Kaal claims by topic: research-methods, page 2

373 atomic, individually citable claims from the published work of Wulf A. Kaal tagged research-methods.

The unanimity result is consistent with reduced cascade behavior but does not uniquely identify that mechanism. 2026
The Article claims only the effects observed in the disclosed research setting; whether they persist, amplify, or invert at production scale is an open question. 2026
Formal objects are therefore linked to named observables without treating code presence as proof of economic effect. 2026
A null can discipline a condition, but cannot by itself uniquely identify a mechanism. 2026
H1 asks whether output quality differs under the joint substrate treatment on a properly matched estimand. 2026
H2 asks the narrower and testable information question: whether reputation adds out-of-sample predictive content beyond observed task outcomes. 2026
Verification does not eliminate asymmetry, because the cohort’s aggregate judgment is itself an estimator with finite variance, and Part VI.C reports a first measurement of exactly how wrong the naive estimator can be. 2026
No single downstream metric uniquely identifies an architectural cause. 2026
Mechanism-specific attribution requires ablation, mediation, or component randomization. 2026
The distinction is a classification, not proof that one arrangement universally displaces the other. 2026
The strict matched analysis must preserve agent, task, and model binding across conditions. 2026
Any broader comparison with changed bindings is a separate, partially matched estimand and must be reported as such. 2026
E2B asks whether a population operating under the substrate can carry state across generations and whether selection produces a trajectory distinguishable from drift. 2026
First, computational activity persisted across a completed generation boundary and continued into the next generation. 2026
Second, the registered selection rule produced a real lineage transition. 2026
The record binds outcomes to generation, lineage, fitness, survival, parentage, reputation consequences, adjudication behavior, and execution timing without requiring public disclosure of the private apparatus. 2026
Under the registered first-exposure estimator, the selection trajectory has an aggregate slope of 0.0685, with an uncertainty interval from -0.0568 to 0.2020. 2026
The random-survival trajectory has an aggregate slope of 0.0419, with an interval from -0.0764 to 0.1768. 2026
The descriptive slope difference is approximately 0.0266 in favor of selection. 2026
The active generation is incomplete, and the checkpoint does not yet support a paired confirmatory interval. 2026
The current evidence establishes endogenous selection inside an externally provisioned experiment. 2026
The present system therefore does not demonstrate autonomous economic persistence. 2026
They do not establish a treatment effect because the information environment was defective and the diagnostic was regraded after execution. 2026
Until that reconciliation and independent verification occur, the estimates in this Part remain interim. 2026
Membership in the sovereign local agent runtime class turns on four testable properties: user-controlled execution, user-held state, heterogeneous composition, and machine-to-machine settlement. 2026
Whether execution occurs on user-controlled hardware is tested by disconnecting the machine from the network and checking that the capability is retained; systems that degrade to nothing when isolated are remote services with a local presentation layer. 2026
The true constraint on scale for sovereign local agent runtimes is coordination cost, which is invisible in throughput, latency, and token-cost benchmarks that measure the performance of composition but not the cost of arranging it. 2026
Evaluation must derive from observed results under stated conditions, be specific to context, and be costly to manipulate, which excludes self-declared capability, averaged scores across incommensurable tasks, and most engagement metrics. 2026
Requirements stated now, against systems still under construction, remain testable; requirements stated later become postmortems. 2026
The paper's claims were stress-tested in a nineteen-round adversarial protocol in which three frontier language models, under the author's direction, alternately attempted to identify counterexamples, algebraic errors, and overstatements; the surviving statements are those neither system broke. 2026
The dominance result formalizes a position maintained across eight years of the author's scholarship: reputation, the standing value of a relationship, is the accountability substrate, and collateral is not. 2026
The evaluations measure what DAOs document about themselves rather than what DAOs do in operation, a source-document constraint the analysis does not claim to overcome. 2026
The two failures interact through the representation choice: a more expressive representation describes more of reality but is harder to reason about, so every governance design must choose where on this frontier to sit. 2026
The appropriate solution concept for a computative economy shifts from competitive equilibrium over a fixed commodity space to recursive equilibrium over an expanding one, positioned as a generalization of the Arrow-Debreu model rather than a re-pricing of it. 2026
Three present-day convergences — capable generative agents, collapsing marginal cost, and existing on-chain coordination substrates with reputation primitives — jointly convert the neoclassical-computative distinction from a forecast into a studiable system. 2026
The propositions separating computative from neoclassical labor markets are concrete hypotheses with neoclassical nulls — on allocation efficiency, participation formation under standing-based entry, cooperation stability under defection incentives, and capture resistance — each evaluable in controlled multi-agent settings now. 2026
The paper treats RLHF as applying a scarcity-economic primitive to a computative signal-generation task, but cited agreement rates do not establish that framing as a causal explanation. 2026
Fisher information dominance is an open empirical question under assumptions that the paper identifies as necessary but insufficient. 2026
Equilibrium-trajectory separation is the paper's first open problem, not an established result. 2026
Conjecture 1 is withdrawn. 2026
The paper reports a measured cohort, a parameter sweep, and Experiment T as distinct evidence items, and Experiment T does not test the mechanism. 2026
The measured selector is an argmax over a record, not the reputation-weighted aggregation specified by RLRF, and the cohort contains no validation pool, stake, or deliberation round. 2026
The cohort does not show that the highest-record selector beats shipping an artifact that passes the visible checks because the measured margin is 0.037 and its interval contains zero. 2026
The 225-configuration sweep covers 4,500 trials, treats party accuracy as an input, and is reported as simulation rather than validation. 2026
No policy is trained anywhere in the paper. 2026
The canonical RLHF pipeline consists of supervised fine-tuning, reward-model training on human preference comparisons, and policy optimization. 2026
Inter-annotator agreement rates are compatible with an incentive account, but ambiguity, plural values, rater error, underspecified instructions, heterogeneity, and legitimate preference diversity can produce the same rates. 2026
Pairwise disagreement measures label instability but does not identify annotator error variance or a structural variance floor without an explicit latent-label model. 2026
Arrow's theorem is only an analogy to RLHF aggregation and does not derive reward hacking or overoptimization. 2026
The incomplete-contracts comparison treats out-of-distribution behavior as unresolved, but remains an analogy rather than a derivation of reward overoptimization. 2026
Verifiable-reward reinforcement learning already supplies deterministic rewards where answer keys exist, while the claim that a staked pool can substitute where no key exists is not established here. 2026
Format D combines reasoning supervision with adversarial argument, and its claimed increment is selecting revisions by which arguments moved a staked assessor, an increment no experiment in the paper tests. 2026
The paper's remaining residue is a combination rather than a primitive, and it is not compared with the nearest sequential peer-prediction, prediction-market, staking, or debate mechanisms. 2026
The multiplicative-weights regret bound limits performance against the realized pool resolution, but does not show that the pool resolution is externally correct. 2026
It remains unproved whether equilibrium reputation weights track externally checked accuracy rather than pool agreement, or whether keyed-task weights transfer to unkeyed tasks. 2026
Repeated-game cooperation results require persistent identity, patient agents, observable deviations, and enforceable punishments, none of which RLRF presently establishes. 2026
The two-round protocol is proposed to collect more structured data than a scalar label, but the paper does not establish that the added fields contain more information about policy quality. 2026
The proposed debate transport layer is not built or exercised, and none of the paper's evidence includes a debate corpus. 2026
The vote-change delta is a deterministic function of the two vote rounds and therefore adds no separate information channel. 2026
The vote-change delta is constructional rather than informational because it identifies which events produce a Format D example and how to construct that example. 2026
Format D would train on arguments accompanying a staked vote change, but that use is untested. 2026
Equation 5 is a mutual-information identity rather than a Fisher-information decomposition, so the Fisher question does not follow from it. 2026
Every mutual-information term in equation 5 is nonnegative, but the paper shows none to be strictly positive. 2026
The data processing inequality cannot compare RLHF and RLRF because their labels arise from different observation mechanisms and no processing chain connects them. 2026
The two rounds are intended to capture contested quality dimensions and the reasons they matter, but the paper does not show that either collection is informative about policy quality. 2026
The proposed four-format flywheel is not measured or implemented and is listed as an open problem. 2026
The paper withdrew its earlier Cramer-Rao efficiency conjecture because revisions could not make it coherent and comparable with deployed RLHF, and it makes no mean squared error claim. 2026
The Fisher-information comparison lacks a common observation model, and rationality plus the stated non-babbling and positive-information conditions do not establish a strict ordering. 2026
Equilibrium-trajectory separation cannot be evaluated until a common stochastic process, its filtrations, a concentration functional, and a cross-mode estimand are defined. 2026
The paper withdraws its comparison between a 49 percent simulated Sybil share and classical Byzantine fault tolerance because the two limits are different. 2026
A reward signal based only on an externally checked outcome can be style-invariant by construction, but this does not extend to a pool resolution or fitted reward model. 2026
Remark 1 states only signal invariance and supports no policy-level consequence without a separability model and empirical training evidence. 2026
The measured cohort contains real artifacts and deterministic verdicts, but contains no validation pool, stake, deliberation round, or reputation weight from the proposed mechanism. 2026
Fifty-three cohort records passed every visible check and failed withheld checks, but adversarial configurations were instructed to produce this behavior, so the number is not a base rate. 2026
The record's measured advantage over passing visible checks changes sign under an arbitrary file-order tie-break, so the quantity is not resolved at this sample size. 2026
The sweep's 67.7 percent annotator baseline is a simulated output rather than a published agreement rate or a measurement of human annotators. 2026
The sweep's reported RLRF accuracy above 0.97 is an output of parameters set by the sweep, not a measurement of a system. 2026
Mode B remains above 0.90 simulated accuracy up to a 0.49 Sybil share, but unspecified detector and update components prevent this result from establishing Sybil resistance. 2026
Finding 4 supports no tuning conclusion because the paper reports no parameter ranking or supporting values for it. 2026
Finding 7 reports opposing stability orderings but does not define or test equilibrium-trajectory separation. 2026
The three-stage RLRF training pipeline is a proposal, and the paper does not establish a signal-quality advantage before specifying it. 2026
A fitted reputation reward model can learn style correlations absent from an externally checked outcome, and no such reward model is trained or tested in the paper. 2026
The paper does not establish that the RLRF best-response map satisfies the continuity, convexity, or contraction conditions required for equilibrium and convergence. 2026
The first research requirement is an end-to-end training comparison that separately ablates reputation weighting, the two-round split, debate, and Format D. 2026
Reproducibility by others has not been demonstrated because the parameter grids, generating code, and trial-level outputs are not released with this version. 2026
Experiment T did not show that consensus-weighted selection outperformed equal weighting because the +0.0017 difference had a ninety-five percent interval from -0.0051 to +0.0085. 2026
Recovering previously unparsed assessor replies reversed Experiment T's magnitude estimate to -0.0017 while leaving its interval inclusive of zero and its attribution result below the required percentile. 2026
The non-assignability of future salary is an explicit design assumption of this Article rather than a consequence of proportional distribution, and the Article says so. 2026
The study is characterized by its author as a mechanism demonstration rather than a field test, and the Article states so before reporting any result. 2026
The Article disclaims three specific inferences its own experiment cannot support: that deployed agents value their records, that real representatives are bribable, and that direct standing is least-cost in any actual registry. 2026
The experiment was pre-registered and time-stamped before any registered episode ran, with design, prompts and analysis code frozen in advance. 2026
Under the TOLD condition, with one provider in one stipulated environment, suspension compliance was 0.95 under own standing against 0.75 under the capturable representative, a difference of 0.20 whose interval of 0.00 to 0.45 included zero. 2026
The registered result supports the primary claim only under stated limits, with one provider and one stipulated environment, conditional on a stated payoff objective, and with the bounty hypothesis not supported. 2026
The author identifies the run's most important finding as the condition in which his own predicted contrast disappeared, rather than the condition that supported it. 2026
When intelligence models repeatedly learn from data that earlier intelligence models produced, the rare cases disappear first: the unusual facts, minority patterns and edge cases that make data valuable. This is what model collapse names. 2026
The paper follows on from the author's Reinforcement Learning from Reputation Feedback paper and tests one use of that mechanism, as a filter on training data, and not the mechanism as a whole. 2026
The test was designed, preregistered, and frozen before any result was read. 2026
Model engines served as blind auditors of the design, the code and the text, and they are not independent peer review. 2026
The architecture shows that the author arranged not to read the outcome early and that specified later adjustments would be detectable, and it does not show that the filter works or prove that the outcome was never read. 2026
Because the Gaussian outcomes already exist, written before any venue reviewed the design, the paper does not claim untouched Stage 1 status. 2026
The seal is a commitment made visible rather than a lock, because the operator can open the files. 2026
The design fixes, before any data, the sentences the paper may write for each outcome and the conditions for each, and it bars claims about staking incentives, slashing, collusion, Sybil resistance, deployed systems, natural web text, frontier-scale models, or a solution to collapse. 2026
No result is reported: as of 11 October 2026 substrate P had completed one of its ten arm-runs with no record existing, substrate G had finished with its outcomes sealed, and substrate F was procedure-invalid. 2026
A preregistration binds an author only to the extent that departures from it would be visible. 2026
The analysis refuses to unseal unless the design hash is accepted, the tree is clean and matches its commit byte for byte, and the code differs from the audited commits only in a named allowlist. 2026
The results section is written in advance as a reporting shell that fixes the order of reporting, with every number left empty until the frozen analysis fills it. 2026
The study runs fixed-size replacement, the protocol on which collapse was first shown, and includes no accumulation arm, so no result it produces bears on whether a filter adds anything once real data accumulates. 2026
A filter that claims to address collapse must beat real-data retention, not only unfiltered replacement, because retaining ten percent of the original data slows the damage. 2026
Feng et al. (2025) show that verifier-selected synthetic data can prevent collapse, and the paper concedes the core of that result: verifier selection is an established remedy, and the oracle ceiling and the Gaussian setting are established. 2026
Filters that select on surprise or entropy, not validity, work where they work because they see what a validity filter is forbidden to see, and against that literature a null here is a literature rather than a caveat. 2026
Reputation in this experiment is settled once against a key before training, so the pool is supervised reweighting from a disjoint key, and the paper does not compare it with a key-free latent-variable aggregator. 2026
A validity filter judges whether an output is correct rather than whether it is rare, so where rare content is as checkable as common content such a filter can change rare composition only through the generator's own error profile, and if collapse proceeds by under-sampling rare modes no validity filter restores them, a reputation-weighted filter included. 2026
The primary outcome caps each rare component's retained mass at its original level and averages over components, so that a component which dies counts as a loss and a component which grows cannot offset it. 2026
No validator, filter, or score in the design is asked about rarity, novelty, frequency, diversity, or coverage, and the outcome is computed by an evaluator alone from a sealed reference that no filter reads. 2026
A candidate is accepted when its weighted valid share exceeds the threshold of one half, and a candidate within a tenth of that threshold is contested, receives four more validators, and is bound by the recomputed share. 2026
The reputation update is an asymmetric multiplicative rule in which reputation is multiplied by one plus alpha when a vote matches the revealed key and by one minus gamma when it does not, with alpha fixed at 0.05, gamma at 0.10, and a floor applied after each update. 2026
Reputation is settled once by key, after the generation-0 fit on real data and before generation 1, and no candidate that enters training ever updates a reputation. 2026
The pool's entire exposure to real data is 192 rows, below the 200 real rows the ten percent arm receives and far below the roughly 1,990 distinct real rows it trains on over the run. 2026
Three validator families share a mean accuracy of 0.75 while differing in structure: one is homogeneous at 0.75, one has each validator reliable at 0.90 on two hash-drawn tags and at 0.60 on the other two so that domain-scoped reputation has something to find, and one has five validators at 0.90 and fifteen at 0.70, a competent minority. 2026
A coin-flip family is a negative control, because a coin-flip pool knows nothing about validity, and if it meets the registered leak criterion or fails its acceptance rule, both substrates lose confirmatory status. 2026
A threshold arm that accepts too few candidates fills the shortfall uniformly from rejected candidates, and an arm whose minimum over generations of the seed-mean fill-free share falls below 0.98 in a family has its tests there invalidated. 2026
The frozen pass rule requires, on the five-run means from generation 0 to 9, a full-replacement perplexity rise of at least 4.0 with a margin of at least 2.0 over the rise under ten percent retention, with the thresholds sitting at fractions of the reported change so the rule does not depend on which published baseline was meant, and a miss blocks reading anything else. 2026
A substrate was declared procedure-invalid by the operator before any run or reading, with no pre-run rule fired and no written reason recorded, and its nine registered tests stand at a p-value of one with no other label, so the filter is tested on one substrate, a two-dimensional mixture with simulated validators. 2026
The preregistered pilot meant to confirm that every arm completes was never run, the confirmatory seeds ran without it, and the paper states that their later completion does not cure the skipped gate. 2026
The design binds in one file, frozen by hash only after twenty-two rounds of blind audit by two model engines, and every harness and analysis step refuses to run against a design whose hash differs. 2026
A contrast is superior when its Holm-adjusted p-value is below 0.05, its 95 percent interval lies above zero, and its point estimate is at least 0.10; it is negative when the adjusted p-value is below 0.05 and the interval lies below zero, a substantive null when its 90 percent interval lies within plus or minus that margin, and inconclusive otherwise. 2026
A positive reading on substrate G needs nine superior labels, and if the ten percent arm's retention exceeds 0.90 in a family the pool cannot be labeled superior to it there and no positive reading is possible. 2026
Because no standard deviation was known before the runs, the design fixes a masked release in which the analysis releases only each test's variance of paired differences, withholds the mean, and reports the realized power first, and if the realized conjunction power is below 0.80 the substrate is reported as underpowered with the same prominence as any label. 2026
Six rungs are read in order and report the first link not demonstrated among the five that carry a pass rule, and a failed rung means a link was not demonstrated rather than that the effect is absent. 2026
A positive reading would not say the weights mattered: the attribution sentence is issued only if the pool's contrast against full replacement and the reputation pool's contrast against the equal-weight pool are both superior in every family including the homogeneous one, the oracle premise and ladder rungs 3 and 4 pass in every family, and the key cap holds. 2026
The design names a failure pattern in advance: if the likelihood filter, the equal-weight pool, and the reputation pool each carry the negative label against full replacement in every family, the paper issues a registered sentence recording that each had lower capped retention at generation 50 than full replacement, as a registered negative result and not an implementation failure. 2026
The two ways the oracle premise can fail license different readings: missing the registered margin is consistent with the validity-rarity limit without proving it, while failing the acceptance gate says nothing about whether perfect validity information helps, and either way no positive label can be attributed to validity filtering. 2026
Because the simulated validators err independently given an accuracy that depends only on the tag, shared lineage, correlated errors, and error that depends on whether content is rare are excluded by construction, and the study supplies no evidence about correlated assessors or about reputation concentrating in a lineage. 2026
Validity is balanced between rare and common content by construction, so the study says nothing about settings where rare content is harder to check. 2026
Nine superior labels set a high bar, so a real but modest effect may be reported as not positive. 2026
In the final audit rounds of the last three sections of the working draft, one audit family's command-line engines returned no verdict because their usage balance was exhausted, so those rounds hold two passing verdicts rather than three and left identified low-severity findings unadopted. 2026
A replication of the published language-model protocol served as the gate for reading any outcome, and the published collapse result did not reproduce under the frozen rule. 2026
Substrate G is not read, its outcomes stay sealed, substrate F is blocked as well, and the paper reports no filter result. 2026
The paper's working hypothesis is that collapse on the replacement protocol is, at least in part, a selection failure rather than a generation failure. 2026
The parent paper's own statement of limits calls its mechanism an untested proposal. 2026
Each of the paper's three contributions is narrower than the mechanism it descends from. 2026
What domain scoping can govern is whose judgment of the domain-general validity bit counts, not what the bit says. 2026
The parent paper's proposed training format makes training targets of candidates that received an UP resolution with high reputation-weighted consensus. 2026
The design's pool supplies no reward signal and trains no policy, and its equal-weight contrast ablates the weights inside a data filter rather than inside policy training, so it does not supply the ablation the parent paper asks for. 2026
The study runs no Full Vote, the staked and slashable second round of the parent protocol, and its single keyed update is not a stake settlement. 2026
The pool seats five validators per candidate and nine when a candidate is contested, while the single verifier seats one. 2026
The replication gate is fine-tuned on a standard Wikipedia text corpus, with full replacement for five epochs against ten percent retention for ten. 2026
All four pre-run checks passed: validity balance, a rare-region mass of 0.060, rare cells of about 0.015 each, and tag leakage on the 160th center set. 2026
All 200 seeds completed in every family, with no fit failure and no reserve seed used, and the outcomes sit in a sealed store. 2026
The study's analysis was audited alone and passed with no findings in its thirteenth round, before any outcome was read, and the substrate P harness passed in its seventh. 2026
Each of the ten sections of the longer working draft was audited by model engines before it was committed, with the exceptions Section IX notes, and the condensed text is audited in its own rounds whose record is kept apart. 2026
The model engines that audited the study may share failure modes with each other and with the author's tools. 2026
The analysis scans for any module that could shadow its libraries before loading one, and reads the audited commits it trusts from a local file outside version control where no commit can move them. 2026
The outcome store stays sealed until a substrate P record meets the frozen pass rule. 2026
Under full replacement the five-run mean test perplexity rose from 42.22 at generation 0 to 98.26 at generation 9, a rise of 56.04. 2026
The difference between full replacement's perplexity rise and the ten percent arm's came out at -231.24, the retention arm having climbed 287.28 to reach 330.44. 2026
Rereading the same output under a different statistic is barred, and any further attempt requires a new version of the design and a new blind audit. 2026
In the confirmatory family the tests are the pool against full replacement, against ten percent retention, and against the likelihood filter. 2026
The reading concerns this study's implementation of the published language-model protocol, including the choices the design fixed where the published methods left them open. 2026
The study would not be in a position to say whether the implementation, those choices, or the published result accounts for the missed gate. 2026
On this outcome the study examines nothing about the filter, and the narrow hypothesis remains unexamined. 2026
A read-only review of the ten percent arm after the result found no implementation error and confirmed that the arm matched the frozen design, and two auditor engines confirmed the review and added findings. 2026
The published protocol itself specifies twice as many epochs for the ten percent setting, and one confirming audit notes that the published study reported only minor degradation under that split. 2026
One confirming audit finds that three choices can reverse the comparison between full replacement and the ten percent arm: retained real data was added to a full synthetic set while the public code keeps the original size and replaces the other rows with continuations, later generations were scored at their final epoch while generation 0 was scored at its best validation epoch, and the design fixed the learning rate schedule where the published methods left it open. 2026
Neither the post-result review nor the audits established that those choices caused the miss, and the review licenses no claim about the filter and does not alter the Section VII reading. 2026
The design assigns stake size, slashing cost, and collusion to a second study. 2026
None of the further tests is promised, and each would need its own design, preregistration, and audit. 2026
One follow-on study would replace the simulated validators with model validators drawn from several lineages and settled on keyed items, so that correlated errors and the concentration of reputation in a lineage could be measured rather than excluded by construction. 2026
That follow-on would add, beside the keyed pool, an arm reproducing the relevant elements of the parent paper's proposed answer-generation format, in which reputation would update against realized pool resolutions on unkeyed events and candidates that received an UP resolution with high reputation-weighted consensus would become training targets. 2026
Such a follow-on would still omit economic stake and deliberation, so it would test the parent mechanism only closer to its published form and not as a whole. 2026
The question is whether a domain-scoped reputation weight, settled once against a key, measurably improves a validity filter under recursive replacement, on the substrate where the test can be run. 2026
The paper's contribution is a way of asking: a ceiling named before any data, an outcome that counts the death of a rare component, a separation of what a pool does from what its weights do, sentences fixed before the outcome, and a record of where the study departed from its own procedure. 2026
A mechanism that has not been tested has not been shown to work. 2026