Pith. sign in

REVIEW 5 major objections 4 minor 18 references

Optimizing Decentralized Online Learning for Supervised Regression and Classification Problems

T0 review · 5 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A single calibrated parameter set—$p=3$ for regression, $p=5$ for classification, $\alpha=0.1$—keeps decentralized network loss near its minimum.

desk verdict Useful parameter calibration for a decentralized inference protocol, but the headline classification default (p=5) is a boundary choice that its own refined grid doesn't support. read the letter →

arxiv 2501.16519 v1 pith:2NNXETXT submitted 2025-01-27 cs.LG cs.DCcs.MA

J. M. Diederik Kruijssen (1) , Renata Valieva (1) , Steven N. Longmore (1 , 2) ((1) Allora Foundation , (2) Liverpool John Moores University) This is my paper · ORCID
classification cs.LGcs.DCcs.MA
keywords decentralizedlearninginferencesynthesisparameteroptimizationregressionclassificationexponentialmovingaveragerewarddistributionnetworkloss
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the free parameters governing decentralized inference synthesis—the slope that maps a participant's historical performance to its weight, the length of the memory used to score that performance, and the slopes that map performance to rewards—can be calibrated once and reused across network sizes. The authors extend the Allora Network (a decentralized inference-synthesis protocol) to classification, build a simulator that generates synthetic regression and classification tasks, and grid-search the parameter space over eleven participant compositions. They find that, within the explored grid, the network loss is minimized by a weight-mapping slope of $p=3$ for regression and $p=5$ for classification, by an exponential moving average with $\alpha=0.1$, and by reward slopes $p_i=3$, $p_f=3$, $p_r=1$. If the result holds, operators of decentralized learning networks do not need to retune per composition; they only need to pick the task-specific slope. The practical payoff is a calibration recipe that applies to any network that synthesizes one inference from many participants and rewards them by performance.

What carries the argument

The load-bearing mechanism is the soft-selection potential function $\phi_{p,c}(x)=\ln(1+e^{p(x-c)})$, whose gradient $\phi'_{p,c}(x)=p/(e^{-p(x-c)}+1)$ converts normalized regret differences into inference weights. The slope $p$ controls whether the network averages all participants or selects the best one: low $p$ blends, high $p$ selects. An exponential moving average $R_{i,l}=\alpha(\log L_i-\log L_{i,l})+(1-\alpha)R_{i-1,l}$ defines how far back 'historical performance' reaches, and the same potential function maps normalized scores to reward fractions, with a reward-entropy rule distributing rewards among tasks. The optimization sweeps $p$, $\alpha$, $p_i$, $p_f$, and $p_r$ on a grid and compares the 1000-epoch network loss and the cross-task reward spread across eleven network compositions.

What would settle it

Replay the parameter sweep on logged real-world data from a live decentralized topic, using actual worker inferences, losses, and skill trajectories; if the loss-minimizing weight slope is not $p=3$ for regression and $p=5$ for classification with $\alpha=0.1$, the defaults are an artifact of the synthetic skill dynamics rather than a general calibration.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single default parameter set keeps network loss near its minimum across every participant composition tested: $p=3$ for regression, $p=5$ for classification, $\log_{10}(\alpha)=-1.0$, $p_i=3$, $p_f=3$, $p_r=1$. For regression the loss is minimized at $p=3$; for classification the loss decreases monotonically with $p$ and saturates beyond $p=5$, so $p=5$ is chosen as the working default. Classification needs the steeper slope because label probabilities are bounded, so the regret-to-weight mapping must differentiate finely between models to select the best label, whereas regression benefits more from averaging. The memory parameter $\alpha=0.1$ corresponds to weighting roughly the last ten inference cycles, balancing adaptability against stability. The reward slopes keep the spread of mean rewards across inferers, forecasters, and reputers small, and the paper presents the resulting recipe as a generalizable calibration for any inference-synthesis-based decentralized learning network.

Load-bearing premise

The recommended defaults stand or fall on the assumption that the synthetic dynamics of worker skill in the simulator—steady exponential improvement plus rare random outperformance, and the hand-chosen difficulty of the synthetic classification tasks—faithfully represent how real participant quality changes over time.

Editorial extensions

If this is right

  • Network operators can fix $\alpha=0.1$ and reward slopes $p_i=3$, $p_f=3$, $p_r=1$ without retuning as participant counts change.
  • Classification networks gain little from adding participants beyond $N_i\approx 10$ and $N_f\approx 5$, whereas regression networks continue to improve up to at least $N_i=30$ and $N_f=15$.
  • The task-type difference implies that classification topics should use steeper performance-to-weight mappings than regression topics in any inference-synthesis network.
  • A single model-selection-versus-averaging rule—high $p$ for bounded outputs, low $p$ for unbounded outputs—can be carried over to other decentralized protocols with the same parameter structure.
  • The chosen reward slopes keep mean rewards close across inferers, forecasters, and reputers, so the defaults are compatible with a decentralized incentive structure.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the optimal $\alpha$ is tied to the simulated skill-improvement timescale (exponential decay at rate 0.03 and random outperformance factor 0.3), real networks whose participant quality drifts faster or slower would likely need a different or adaptive memory window; this is testable with logged worker histories.
  • The reward-spread criterion is one fairness proxy; minimizing worst-case participant regret or maximizing long-run participation could shift the recommended reward slopes, and the same simulator could be rerun under those objectives.
  • The saturation of classification gains at small participant counts suggests that classification topics should spend resources on participant selection or quality filtering rather than on adding workers, while regression topics can keep scaling worker count.
  • The design extension to classification with unbounded label sets points toward a common parameterization for other bounded-output tasks such as ranking or probability estimation, though the paper does not test those tasks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper presents an optimization framework for calibrating free parameters of decentralized online learning networks, using a simulator modeled on the Allora Network. The parameters are the slope p of the regret-to-weight mapping, the EMA parameter alpha for historical performance, and the reward-mapping slopes pi, pf, pr. The paper extends Allora's inference synthesis to classification tasks and conducts grid-search experiments over synthetic regression and classification data. The central claims are that the network loss is minimized for p=3 in regression and p=5 in classification, with log10(alpha)=-1.0, pi=3, pf=3, and pr=1, and that these defaults hold nearly independently of network composition. The paper further argues that the framework generalizes to any inference-synthesis-based decentralized AI network.

Significance. If the claims are correct, the paper would provide a useful, systematic calibration for a class of decentralized inference protocols and would extend a published regression-only design to classification. The manuscript is commendably explicit: the equations for inference synthesis, confidence intervals, reward distribution, and synthetic data generation are laid out in detail, and the refined grid search in Section 4.2 uses 10 seeds, which is a step toward reproducibility. The main scientific value lies in the falsifiable defaults it proposes. However, the load-bearing classification recommendation p=5 is not supported by the paper's own refined grid, and the broader generalization claim rests on a single synthetic simulator without robustness analysis. These issues are substantive and require revision.

major comments (5)
  1. [Section 4.2, Figure 5] The refined classification grid contradicts the stated default of p=5. The text reports mean optimal p = 4.79 ± 0.04 across all compositions, p = 4.85 ± 0.05 for inferer-heavy compositions, and p = 4.70 ± 0.08 for forecaster-heavy compositions. These values are neither equal to 5 nor piled up at the right edge of the grid (p=6), so they do not support a monotone decrease through p=5 or a saturation for p>5. The paper should report the actual loss differences among p=4.6, 4.8, 5.0, 5.5, and 6.0, together with seed-to-seed variance, before claiming that p=5 is optimal or that larger p is not beneficial.
  2. [Section 4.1, Figure 2] The classification default p=5 is a boundary optimum of the initial grid, which only considers p from 1 to 5. The text states that the loss 'monotonically decreases with increasing p' and that the decrease 'saturates for p > 5', but explicitly notes that this saturation 'is not shown in Figure 2'. The promised demonstration in Section 4 is not provided: Section 4.2 reports a refined range of 3 to 6, but the minima in Figure 5 average 4.79, not 6. A boundary optimum cannot be claimed as an interior optimum without an extended grid and a quantitative saturation check.
  3. [Section 3.1 and Section 3.2] The recommended defaults are calibrated entirely on a simulator whose data-generation constants are hand-chosen: the worker improvement timescale of 0.03 in Eq. 17, the random outperformance factor of 0.3, the classification stochasticity boost fclass=3 in Eq. 34, and the Dirichlet parameterization in Eq. 30. No sensitivity analysis is reported for these constants. Since the abstract and Section 5 generalize the results to 'any inference synthesis-based, decentralized AI network', the lack of robustness tests makes the defaults simulator-specific. At minimum, the authors should vary the improvement rate, the outperformance factor, and the classification boost and report whether p=3/5 and alpha=0.1 persist.
  4. [Section 3.1, Eqs. 15 and 13-14] The parameter values in the regression data-generation setup are internally inconsistent. The text sets the return standard deviation to sigma_rho = 0.01 but writes 'from N(0.01, 0.1)'. Equation 15 states log sigma_j ~ N(log 2 sigma_rho, log 1.5) = N(-0.699, 0.176), which holds only if sigma_rho = 0.1 and logarithms are base-10. If natural logarithms are used, log(1.5) is 0.405, not 0.176. This ambiguity changes the scale of worker errors and affects the simulation results; it must be corrected and checked for impact on the optimized parameters.
  5. [Section 4.1 and Section 4.2] Most performance comparisons, including Figures 2, 3, 4, and 6, are box plots over 1000 epochs but do not report variance across random seeds. Only the refined grid in Figure 5 uses 10 seeds. Consequently, statements such as the impact of pf and pr being 'minor (<12%)' or the reward spread being 'weak enough' are not supported by error bars or significance tests. The reward-spread recommendation pi=3, pf=3, pr=1 needs seed-level uncertainty to be credible.
minor comments (4)
  1. [Section 4.1, Table 2] The text says each parameter set is evaluated for 11 different network compositions, but Table 2 lists only six compositions; this discrepancy should be corrected.
  2. [Section 3.1, Eq. 23] Equation 23 uses Lijk on both sides: the left-hand side is the logarithm of the forecasted loss while the right-hand side uses the same symbol for the target loss. Rename one of these quantities to avoid confusion.
  3. [Section 3.1] The notation for logarithms is inconsistent: Equations 13-14 use exp and ln, while Equations 15-29 appear to use base-10 logarithms. The base should be stated explicitly at first use.
  4. [Section 4.1] The sentence 'We will show in §4 that it is not beneficial to increase p beyond p = 5' is vague; it should cite the specific figure or table that contains this evidence, which is currently absent.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the recommended parameters are grid-search outputs of an explicitly specified simulator, and the classification p=5 default is an unsupported boundary/rounding issue rather than a circular derivation.

full rationale

The load-bearing claims of the paper are empirical calibration results, not derivations: the selected defaults (p=3 or 5, log10(alpha)=-1, pi=3, pf=3, pr=1) are minima or balance points of the explicitly stated objective functions (Eqs. 35-37), evaluated over the parameter grid in Table 1 on the simulator defined by Eqs. 10-34. None of these equations defines the default parameters in terms of the conclusions, and the regression fiducial p=3 from Kruijssen et al. (2024a) is expressly re-optimized rather than imposed, so the recovery of p=3 is a genuine output rather than a fitted input renamed as a prediction. The self-citations to Kruijssen et al. (2024a) identify the protocol being mimicked and are not load-bearing as evidence for the optimality claims; the simulator equations are largely restated in this paper. The classification p=5 default is weaker: Table 1's p-grid stops at 5.0, Section 4.1 asserts without displaying evidence that the decrease saturates for p>5, and Section 4.2's refined search (p=3-6) reports a mean optimal p=4.79±0.04, with p=4.70±0.08 for forecaster-heavy compositions. This is a genuine internal-consistency and statistical-support defect that should be raised as a correctness concern, but it is not circularity: the paper's own minimization does not force p=5 by construction, and the refined grid values are not used as premises to define the default. Similarly, the generalizability to arbitrary decentralized networks is an inductive assertion supported by argument rather than by a self-citation uniqueness theorem, so it does not fall under the circularity patterns. Overall, the derivation chain is self-contained empirical calibration, and no load-bearing step reduces to its own inputs.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim (the recommended defaults) is a fitted result over a simulator whose data generation and dynamics contain multiple hand-chosen constants (improvement timescale 0.03, outperformance factor 0.3, fclass=3, Dirichlet constants). These constants are the unaccounted-for degrees of freedom that determine the outcome.

free parameters (5)
  • Ground-truth return distribution (mu_rho, sigma_rho) = N(0.01, 0.1)
    Sets the scale of regression losses; the resulting loss magnitudes determine which potential-function slope p is selected (Eqs. 14-15).
  • Worker skill improvement timescale = 0.03 per epoch (Eq. 17)
    The exponential decay of worker error and bias over ~30 epochs creates the temporal dynamics that make alpha=0.1 the best memory window; no real data supports this timescale.
  • Context-dependent outperformance factor = 0.3 (Eq. 19)
    A randomly selected worker's error and bias are multiplied by 0.3 each epoch; this event drives the value of the selection-vs-averaging slope p.
  • Classification stochasticity boost fclass = 3 (Eq. 34)
    Ad hoc multiplier chosen to match the stochasticity of regression tasks; directly influences the optimal p=5 for classification.
  • Dirichlet ground-truth parameters = v, b in [0,1], with constants 2 and 10 in Eq. 30
    The balance and concentration of synthetic classification labels are set by hand; the difficulty of the resulting classification problem shapes the fitted p=5.
assumptions (6)
  • domain assumption Decentralized inference can be reduced to a performance-to-weight mapping, a score-to-reward mapping, and an EMA memory, all using the potential function phi_{p,c}(x) = ln(1 + e^{p(x-c)}).
    The paper assumes the Allora design space is the relevant one (Eqs. 10-12) and does not compare to other aggregation schemes; generalization to 'any' network inherits this assumption.
  • ad hoc to paper The synthetic data generation process, including the Dirichlet-modeled ground truth and the worker error/bias dynamics, captures the real distribution of inference quality in decentralized networks.
    Eqs. 15-34 are chosen by the authors without real-world validation; all conclusions are conditional on this process.
  • domain assumption Minimizing the spread of mean rewards across participant classes (Eq. 37) is the correct objective for incentive design.
    The optimization of pi, pf, pr depends on this fairness objective, which is a design choice, not a mathematical consequence.
  • domain assumption The loss functions (MSE for regression in Eq. 35, MSE over label probabilities in Eq. 36) are the appropriate performance metrics for both task types.
    Using MSE for classification is justified for comparability, but alternative losses (cross-entropy, hinge) would likely change the optimal p.
  • domain assumption The 1000-epoch experimental horizon is long enough for the EMA and the worker-improvement dynamics to reach equilibrium.
    The paper notes that very low alpha values 'cannot impact the network within the 1000-epoch duration', indicating that the horizon materially affects the alpha selection.
  • standard math The stochastic processes (normal draws, Dirichlet draws, Pareto-distributed stakes) are independent and correctly implemented.
    The simulator relies on standard sampling algorithms; no explicit verification of the stochastic implementation is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimizing Decentralized Online Learning for Supervised Regression and Classification Problems." pith.science (2026). https://pith.science/paper/2NNXETXT

@misc{pith2026250116519,
  author       = {Pith},
  title        = {Pith review of: Optimizing Decentralized Online Learning for Supervised Regression and Classification Problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2NNXETXT}},
  note         = {Machine review of arXiv:2501.16519}
}
read the original abstract

Decentralized learning networks aim to synthesize a single network inference from a set of raw inferences provided by multiple participants. To determine the combined inference, these networks must adopt a mapping from historical participant performance to weights, and to appropriately incentivize contributions they must adopt a mapping from performance to fair rewards. Despite the increased prevalence of decentralized learning networks, there exists no systematic study that performs a calibration of the associated free parameters. Here we present an optimization framework for key parameters governing decentralized online learning in supervised regression and classification problems. These parameters include the slope of the mapping between historical performance and participant weight, the timeframe for performance evaluation, and the slope of the mapping between performance and rewards. These parameters are optimized using a suite of numerical experiments that mimic the design of the Allora Network, but have been extended to handle classification tasks in addition to regression tasks. This setup enables a comparative analysis of parameter tuning and network performance optimization (loss minimization) across both problem types. We demonstrate how the optimal performance-weight mapping, performance timeframe, and performance-reward mapping vary with network composition and problem type. Our findings provide valuable insights for the optimization of decentralized learning protocols, and we discuss how these results can be generalized to optimize any inference synthesis-based, decentralized AI network.

Figures

Figures reproduced from arXiv: 2501.16519 by the authors.

Figure 1
Figure 1. Probability distributions generated from the Dirichlet distribution D for a set of three labels Nic = 3 (where c Pic = 1 implies that the parameter space is a two-dimensional triangle), using different concentration parameter vectors ⃗α. A comparison between the two top panels shows that the normalization of ⃗α controls the overall concentration of probabilities, with higher ∥⃗α∥/ √ Nic resulting in a higher degree … view at source ↗
Figure 2
Figure 2. Variation in network loss as a function of the regret-to-weight mapping slope p (left) and the EMA parameter log10(α) (right), for regression (top) and classification (bottom). The box plots show the network loss over 1000 epochs and across all network compositions, with a red line showing the median and whiskers extending to the 10th and 90th percentiles. Beyond these percentiles, outliers are shown as transparent … view at source ↗
Figure 3
Figure 3. Variation in the instantaneous standard deviation of the mean reward received across each class of activity at epoch i, as a function of the score-to-reward mapping slope pi (left), pf (middle), and pr (right), for regression (top) and classification (bottom). The box plots show the reward spread over 1000 epochs and across all network compositions, with a red line showing the median and whiskers extending to the 10… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Variation in median network loss for different network compositions, as a function of the regret-to-weight mapping slope p (left) and the EMA parameter log10(α) (right), for regression (top) and classification (bottom). The median is taken over 1000 epochs, with each l…
Figure 5
Figure 5. Figure 5: Variation of the regret-to-weight mapping slope p that minimizes the network loss with network composition, represented as a function of the number of inferers Ni and forecasters Nf for regression (left) and classification (right). Each value is obtained by taking the …
Figure 6
Figure 6. Figure 6: Variation in median reward spread for different network compositions, as a function of the score-to-reward mapping slope pi (left), pf (middle), and pr (right), for regression (top) and classification (bottom). The median is taken over 1000 epochs, with each line repre…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 10 canonical work pages

  1. [1]

    2024, Information, 15

    Bhumichai, D., Smiliotopoulos, C., Benton, R., Kambourakis, G., & Damopoulos, D. 2024, Information, 15. https://doi.org/10.3390/info15050268 https://doi.org/10.3390/info15050268 -3pt

    work page doi:10.3390/info15050268 2024

  2. [2]

    2014, Ethereum: A next-generation smart contract and decentralized application platform

    Buterin, V. 2014, Ethereum: A next-generation smart contract and decentralized application platform. https://ethereum.org/content/whitepaper/whitepaper-pdf/Ethereum\_Whitepaper\_-\_Buterin\_2014.pdf https://ethereum.org/content/whitepaper/whitepaper-pdf/Ethereum\_Whitepaper\_-\_Buterin\_2014.pdf -3pt

    work page 2014

  3. [3]

    2017, Numeraire: A Cryptographic Token for Coordinating Machine Intelligence and Preventing Overfitting

    Craib, R., Bradway, G., Dunn, X., & Krug, J. 2017, Numeraire: A Cryptographic Token for Coordinating Machine Intelligence and Preventing Overfitting. https://numer.ai/whitepaper.pdf https://numer.ai/whitepaper.pdf -3pt

    work page 2017

  4. [4]

    Dinh, T. N. & Thai, M. T. 2018, Computer, 51, 48. https://doi.org/10.1109/MC.2018.3620971 https://doi.org/10.1109/MC.2018.3620971 -3pt

    work page arXiv 2018

  5. [5]

    Do g an, U., Glasmachers, T., & Igel, C. 2016, J. Mach. Learn. Res., 17, 1550–1831. https://doi.org/10.5555/2946645.2946690 https://doi.org/10.5555/2946645.2946690 -3pt

    work page arXiv 2016

  6. [6]

    Harris, J. D. & Waggoner, B. 2019, in 2019 IEEE International Conference on Blockchain (Blockchain), 368--375. https://doi.org/10.1109/Blockchain.2019.00057 https://doi.org/10.1109/Blockchain.2019.00057 -3pt

    arXiv 2019

  7. [7]

    Kruijssen, J. M. D., Emmons, N., Peluso, K., et al. 2024 a , Allora Decentralized Intelligence, 1, 1. https://doi.org/10.70235/allora.0x10001 https://doi.org/10.70235/allora.0x10001 -3pt

    work page doi:10.70235/allora.0x10001 2024

  8. [8]

    Kruijssen, J. M. D., Valieva, R., Peluso, K., Emmons, N., & Longmore, S. N. 2024 b , Allora Decentralized Intelligence, 1, 20. https://doi.org/10.70235/allora.0x10020 https://doi.org/10.70235/allora.0x10020 -3pt

    work page doi:10.70235/allora.0x10020 2024

Show all 18 references
  1. [9]

    & Leibler, R

    Kullback, S. & Leibler, R. A. 1951, The Annals of Mathematical Statistics, 22, 79 . https://doi.org/10.1214/aoms/1177729694 https://doi.org/10.1214/aoms/1177729694 -3pt

    1951

  2. [10]

    2017, in 2017 IEEE International Conference on Computer Vision (ICCV), 2999--3007

    Lin, T.-Y., Goyal, P., Girshick, R., He, K., & Dollár, P. 2017, in 2017 IEEE International Conference on Computer Vision (ICCV), 2999--3007. https://doi.org/10.1109/ICCV.2017.324 https://doi.org/10.1109/ICCV.2017.324 -3pt

    2017 doi

  3. [11]

    2008, Bitcoin: A Peer-to-Peer Electronic Cash System, https://bitcoin.org/bitcoin.pdf

    Nakamoto, S. 2008, Bitcoin: A Peer-to-Peer Electronic Cash System, https://bitcoin.org/bitcoin.pdf. https://doi.org/10.2139/ssrn.3440802 https://doi.org/10.2139/ssrn.3440802 -3pt

    2008 doi

  4. [12]

    & Rubin, H

    Olkin, I. & Rubin, H. 1964, The Annals of Mathematical Statistics, 35, 261 . https://doi.org/10.1214/aoms/1177703748 https://doi.org/10.1214/aoms/1177703748 -3pt

    1964

  5. [13]

    2023, Artificial Intelligence Review, 56, 3951

    Qammar, A., Karim, A., Ning, H., & Ding, J. 2023, Artificial Intelligence Review, 56, 3951. https://doi.org/10.1007/s10462-022-10271-9 https://doi.org/10.1007/s10462-022-10271-9 -3pt

    2023 doi

  6. [14]

    2021, BitTensor: A Peer-to-Peer Intelligence Market

    Rao, Y., Steeves, J., Shaabana, A., Attevelt, D., & McAteer, M. 2021, BitTensor: A Peer-to-Peer Intelligence Market. https://doi.org/10.48550/arXiv.2003.03917 https://doi.org/10.48550/arXiv.2003.03917 -3pt

    2021 doi arXiv

  7. [15]

    2021, Security and Communication Networks, 2021, 6126247

    Wang, R., Luo, M., Wen, Y., et al. 2021, Security and Communication Networks, 2021, 6126247. https://doi.org/10.1155/2021/6126247 https://doi.org/10.1155/2021/6126247 -3pt

    2021 doi

  8. [16]

    2024, Automation in Construction, 166, 105669

    Zheng, C., Tao, X., Dong, L., et al. 2024, Automation in Construction, 166, 105669. https://doi.org/10.1016/j.autcon.2024.105669 https://doi.org/10.1016/j.autcon.2024.105669 -3pt

    2024

  9. [17]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter doi edition editor howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.se...

  10. [18]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.