On The Role of Intentionality in Knowledge Representation: Analyzing Scene Context for Cognitive Agents with a Tiny Language Model

Mark Burgess

REVIEW 4 major objections 5 minor 23 references

On The Role of Intentionality in Knowledge Representation: Analyzing Scene Context for Cognitive Agents with a Tiny Language Model

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a training-free, language-agnostic score based on repetition, writing effort, and a fixed cognitive coherence scale can split narrative text into intended fragments and ambient context.

desk verdict A cheap, training-free intentionality heuristic with new formulas but qualitative evaluation; the fixed cognitive constants are the main risk, but the paper is honest and deserves a careful referee. read the letter →

arxiv 2507.10000 v1 pith:IMGSKBM3 submitted 2025-07-14 cs.AI cs.CL

Mark Burgess This is my paper

classification cs.AIcs.CL

keywords intentionalityknowledgerepresentationtinylanguagemodelcoherencen-gramfractionationtraining-freeextractioncognitiveagentssemanticspacetime

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 intentionality, the degree to which a fragment of a signal is produced for a purpose, can be read off a stream of symbols by counting how often patterns recur, how much work they cost to produce, and how their recurrences are spaced in time. The proposed score needs no knowledge of the language and no training: it ranks each n-gram by frequency times writing effort, suppressed once repetition becomes habitual, and then uses large gaps between recurrences to mark a phrase as intended rather than ambient. If this works, an agent can extract proto-concepts and knowledge-graph entry points from narrative at negligible computational cost, a step toward representing meaning without ontologies or large language models. The paper applies the method to several English texts and reports that the split between intended and contextual fragments matches what a human reader would call themes and names.

What carries the argument

The load-bearing object is the intentionality score $I(w,\Phi) = \Phi(w)W(w)/(1+\exp(\Phi(w)/\Phi_0 - \rho))$, built from three ingredients: a frequency count $\Phi(w)$ of an n-gram in a scene, a work value $W(w)$ measuring the cost of producing it (word length, or stroke count in character scripts), and a sigmoid that discounts repetition beyond a coherence interval. The coherence interval is set by a cognitive-attention constant, taken as roughly 45 sentences, which fixes both $\Phi_0 = 1/D_{30}$ and the reciprocal gap scale $\lambda = 1/45$ in the running form of the score. The fractionation machinery splits a text into n-gram streams ($n = 1, \ldots, 5$) and applies the score at each scale, so that longer, rarer fragments behave like proper names and short repeated fragments form the ambient soup. The longitudinal gap criterion adds the temporal dimension: a phrase whose occurrences are separated by a large gap after a quiet stretch is treated as intentionally reintroduced rather than habitually repeated.

What would settle it

Take a multilingual corpus whose intended fragments are known independently from headlines, quoted key phrases, or reader annotations, compute the score with the coherence scale varied by a factor of two on either side of the proposed 45-sentence value, and check whether the ranking of intended versus ambient fragments stays fixed and beats chance agreement with the human labels; if the ranking flips under the constant change, or does not beat chance, the central claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a scalar intentionality score separates 'intended' fragments from 'ambient context' in a stream of symbols. For a fragment $w$ with occurrence frequency $\Phi(w)$ and work cost $W(w)$, the score is $I(w,\Phi) = \Phi(w)W(w)/(1+\exp(\Phi(w)/\Phi_0 - \rho))$, where $\Phi_0$ is a coherence-frequency scale set by a cognitive attention limit and $\rho$ is a threshold near the expected density of ideas over a coherence interval. A complementary longitudinal criterion marks an n-gram as intentionally reintroduced when the largest gap between its occurrences exceeds the smallest gap plus the mean gap, indicating that an idea was revisited in a new context after a hiatus. Together the two criteria sort fragments into a 'contextual' class of frequent, evenly spaced, low-work items and an 'intentional' class of rarer, bursty, higher-work items. The paper argues this works without language knowledge or training, because the scales come from cognitive process limits rather than from document statistics.

Load-bearing premise

The result depends on the assumption that human cognitive attention has a stable scale, about 45 sentences for a coherence interval and a corresponding frequency threshold, and that this one scale applies across documents, genres, and languages; if that constant is wrong for a reader or text, the intended/ambient split becomes an artifact of the fixed partitioning rather than a property of the stream.

Editorial extensions

If this is right

A text can be indexed by its high-intentionality fragments, giving random-access entry points for knowledge graphs without ontologies or supervised training.
The same procedure applies to any symbol stream, such as DNA, sensor readings, or computer code, because it never relies on word meanings, only on recurrence, work, and spacing.
The method's computational cost is small enough for simple cognitive agents and real-time processing, in contrast to probabilistic batch training.
Long documents produce a stable linear calibration of intentionality, while short documents are the least reliable because they lack enough repetition to build a coherence assessment.
Intentionality is relative within a stream, so the score ranks each source's effort against itself rather than assigning universal importance values.

Reading between the lines

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

Editorial inference: if the ranking is robust across scripts with different writing costs, the work term $W(w)$ could be replaced by measured production effort, such as keystrokes, strokes, or speech duration, to transfer the method to spoken and gestural streams.
Editorial inference: the cognitive constants (a 45-sentence coherence interval and a $1/D_{30}$ frequency threshold) are claims about human attention; a direct experiment varying reader expertise or text genre could test whether the split tracks human judgments better than chance.
Editorial inference: the score's success at picking out themes such as anger in the paper's example texts suggests a low-level energetic proxy for what is usually called salience, but that link to psychological salience is not established in the paper.
Editorial inference: one could pit this training-free score against a large language model's attention weights on the same extraction task; if they correlate, attention mechanisms may be learning an approximation of this coherence-work ranking.

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.

Desk Editor's Note

read the letter

Here's the short version. The paper proposes a training-free, low-cost heuristic for labelling n-gram fragments as 'intended' or 'ambient' in narrative text. The specific formulas—work-weighted frequency with a saturating sigmoid (Eq 8) and the longitudinal gap criterion (Eqs 10-11)—are new relative to the author's earlier papers, and the method is fully specified with code released. Applied to Darwin and Moby Dick, the output is plausible: the ambient fragments read like ordinary collocations, the anomalous ones like content-bearing phrases. The author is also candid that the assessment is 'somewhat ad hoc' and relative, not a rigorous computation.

What the paper does well: it commits to a concrete operationalization of intentionality and shows it running on real texts at negligible cost. That is more than most philosophy-adjacent AI papers do. It also distinguishes intent from frequency in a sensible way: rare long phrases get high intentionality, over-repeated padding gets suppressed.

The soft spots are real and concentrated in the calibration. The constants Φ0=1/D30 and λ=1/45 are imported from Dunbar numbers without any sensitivity analysis. The stress-test is correct: if those constants move, the intended/ambient split can move with them, and Figure 2's apparent stability owes a lot to Eq (11), where 1-e^{-λ(τ-τlast)} saturates by construction. The linearity being 'partly a choice' is admitted in the text. There is also no baseline—no TF-IDF, no frequency-only ranking—so we can't tell how much work the intentionality weighting is actually doing. The evaluation is qualitative across six texts, with no error bars.

None of this kills the idea. The paper is honest about its limits, and the problem it addresses (cheap seed concepts for knowledge graphs) is real. The right fix is to add a baseline comparison and a sensitivity sweep over the two constants; that would convert a plausible demonstration into a testable method. As it stands, the central claim is under-supported but not broken.

Who should read it: people building tiny cognitive agents or knowledge-graph seeds from raw text, and anyone interested in non-Markov text scoring. Readers who need rigorous NLP evaluation will be unsatisfied. I'd send it to peer review with a request for those experiments; it deserves a serious referee.

Referee Report

4 major / 5 minor

Summary. The paper proposes a 'tiny language model' based on Promise Theory's Semantic Spacetime that estimates the intentionality of n-gram fragments in narrative text without knowledge of the language and without training. The central quantity is the intentionality score I(w, Φ) = Φ(w) W(w) / (1 + exp(Φ(w)/Φ0 − ρ)) in Eq. (8), together with a longitudinal gap criterion in Eq. (10) and a running variant in Eq. (11). The author applies the method to six English texts, separating fragments into 'ambient context' and 'anomalous/intended' categories, and discusses how these fragments can provide entry points for knowledge graphs. The paper claims that the method is low-cost, training-free, and cognitive-universal, with caveats that the assessment is approximate and depends on the observer's memory capacity.

Significance. If the claimed separation is robust, the method could offer a genuinely cheap, unsupervised heuristic for extracting proto-concepts from text, complementing LLM-based approaches and requiring no probabilistic training. The paper is explicit about its idealized nature and provides a software implementation (reference [8]), which is a strength. However, the current evidence is only qualitative: six texts, no baseline, no non-English language, no error bars, and no sensitivity analysis. The significance is therefore conditional on a much stronger empirical demonstration than is presented here.

major comments (4)

[§4.1–4.2, Eqs. (8) and (11)] The claimed language-independent, training-free separation rests entirely on hand-set cognitive constants Φ0 = 1/D30 and λ = 1/45, with ρ stated only as 'of order 1/D30'. No sensitivity analysis is provided. Since the ranking of ambient vs. anomalous fragments is determined by these constants, the authors should vary them over a plausible human range (e.g., D30 between 20 and 100) and show that the resulting ranking is stable. Without this check, the separation may be an artifact of the calibration rather than a property of the text.
[§5.2, Figure 2 and Eq. (11)] The stability and convergence shown in Figure 2 are to a large degree a mathematical consequence of the functional form of Eq. (11): as τ grows, the factor (1 − e^{-λ(τ−τ_last)}) tends to 1, so the score converges to a constant involving W(w). The author acknowledges in §5.2 that the linearity is 'partly a choice' due to linear partitioning. Figure 2 therefore cannot serve as empirical validation of the method's discriminating power. A direct comparison of the extracted fragments against a human-annotated reference set or a simple baseline (e.g., TF-IDF, frequency thresholds) would be more informative.
[§6 and overall evaluation] The demonstration is limited to six English texts, and the assessment is qualitative ('remarkably well to a human eye' in §7). The central claim of language independence is not tested, and no quantitative agreement measure (e.g., precision, recall, rank correlation) is computed. A minimal evaluation with a few non-English texts and a quantitative comparison with at least one baseline is needed to support the assertion that the method works 'without knowledge of the specific language.'
[§4.1, §5.2, and §7] The paper does not report the exact numerical values of the parameters used in the experiments, particularly Φ0, ρ, and λ. Since the ranking depends on these values, the results cannot be reproduced from the description alone. The author's own caveat in §7 that the intentionality measure is 'a somewhat ad hoc assessment, not a rigorous computation' further underlines the need for a transparent parameter specification and a sensitivity analysis.

minor comments (5)

[§4.1, after Eq. (7)] The sentence beginning 'Any intentional phenomenon' is a sentence fragment; it should be completed or merged with the following sentence.
[§4.2, Eq. (11)] The notation 'I(w, ; τi, τf)' contains a stray semicolon; it should read 'I(w; τi, τf)'.
[§4.1] The phrase 'This we define a coherence frequency scale' is ungrammatical; consider 'We define this as a coherence frequency scale'.
[Figure 2] The caption is difficult to parse: it refers to 'left hand average results' and 'right hand side' but the panels are not clearly labeled within the figure itself, and the top-left panel appears to lack a legend.
[References] Reference [9] is missing volume and page numbers ('European Economic Review, 2025'), and reference [6] cites 'notes available on Researchgate', which is not a stable or citable source.

Circularity Check

2 steps flagged · score 6.0 of 10

The claimed stability/convergence of intentionality is built into Eq. 11, and the intended/ambient split is the ranking score renamed; partial circularity.

self definitional [§4.2, Eq. (11); Fig. 2 caption; §5.2]
"Given the expected high noise levels in text, the remarkable linearity of the calibration over long texts (figure 2) is a result of the linear partitioning into coherence regions. This is partly a choice [...] So, I(w, ; τi, τf ) = W (w)(1 − e−λ(τ −τlast)), τ > τlast. (11)"

Equation (11) is an exponential relaxation to the constant W(w): as τ−τlast grows, the factor 1−e^{−λ(τ−τlast)} tends to 1, so I tends to W(w). The 'Convergence to Stability' plotted in Figure 2 and called 'remarkable linearity' is therefore a mathematical property of the chosen formula together with the linear partitioning, not an empirical discovery about the texts. The paper itself concedes the linearity is 'partly a choice'. The stability result is the input definition wearing the costume of a finding.
self definitional [§4.1 Def. 4 and Eq. (8); §5.2 classification lists]
"Definition 4 (Intentionality of a string) The intentionality of a string is defined to be a weighted function of the work invested in representing the string in terms of its n-grams. [...] If we are in doubt, a change is counted as intentional, so the cognitive work is not so much about ranking intent as in discovering what to ignore in the ambient field."

The labels 'ANOMALOUS' and 'AMBIENT' in Section 5.2 are produced by thresholding the same score I that Definition 4 and Eq. (8) define; no independent annotation of 'intended' is used. The claimed separation of intended content from ambient context is therefore the score's definition restated as a result. The only offered validation is 'Inspecting the output of fragments with a human eye' plus post-hoc thematic readings of Darwin and Moby Dick, which does not test whether the score tracks an externally defined property.

full rationale

The clearest circular step is the convergence/stability claim: Eq. (11) is constructed so that the intentionality score asymptotes to W(w), so Figure 2's convergence is forced by the formula, as the paper partially concedes. The intended/ambient distinction is also definitional: the score defines which fragments count as intentional, and then those fragments are presented as the successful output. The Dunbar-based constants Φ0=1/D30 and λ=1/45 are load-bearing and are imported from the author's prior work with Dunbar, but that is an externally published empirical study rather than a self-asserted uniqueness theorem, so I treat it as a calibration risk rather than full circularity. Overall, the central separation claim retains some independent content as a proposed heuristic, but the paper's headline stability evidence reduces by construction, giving a partial circularity score of 6.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim relies on five imported or postulational components: the Promise Theory framework, the definition of intentionality as work and non-randomness, the Dunbar-derived calibration constants, the sentence-based proper time, and the additive work model. None of these is given independent validation in this paper; the first three come from the author's own prior work.

free parameters (4)

λ (reciprocal coherence length) = 1/45 per sentence
Chosen from Dunbar number for attentive work; sets the longitudinal decay rate in Eq (11). No sensitivity analysis is performed.
Φ0 (coherence frequency scale) = 1/D30, approximately 1/45
Set in Section 4.1 as the threshold separating intended from habitual frequency in Eq (8).
ρ (intentionality threshold shift) = order of 1/D30
Used in the sigmoid denominator of Eq (8); the paper does not give a precise value.
n-gram order range (n=1..5) = 5
The method truncates n-grams at n=5 due to power-law sparsity; this affects what counts as a long fragment.

assumptions (5)

domain assumption Promise Theory provides a valid causal and semantic model for agents
The entire framing assumes Promise Theory's promiser and observer structure, introduced in Section 2.
ad hoc to paper Intentionality can be operationalized as non-randomness plus work cost
This is the paper's central postulate, introduced in Section 3 and Definition 5, not derived from external theory.
domain assumption The Dunbar number for attentive work (45) is a universal cognitive scale for narrative coherence
Imported from references [9] and [23] and used to set λ and Φ0 in Sections 4.1 and 4.2.
domain assumption Sentence count is an appropriate 'proper time' coordinate for all texts
All measurements use sentence positions rather than word or character positions, as stated in Section 4.
ad hoc to paper Symbolic work W(σ) is proportional to string length, with a small premium for uppercase
Definition 5 assumes writing effort is additive and length-dominated, which is language-specific and untested.

how reviews work

0 comments

Cite this review

Pith. "Pith review of On The Role of Intentionality in Knowledge Representation: Analyzing Scene Context for Cognitive Agents with a Tiny Language Model." pith.science (2026). https://pith.science/paper/IMGSKBM3

@misc{pith2026250710000,
  author       = {Pith},
  title        = {Pith review of: On The Role of Intentionality in Knowledge Representation: Analyzing Scene Context for Cognitive Agents with a Tiny Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IMGSKBM3}},
  note         = {Machine review of arXiv:2507.10000}
}

read the original abstract

Since Searle's work deconstructing intent and intentionality in the realm of philosophy, the practical meaning of intent has received little attention in science and technology. Intentionality and context are both central to the scope of Promise Theory's model of Semantic Spacetime, used as an effective Tiny Language Model. One can identify themes and concepts from a text, on a low level (without knowledge of the specific language) by using process coherence as a guide. Any agent process can assess superficially a degree of latent `intentionality' in data by looking for anomalous multi-scale anomalies and assessing the work done to form them. Scale separation can be used to sort parts into `intended' content and `ambient context', using the spacetime coherence as a measure. This offers an elementary but pragmatic interpretation of latent intentionality for very low computational cost, and without reference to extensive training or reasoning capabilities. The process is well within the reach of basic organisms as it does not require large scale artificial probabilistic batch processing. The level of concept formation depends, however, on the memory capacity of the agent.

Figures

Figures reproduced from arXiv: 2507.10000 by the authors.

**Figure 1.** The spacing between repeated signals is further evidence of intentionality. Frequent repetitions all on the same scale tend to indicate habitual usage, padding signals (like spacing etc). However a hiatus followed by a repeated use over different timescales indicates that there is intent to repeat in a new context. At each new scale or level of aggregation Λ, the alphabet of states Σ Λ defines the set of acceptable … view at source ↗

**Figure 2.** Plot giving a heuristic impression of the degree of overlap in fragments selected by I(w, ; τi, τf ) and I(w, ; Φ) for the documents studied. Upper panels show average and deviation of diverse document lengths and types, giving a rough idea of the spread of intentionality using different models. The left panels are calculated using the expression in (8); on the right are the counterproposal for intentionality decayi… view at source ↗

**Figure 3.** The process of transmuting episodic sensory signals info fractions, then low level impressions associated with them, then collective impressions about a coherent interval of time, and finally distilling these into long term persistent concepts over a much longer time frame. Inspecting the output of fragments with a human eye, someone’s one method appears appealing, sometimes the other. The longitudinal coherence met… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Plot of running I(w, ; τi, τf ) over a long document, showing that there is no obvious pattern to the level of intentionality in consecutive sentences. Intentionality appears as a random variable, because natural language is highly dense in associative meaning. Samplin…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 16 canonical work pages

[8]

M. Burgess. Sstorytime project. https://github.com/markburgess/SSTorytime/ tree/main, 2025

work page 2025
[1]

Bergstra and M

J.A. Bergstra and M. Burgess. Promise Theory: Principles and Applications (second edition) . χtAxis Press, 2014,2019

work page 2014
[2]

M. Burgess. On the theory of system administration. Science of Computer Programming , 49:1, 2003. 18

work page 2003
[3]

M. Burgess. Spacetimes with semantics (i). arXiv:1411.5563, 2014

arXiv 2014
[4]

M. Burgess. Spacetimes with semantics (ii). arXiv.org:1505.01716, 2015

arXiv 2015
[5]

M. Burgess. Spacetimes with semantics (iii). arXiv:1608.02193, 2016

arXiv 2016
[6]

M. Burgess. Motion of the third kind (ii) notes on kinematics, dynamics, and relativity. (notes available on Researchgate), 2022

work page 2022
[7]

M. Burgess. Notes on trust as a causal basis for social science. SSRN Archive, available at http://dx.doi.org/10.2139/ssrn.4252501 (DOI: 10.2139/ssrn.4252501), August 2022

doi:10.2139/ssrn.4252501 2022

Show all 23 references

[9]

Burgess and R.I.M Dunbar

M. Burgess and R.I.M Dunbar. A quantitative model of trust as a predictor of social group sizes and its implications for technology. European Economic Review, 2025

2025
[10]

Mark Burgess. Testing the quantitative spacetime hypothesis using artificial narrative comprehen- sion (i) : Bootstrapping meaning from episodic narrative viewed as a feature landscape, 2020, 2010.08126

2020 arXiv
[11]

Mark Burgess. Testing the quantitative spacetime hypothesis using artificial narrative compre- hension (ii) : Establishing the geometry of invariant concepts, themes, and namespaces, 2020, 2010.08125

2020 arXiv
[12]

Agent semantics, semantic spacetime, and graphical reasoning, 2025, 2506.07756

Mark Burgess. Agent semantics, semantic spacetime, and graphical reasoning, 2025, 2506.07756

2025 arXiv
[13]

Cover and J.A

T.M. Cover and J.A. Thomas. Elements of Information Theory. (J.Wiley & Sons., New York), 1991

1991
[14]

R. Dunbar. Grooming, Gossip and the Evolution of Language. Faber and Faber, London, 1996

1996
[15]

Trust, semantic learning, and monitoring

NLnet Foundation. Trust, semantic learning, and monitoring. NLnet sponsored project website
[16]

Kahneman

D. Kahneman. Thinking, Fast and Slow. Penguin, London, 2011

2011
[17]

Lewis and C

H. Lewis and C. Papadimitriou. Elements of the Theory of Computation, Second edition . Prentice Hall, New York, 1997

1997
[18]

Manning and H

C. Manning and H. Schutze. Foundations of Statistical Natural Language Processing. MIT Press, 1999

1999
[19]

A review on the attention mechanism of deep learn- ing

Zhaoyang Niu, Guoqiang Zhong, and Hui Yu. A review on the attention mechanism of deep learn- ing. Neurocomputing, 10:48–62, 2021

2021
[20]

J.R. Searle. Intentionality. Cambridge, 1983

1983
[21]

Shannon and W

C.E. Shannon and W. Weaver. The Mathematical Theory of Communication. University of Illinois Press, Urbana, 1949

1949
[22]

Vaswani, N

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez, L. Kaiser, and I. Polo- sukhin. Attention is all you need, 2023, 1706.03762

2023 arXiv
[23]

W.X. Zhou, S. Sornette, R.A. Hill, and R.I.M. Dunbar. Discrete hierarchical organization of social group sizes. Proc. Royal Soc., 272:439–444, 2004. 19

2004

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [8]

M. Burgess. Sstorytime project. https://github.com/markburgess/SSTorytime/ tree/main, 2025

work page 2025

[2] [1]

Bergstra and M

J.A. Bergstra and M. Burgess. Promise Theory: Principles and Applications (second edition) . χtAxis Press, 2014,2019

work page 2014

[3] [2]

M. Burgess. On the theory of system administration. Science of Computer Programming , 49:1, 2003. 18

work page 2003

[4] [3]

M. Burgess. Spacetimes with semantics (i). arXiv:1411.5563, 2014

arXiv 2014

[5] [4]

M. Burgess. Spacetimes with semantics (ii). arXiv.org:1505.01716, 2015

arXiv 2015

[6] [5]

M. Burgess. Spacetimes with semantics (iii). arXiv:1608.02193, 2016

arXiv 2016

[7] [6]

M. Burgess. Motion of the third kind (ii) notes on kinematics, dynamics, and relativity. (notes available on Researchgate), 2022

work page 2022

[8] [7]

M. Burgess. Notes on trust as a causal basis for social science. SSRN Archive, available at http://dx.doi.org/10.2139/ssrn.4252501 (DOI: 10.2139/ssrn.4252501), August 2022

doi:10.2139/ssrn.4252501 2022