CodeStream #005.1 ~ Inside the template/ studio

right. All right. Been a while. It is May 27th, 2026. 2026. This is inside the template studio going into the Doxology template repo as it currently exists. and showing a few ways that I'm using it right now to do things like generate this modular textbook which we'll look through as well as publish with this cursor dispatch an active inference manuscript that just from a code perspective we'll look through in this stream. So write any questions in the chat. Um I will go through the template repo overall then how it's used to render the templates which are the template projects that are like control positives inside the repo and then the biology textbook and hopefully by then show the publishing branch with the double referential publishing between Zenotto and GitHub of a new paper and it's

running right now. Let's go to the template overall. Okay, so this repo uh was described in a paper in March 2026 and I've continued to use it for different projects and and versioned it most recently using cursor composer 2.5 and claude code with opus 4.7 as the main LLMs and using claude code and Hermes agents and made several uh uh refactorings and additions that give new functionalities and services for people and agents to work with this repo quite functionally. On the left hand side are the top level folders. So all the dot folders are uh of secondary relevance. Let's just look at these folders that aren't starting with a dot. The docs folder has a variety of documents that you can point your LLM at that are hopefully kept well linted and signposted accurately. HTML

cove is a testing artifact. It's not relevant. Can be deleted. infrastructure is the codebase for the template repo itself. So to say that we're in the printers studio or template studio or publishing factory or publishing factory for publishing materials materials that's top level code from the infrastructure infrastructure and at its core is a set of methods that is going to do firstly running the repo level test suite in infrastructural tests. So confirming that the factory itself that the infrastructural code generation methods themselves are modular and functional using its own top level test suite that can be run or not not and then when run.sh SH or any other surface of calling these methods is deployed which can be a um API like or an import like an MCP or skill like the repo will iterate

over the projects that are active in projects and then there's projects in progress which is kind of like the back burner and projects archive. So, and there's strong confidentiality guard rails. So, all of the projects, which right now I'm just showing the two that we'll focus on in this stream, which are examples of how they all are in the archive and in progress, which which you could make any other folder system you want. There these are all arrows representing sim links. So the template repo itself which is open source source it rides with these five template projects projects included in the repo but also published a certain way as all the other projects are if if you want and all of the actual projects themselves like the things that I don't want to go to

template main are in a sidecar repo. That is you can see from the mouse over is slash projects slash and then the projects. So slash projects is like a private mono repo I have that either includes or doesn't all other projects ones that I'm using this system for or not. And then through sim links in the template repo you can specify which projects do you want discoverable on the front burner iteratable when you want to hit that one run button and do each project in series everything which of a process that is going to be described once we jump into the project level. But for each project that is selected whether just one or set by choice or iterate over all in that projects folder just for simplicity and completeness. completeness. Um Um run for

each of these projects according to a similar structure. run that project level test suite with acceptance or um s stopping or continuation uh thresholding right there on through rendering the manuscript and running the full analysis. All the variables where possible are are placeholders in the manuscript. It's kind of like a mad lib. And then during the run from the top level the uh tokens are injected into the markdown for the manuscript so that the final rendered PDF will look like this for these uh active projects and the ones that start with template underscore. So they are real projects. they are published in as if the other projects um are with this GitHub plus cenotto cenotto uh double publication automatic system and also these template ones are special because they're kept in the public main as

examples so that it's really easy to start with a prompt which like if somebody in the live chat writes a suggestion during the stream I'll uh start kick off a project like this. But a common start is like in the format of write a whole new project on and then that will um use the syntax and structure of these template projects which can be really strongly typed. and documented very bizarrely and bespokely bespokely um as examples. And then because these two are open source projects, the biology textbook and the active inference policy entanglement paper um um they're being sim linked. So they're found from running at the top level dot slashrun.sh SH makes a plain text menu and you have a you can pick a letter from doing environment setup and the test suite for

the infrastructure methods methods on through one or all the projects with the core methods which is running the whole project test suite methods themselves and manuscript rendering into the the uh uh formats of PDF, PDF, EPUB, DOCX, etc. or having this additional LLM step which can do things like translate it or um have a reviewer comments like a reviewer persona. Give an LLM comment using a a a panel of persona to review it against like different prompts that can be applied to the final rendered artifacts. each. Okay. Now, let's look at the template projects and their outputs. So, the template projects are of uh five different kinds. And I'm just showing a total slice in time. Check check check back in with what's really there and everything else. uh rather than defer to this exact

snapshot because I'm continuing to make changes and I'm sure it'll change in many ways too and not necessarily um forwards compatible ways per se. The template template project documents this repo itself, but this is an example where it runs code analyzing its own codebase, renders it into this PDF. Template pros project is a prosheavy with just mermaid diagrams or just pros but not doing any simulations This is a codebased project that does a little minimal simple optimization step and just has the placeholders for inserting visualizations and for running simulations in a in a more of a codebase where you're injecting numbers in from the codebases run into the PDF into the market. Um, then there's two more recent ones that were just spun up over the last couple of days and that is the template

auto research project. So this one is injecting variables from a auto researcher loop on the mnest digit using some neural networks training neural networks using auto researcher methods that's what the codebase is doing um and this paper reports reports the results of an auto researcher run run injecting in aspects of the codebase and the typing and has has visualizations and can can represent different aspects of the uh auto researcher run kind of target inference as well as features of the auto researcher um process itself probably very interesting things to study and then there's the template active inference project which is just the total experiment first pass and it looks like the PDF didn't render for that one so let us start an agent to resolve that fully properly renders. So just as a exploring around

uh today I was just putting in that first and last page page of the PDF having this beginning of transmission kind of facts cover letter and then a end of transmission. So in a if this were in a a plain text format or uh could have plain text cryptography or could have images on these known bookmark pages and have that page and then and then I think right now it's displacing the real cover page but then have the first page. So kind of like the first um special pages of the book. So those are the template projects which I will try to make better and more more robust and just more aligned. But there's been a lot of codebase changes. So just sort of settling down the plane, but it's it's 99% there. But on

um rendering and conforming the template projects to the current state of the code. Um, and then the these other projects kind of started in in slightly more open forms of of this earlier forms of this template repo. But between cursor composer 2.5 with plan mode and opus 4.7 and/goal and Hermes agent and several other techniques and developments. Um I I I think a lot has happened on the template repo itself and that is now turning to two of the non non template kind of minimal projects. So the the the goal of and of having and using using and um highlighting these template underscore and calling it template is like and although the top level uh repo I can kind of imagine something uh different than template itself but at least the template projects makes it

very easy stylistically for agents to oneshot projects that can adapt a codebase really really fast and modularly and know that with certain constraints and and norms being followed that there's a lot of flexibility to for example make a generative biology introductory textbook. So, here's the sim link to the bio textbook. So this is uh published yesterday to Pruise hydrogen holographic science and intelligence zodo community and there's a GitHub repo doxology biology textbook and a matching Zenono publication which has just the compressed press PDF. This just one possible Zenoto uh versioning and DOI and and co-reference which is going to come into play uh shortly. shortly. And this book has some of the following structure. Well, ju just to say and and there there's I'm sure more to say on this, but this is partially related

to my experience this semester at College of the Redwoods teaching the two introductory biology classes, which is stories for another day. But suffice to say a lot of thought and discussion and and uh some feedback around what is introductory, what is biology, how to go about it, what kinds of availabilities in the lab and printed material can support what kinds of synchronous experience and learning materials for um study notes and trans media education. ational experiences and and what can open up and and happen in the uh away from the text and just thinking about what all that looks like in this generative era. So cover front matter second page with the William Blake quote the tree which moves some to tears of joy is in the eyes of others only a green thing that stands

in the way. Some see nature all ridicule and deformity and some scarce see nature at all. But to the eyes of the man of imagination, nature is imagination itself and it is a open- source textbook. There is a automatically generated obviously as the rendering is but there's sort of two layers to the generativity at at least here. There's the multiple passes through through feedback amongst agents and with myself scanning and sampling random parts of the book to generate the material. So I didn't handw write any of the primary material of this book uh uh nor the the details of the rendering of how the markdowns which are looking like projects biology textbook manuscript and then this is what the front matter markdowns look like. And then this is what the chapter markdowns looked like. Mermaid

diagrams in line, inserted images in line. And this table of contents itself is uh several dozen pages 60ome pages. Um Um but it's just to just to start with I mean think about local adaptation small model translation on the edge um diving in accordion fractal learning act active learning with zooming in or using online search or connecting to lit more the more literature review methods between the textbook. So, total starting point and uh the chapters themselves are all linked up to click around and the citations are all rendered at the end of the book. There are questions around Bloom's taxonomy of learning ordering of the chapters because for each chapter there we can see for example how many words or lines are in each chapter or do we have what claims correspond to which chapters

and sections if it's relevant or which ones do we have more subsections for or which ones have uh different computational resources because because it's not just manuscript files in the biology textbook folder. There's the SRC which are the source methods for that that are invoked by the test internal to the biology textbook which is what allows it to be a standalone repo separate from the template rendering with all the information in the markdown. Like all the template is is the convenience method in in multiple senses for running the test suite, clearing the outputs if desired, running the full reanalysis if desired, and then rerendering the PDF with some of these kind of cool and fun PDF tools like including steganography and cryptographic methods for the publishing and also automated publishing. to to GitHub and Zenotto

um um through the infrastructure publishing publishing modules. Um so when biology textbook gets called up by run.sh SH or you're just in a clawed code within biology textbook, don't have the template in scope at all and then just say run the whole textbook again. Then it will first run the tests of the repo itself. itself. Then it will run the defined set of scripts scripts which are enforced to be thin orchestrators orchestrators which means that they're not going to define any new methods. All methods are defined in the src modules. It's the same practice as at the higher repo level. The scripts are just like the pseudo codes that thinly get called by uh conditional upon the tests passing if they're required to pass or proceeding if the tests fail if it's if it's allowed

to fail. Then the scripts are called which might range from just generating like a deterministic figure and just having like a map of a region all the way on through invoking like a very large computational run that might involve like checking and confirming that you can connect to a cloud uh job distribution system and then spinning up and verifying that much cloud compute and then updating and actually running the simulation and pulling all the log into it's like as as as big as that analysis is whatever you want the simplest simplest lowest cognitive overhead for context switching into this project for doing these different computational methods from the project that's 99% pros and it's just type setting to the project that is going to be heavily modularized with a variety of outputs that's the kind

of expressivity that I This is just one design play and exploration in that space of that's the biology textbook and let's now check if so policy entanglement is this paper that is sim linked from under projects policy entanglement which is a this is a more computationally intensive project which has a simulationbased PMDP component as well as a lean theorem environment environment uh aspect in in the code itself but it still has that manuscript markdown markdown structure structure src src methods methods the scripts to orchestrate the tests to validate the SRC. So it has that standalone documentation. Let's reload. reload. There it is. So during this very stream, the the GitHub was pushed I guess I could manually edit out that one abstract field. Changing the abstract doesn't change the DOI. So, here we go. Yes,

tiny little rendering. rendering. issues, some little bit of multi-ublication issues, but um it's there and there's a cross reference between the GitHub push and the Zenotto and the DOI. And um first let's look at the PDF of this paper. Then we'll look a little bit at the code of this paper which is a much we will publish an update. Let's do it. So check the title says that we don't want that in PDF. Check the zenotto abstract. There is well we already deleted that first part but there is this. So the end is a bit well. So we the we we this is the only part that really should be changed. just that one one bit weird with that. Okay. So, it'll make those improvements. We'll we'll continue to put some in uh to the

chat as we read this uh cover page should have the DOI email email metadata Check the other chat. Right, here's that active inference template page. Another another day, but I'm super excited about this one. Very excited about this one. Okay. But we'll update the um this to v 0 1.01 or something during this stream the next um 20 minutes and uh double push it to Zenotto and to GitHub. Okay. Um wow. Hyperlinks must be read, of course. But they're they are hyperlinks, but they're not read. But that's kind of interesting. Little bit like little covert hyperlinks. Um, Um, okay. Like what this paper is about? Well, first it's intended as a structured artifact, not necessarily something where the the exact wording matters as much as the real process that leads to this type of multiple

referential technical structured research document. So, it's not even necessarily only about the PDF or any specific part of it. It's about having this project as a versioned modular pros large technical analytical documentation that's just like in the repo but it's kind of parah to the manuscript even the supplements and the tested codebased thin orchestrated um all these other kinds of of surfaces. this. So, here's the uh oh, abstract is inserted twice. Um here's the beginning part which I did review multiple times to So picking up here, active inference is mathematically compact, biologically motivated, and increasingly used across perception, decision-making, robotics, and computational psychiatry. While its direct empirical status remains a comparative modeling question, an agent of any non-trivial complexity, however, is not best described by a single policy variable. So with reference to however you're

going to fit that policy variable reward precision this or that architecture this or that composing down or projecting down to this or that free energy or energy landscape whatever your input perception like space andor output action-like space but focusing here on the action space uh inference a real agent juggles many concurrent policies motor and intentional so for example policies in the body and in the mind fast and slow for multiple time scales of different policy responsiveness responsiveness modality specific. So for example to um if if you can change a object's shape, color, smell, size like all the senses the different modalities there could be policies across different modalities like choosing to one of your sensors is an air air horn sound, one of them's a light, one of them releases a smell, one of them

does some haptic. So those would be like different modalities of policies, not just multiple light actuators. And there's the possibility that in this graph, graph, some streams treat the next move as a oneshot gradient descent. So um a sort of uh mid complexity type conservative Newtonian type actor. So here it could be this is not the only uh type but it could be like a pendulum or something which is just it's exact position and momentum velocity all that exactly determine its next move whether it's a single pendulum double pendulum and so we can still think about its sort of action selection as happening through variational free energy minimization on a basically um convex landscape. ape which is like non-planning. It's not looking ahead. It's not some appeal to qualia. Not an appeal even to as

if planning just exact same results from the same location. Whereas other levels even for one or for more time steps, people are often interested in studying the expected free energy or policy search methods for how different affordances can be chained into policies and re-evaluated at some frequency relative to to that policy. to uh evaluate those different policy sequences in terms of their next move for actionability. looking into a certain sequence that could get reevaluated or not in terms of those policies having uhformational and pragmatic value and and the standard partially observable marov decision process treatment. and SPMDM split inference across independent factors and policy that is they rely on strict meanfield factorizations across hidden state factors observation modalities and the multiream construction here policy variables so we can explore and definitely I'd like to look

into and learn more about where are Are we implicitly dealing with a mean field across modalities or in some of these other situations that that we're discussing in SPM and PMTP and uh making different examples that that speak to different parts of that. The mean field separation is computationally cheap cheap and biologically suggestive. It mirrors functional segregation, sparse cortical connectivity, and the modular character of much of that cortex. So that's to say in terms of approximations go, this is a pretty extreme though natural and often tractable and uh adequate parsing of essentially how are you going to chop up that joint distribution of all by all by all to give some model sparcity rather than picking from the largest possible set of of models which even where there's really some sub some subset of models

that a good performance and not overfitting and that kind of like adaptive model performance. Um, it could be diluted or or made to be below some statistical power correction threshold in very large panels. this. So, so that's to say the meanfield may may work and this framework and discussion that this paper is about recovers the meanfield case. However, this discards a structure that is visible in natural behavior and engineered agents. Action taken on one stream are routinely contingent on, anticipatory of, and instrumentally coupled to actions on others. A drummer's left hand does not freely sample its policy distribution while the right hand plans the next fill. An autonomous vehicle's lateral and longitudinal controllers share predictive structure with a symbolic route planner. A human reaching for a cup while reading does not factor reach policy and

secade policy independently. So imagine like in the moment of reaching even if you have a model where there's some higher order decision to reach or not and then it's going to send a descending prediction to the cicade and to the um arm. That's one way to model that situation and then have the cicade or the eye movement planning and the mo motor planning rollouts happen independently of each other and only uh imping but then intersectional effects of those two could have either conditional impact on policy's outcome. So like if a certain muscle movement is happening then this cicade movement is made different in this way then what the ciccade only roll out might think has high epistemic or pragmatic value may be conditioned upon by changes to to the state or the the causal situation

in the world even of the B kind of transition matrix. So the the causal effects of different policy streams can intersect and and that's kind of on the where where it could go wrong by ignoring this and then where it could kind of go better by engaging this is there might be some lower dimensional manifolds or some mutual information or some heristic that like simplifies the policy inference for one or both if they were able to take into account their co- relationship like their their their up their upstream causal um basis which is captured by the simple factorized meanfield model where there's just a top level decision and then it's going to dispatch down to the cicade and the motor and then here is exploring well what about the cicade and the motor policy streams

having a certain connection with each other. Um, this document develops a tunable, analytically tractable, machine checkable formalism for the parametric entanglement of concurrent policies in a single agent. And where this is going to go is a multi-art Oh. Oh, these are the only literal. These are the literal tokens. But this is the one special time where says if any of these symbols are not found, then there was a bug and not the good kind in the gating of the production of this PDF which is the fourtrack gates. Um the I'll go to the four track gate. Let's republish with that. Um the first part is the background and and theory. Then there's a PIMDP simulation and a set of lean sketches sketches in this active policy entanglement lean folder. So there are uh some visualizations

produced by the simulations as well as on the lean lean outcomes outcomes which I did my best and still would seek a lot of input on how to make that lean environment and reporting and other sort of typing based analyses and uh formalizations formalizations even the DSPY type neurosymbolic um loops how how to make that setup and reporting reporting most effective. Here's the gates. Um the uh this relates to the sort of triple play which is the uh claim either being represented by or uh at least assembled from or spoken to by a formal proof layer, a typed API schema layer, a numerical layer, plain text and visualization And this uh structured document which as with a textbook but in a in a very different way way um um seeks to make a structured document

that in its 170 pages I wouldn't say is necessarily simply the best linear reading. It's a little bit like a little bit uh like not the the linearization of this is shown as a demonstration of capacity and possibility in an artifact where ideally we are making these kind of bookend type first and last page and throughout throughout strong strong guarantees or or bounds or or marks on the production. function of the document because maybe a lot of documents we will engage with a a cousin artifact and just referencing the LLM to a repo with that information already in the most structured way where you can rerun different parts of the repo and and transform it and project it into your other um codebase or implementation. Maybe you did it in some other language or not

PMDP but it's all modular so that the and and worked at from a few different perspectives with the pros executable structure generated visualizations all those kinds of things can translate really well and kind of be that ontologically ontologically plural plural um or or is it onlogically singular with the different faces again kind of referencing something similar. Not always, just these different media representations representations uh uh seeking to make the same claim, especially in in types of information conveyance that that different um textual and visual and typographical information transfer can can share a lot in these different artifacts. And And I guess that will close unless anyone writes a question. But just to summarize what this hopefully short, hopefully kind of relevant uh curious stream covered is the template repo which if you want to just

contribute, we can start to explore what uh h how to have good repo hygiene uh with multiplayer there or take it and use it as it is or make some other first principles take on this sort of shape. But this repo has the template projects which are overdocumented and help you set up a uh uh essentially one promptable transformations of these project archetypes into endtoend run and rendered reports. And if you use things like perplexity API API through through a variety of of possible surfaces then you can get research reports. You can modify the literature review methods and uh then showed these two recent publications. one which was yesterday of the generative biology textbook kind of template/system and then one that was during the stream where the GitHub and Zenot were both pushed in uh

in in aentic aentic operation. uh people looking into auto research, check out the template auto research research project project and see if um there's any interesting motifs there. So, So, hope that's uh interesting. Good luck with your research and see you on the channel or wherever.