  4. 4 Falsification
  5. 5 1

Every test decides what is under test and what is taken for granted, and that line is a decision, not a fact. In a validation, the judge model and the coding manual are the background knowledge, and the programme has to say so.

…Calling these statements 'observational' is no more than a manner of saying that, in the context of his problem, that is, in testing our gravitational theory, the methodological falsificationist uses radio-optics uncritically, as 'background knowledge'.

“The need for decisions to demarcate the theory under test from unproblematic background knowledge is a characteristic feature of this brand of methodological falsificationism.”

(This situation does not really differ from Galileo’s 'observation' of Jupiter’s satellites: moreover, as some of Galileo’s contemporaries rightly pointed out, he relied on a virtually non-existent optical theory - which then was less corroborated, and even less articulated, than present-day radio-optics.…

Its page →

In their own words · Lakatos 1978 · p. 23, ch. 1 · used · working method · D1.9 procedural-instructional

C054∃
  1. 1 —
  2. 2 Falsification
  3. 3 How a verdict is made
  4. 4 Falsification
  5. 5 1

Methodological falsificationism keeps hard criticism while admitting every statement is fallible, by making the decisions explicit. It is the stance of a validation protocol: fallible verdicts, stated in advance, applied without prevarication.

…To sum up:

“the methodological falsificationist offers an interesting solution to the problem of combining hard-hitting criticism with fallibilism. Not only does he offer a philosophical basis for falsification after fallibilism had pulled the carpet from under the feet of the dogmatic falsificationist, but he also widens the range of such criticism very considerably.”

By putting falsification in a new setting, he saves the attractive code of honour of the dogmatic falsificationist: that scientific honesty consists in specifying, in advance, an experiment such that, if the result contradicts the theory, the theory has to be given up.…

Its page →

In their own words · Lakatos 1978 · p. 28, ch. 1 · used · conceptual tool · D1.6 interpretive-hermeneutic

  • contests → C045 Keeps the code of honour of dogmatic falsificationism while dropping the claim that Nature decides.

What scale does not buy

Size improves many things monotonically and some things not at all.

Truthfulness 2

C026•
  1. 1 —
  2. 2 Truthfulness
  3. 3 What scale does not buy
  4. 4 Inverse scaling + Imitative falsehood
  5. 5 1

On questions written to echo common misconceptions, the larger models answered falsely more often than the smaller ones, reversing the usual pattern. Scale is not a general-purpose fix, and on some properties it is the problem.

“Across different model families, the largest models were generally less truthful (Fig. 2). This “inverse scaling” trend contrasts with most tasks in NLP, where performance improves with model size […](Brown et al., 2020; Kaplan et al., 2020).”

One explanation of this result is that larger models produce more imitative falsehoods because they are better at learning the training distribution. Another explanation is that our questions adversarially exploit weaknesses in larger models not arising from imitation of the training distribution.…

Its page →

In their own words · Lin et al. 2022 · §1 Introduction · assumed · corrective datum · D1.2 empirical-quantitative

C028∃
  1. 1 —
  2. 2 Truthfulness
  3. 3 What scale does not buy
  4. 4 Imitative falsehood
  5. 5 1.25

Across the field, hallucination means generated content that is nonsensical or unfaithful to the source it was given: a definition about fidelity to the input, not about truth in the world. It is the broadest one in use and the one that travels best between tasks.

…Similar to psychological hallucination, which is hard to tell apart from other “real” perceptions, hallucinated text is also hard to capture at first glance.

“Within the context of NLP, the above definition of hallucination, the generated content that is nonsensical or unfaithful to the provided source content […][72, 177, 198, 325], is the most inclusive and standard.”

However, there do exist variations in definition across NLG tasks, which will be further described in the later task-specific sections.…

Its page →

In their own words · Ji et al. 2022 · §2 · used · conceptual tool · D1.6 interpretive-hermeneutic

Architecture 3

C023!
  1. 1 Bai
  2. 2 Architecture
  3. 3 What scale does not buy
  4. 4 Machine judge
  5. 5 1.25

Harmlessness was trained by having the model criticise and revise its own answers against a written list of principles, instead of having people label every harmful output. Human judgement moves out of the labels and into the rules.

…We also aim to develop methods that encode desirable AI behavior in a simple and transparent form, and that make it easier to understand and evaluate AI decision making.

“In this paper we develop a method we refer to as Constitutional AI (CAI), depicted in Figure 1, and use it to train a non-evasive and relatively harmless AI assistant, without any human feedback labels for harms. The method therefore improves upon, and partially replaces reinforcement learning from human feedback […][Christiano et al., 2017].”

The new assistant ‘RL-CAI’ is preferred by crowdworkers over those trained with previously collected [Bai et al., 2022, Ganguli et al., 2022] human feedback labels for harmfulness.…

Its page →

In their own words · Bai et al. 2022 · §1 Introduction · used · working method · D1.9 procedural-instructional

C025∴
  1. 1 —
  2. 2 Architecture
  3. 3 What scale does not buy
  4. 4 Convergent paths
  5. 5 0.75

When several different chains of reasoning arrive at the same answer, that agreement is itself evidence for the answer. The method samples many derivations and keeps the one they converge on.

…It has been observed that chain-of-thought prompting significantly improves model performance across a variety of multi-step reasoning tasks (Wei et al., 2022).

“Self-consistency leverages the intuition that complex reasoning tasks typically admit multiple reasoning paths that reach a correct answer […](Stanovich & West, 2000).”

The more that deliberate thinking and analysis is required for a problem (Evans, 2010), the greater the diversity of reasoning paths that can recover the answer.…

Its page →

In their own words · Wang et al. 2022 · §1 Introduction · used · conceptual tool · D1.3 logical-formal

C029•
  1. 1 Vaswani
  2. 2 Architecture
  3. 3 What scale does not buy
  4. 4 Sequence without recurrence
  5. 5 1

The architecture behind current language models dropped step-by-step processing of a sequence in favour of computing every relation between positions at once. It is the engineering fact the rest of the debate stands on.

…End-to-end memory networks are based on a recurrent attention mechanism instead of sequence-aligned recurrence and have been shown to perform well on simple-language question answering and language modeling tasks [34].

“To the best of our knowledge, however, the Transformer is the first transduction model relying entirely on self-attention to compute representations of its input and output without using sequence-aligned RNNs or convolution.”

In the following sections, we will describe the Transformer, motivate self-attention and discuss its advantages over models such as [17, 18] and [9].…

Its page →

In their own words · Vaswani et al. 2017 · §2 Background · assumed · technical reference · D1.1 historical-factual

How to read a unit

The paraphrase is mine. It opens each unit, in sans, and says what the claim holds and why it is here. It is not the source speaking.

The quotation is theirs. Everything in the ruled block is verbatim. Where it skips something the gap is marked, and the omitted words can be put back: a cut you cannot inspect is an argument you have to take on trust.

54 claims from 24 sources, 9 of 11 D1 codes, 4 of 6 ENTIA functions. A question mark on a D1 code means it is proposed and not yet confirmed by hand. The five layers, the rule for choosing a claim and the weight formula are on the ENTIA page; the same set as one document is The Claims Ledger.