graphs.sgit.ai → the second edition → The memos

The memos

The founder's voice memos, transcribed and reproduced verbatim, because the house rule is that the founder's voice is source material and is not edited. Each one is followed by the instructions extracted from it and, for each instruction, what it commits the work to. That second half is the agent's reading and is marked as such.

Where the earlier ones are. Briefs 00 to 19 belong to the first edition and froze with it at v0.3.26. They are readable at /v1/documents/ and raw at /v1/briefs/. This section holds the briefs written since, and the numbering continues rather than restarting, because the corpus is one sequence even though the editions are not.

#	Memo	What it gives you	Raw
20	Build the universe first, then find the plot	The memo that inverted the dev pack's construction order. The pack was defining answers before the questions were known; the universe of concepts, claims and evidence has to exist before any altitude can be decided. Also corrects what fractal semantic graphs means, and reverses the verdict on Wardley maps.	.md
21	Review packs, and briefing other agents	A website cannot control what a reviewer reads or in what order. A pack can: one continuous page and a PDF printed from it, read end to end on an iPad or on paper. The specification for the pack family, and a correction I owed about visualisations.	.md
22	The universe viewer, and where it goes next	Feedback on using the reader in earnest: nodes as readable boxes, the doc tree as navigation, families as selectable node packs each with its own peak, stronger and weaker links between concepts, paths to the peaks, and a freeze-and-grow workflow. Twelve instructions mapped, four questions put back to the founder.	.md
23	Visible links, live physics, and exploring the graph	Three notes sent while brief 22 was being built: every link in the source visible and toggleable from the pane itself, one toggle set driving both panes, physics applied as the slider moves, and the explore workflow — focus on a selection, grow it degree by degree towards the peaks, with stats that price the next hop before you pay for it. Answers two of brief 22's questions; three remain open.	.md
24	A document grown from one node	The experiment after the reader: pick one node and compose the document of that concept from what the anchored data verifiably holds — programmatic first, no authored prose, on its own page, printable. The book workflow run from a single word upward; the whole book is the same operation on steroids.	.md
25	Pinned peaks, and the document as a source	Lock the summits at the edges of the canvas and let the physics settle everything else between them: doc root and family peaks left, derived-group summits right, hand-draggable between layout runs. Also: the document becomes a source like any other (all sources off = empty canvas), the derived groups get their own peaks, and the maximised graph finally owns the whole viewport.	.md
26	Fixed nodes, the four areas, and the schema	The stability principle: every node move costs the reader their mental picture, so what is on canvas holds still while newcomers settle. The four border areas with their aligned slots, the drag-and-drop peak board, the maximised view's inspector and type legend, the invisible alignment lines to come, and the schema view that judges the graph at the type level. Nine instructions mapped, four questions back.	.md
27	A narrated walkthrough of the viewer	The first narrated review: ten screen captures with the founder's words spoken over them, words joined to pixels the way anchors join claims to bytes. Six findings connected to their screens — the live graph one click away, a clickable type legend, the explore walk cleansed of the rails' invisible ties, and the peak board docked so the canvas it re-anchors stays visible.	.md
28	Node navigation, reverse verbs, and the path query	Click a node and see its whole universe on the right: every link out with its verb, every link in read through the declared inverse, each row a hop that extends the path trail — the beginnings of the path query. Plus the scale framing (nodes will grow near-exponentially) and two pane bugs fixed where they pointed.	.md
29	The core graph: document to word	The foundational graph under everything else: transform the document all the way down — sections, paragraphs, sentences, words — like an AST driven by the content, every level a node with an ID, because pointing at character 256 breaks the moment the document changes. Bold and links become span nodes, every distinct word gets a counted form node, the tree expands bit by bit from sharded files, the controls move into tabs, and the inspector shows everything known about a node.	.md
30	Words as tokens, and the two-way transform	Each document is its own token universe: every word a unique entry with no vocabulary cap, classified (padding, verb, content), stemmed into families, scored for different-meanings-in-the-same-document, and connected by co-occurrence into the document's own attention map. Plus the instruction addressed sooner more than later: the graph must rebuild the markdown byte-for-byte, with formatting captured in its own graph beside the semantic one.	.md
31	The WCLM: a deterministic transformer over our graphs	The crazy experiment, commissioned: a mini engine in the shape of a transformer where nothing is learned and everything is named — tokens are content hashes, the layers are deterministic transformations with declared roles, attention is the co-occurrence and concept-binding made visible, and the query flips from predict-the-next-word to what-does-this-mean, answered with provenance, blast radius and weights instead of probability. Training is tweaking graph inputs, never fitting numbers. Meaning packs bring the world above the document: QA is part of development.	.md
32	Every box explains itself, both ways	The WCLM verdict (it worked) and the next rounds: every chip at every layer clickable, opening its reason-to-be — caused by upstream, feeding downstream, because it's graphs all over. A side pane for the explanations, layer cards for the layers themselves, run-to-run impact made visible (adding one word made a massive difference and the engine should say exactly how much), and example buttons from strong to weak connectivity. Plus the direction: one of these engines at every abstraction jump of the book, rationally explaining each compression.	.md
33	The detective playbook: a narrated review of the WCLM	Eight narrated moments over the deterministic transformer, and the finding that restructured it: layers must not be jumped — every layer adds or drops evidence, and clicking the answer must show every piece of evidence that got there, transitively. Plus the live experiment that became a block: 'meaning without connectivity' returned the same winner and rightly should not — so operators (without, not, no) now flip binding and surface contradictions. The dictionary and thesaurus move into the early layers, the layers become reusable mix-and-match blocks with a drag-to-order bar, and every weight is labelled opinion or evidence.	.md
34	Words have many meanings: senses, number, and the world model per word	Each word needs a world model: singular is not plural (graph is not graphs, and 'graph of graphs' is not 'graphs of graphs'), and the same word means completely different things across industries — a graph is a network graph here, a chart in a boardroom, a function plot at school. The thought exercise: author three to five senses per word, let people switch to THEIR definition, and programmatically show what stops applying (a graph that is a diagram is not fractal). Two addenda from the same day: layers must hold MULTIPLE engines side by side (pass-through itself an engine, include it or block the stream), and every engine must declare its input/output schema — including a fractal mode where one engine is a full WCLM inside.	.md
35	The fractal nature of the WCLM: world models, analogies, anchored facts	The engines are world-model transformations — every layer asks what does this mean, expand it, contract it, give me more so the analysis continues. The destination: ask a document, ask a paragraph — here's the graph of where I'm going; does it agree, provide evidence, reach the same conclusion? New and named: ANALOGIES — to explain this to somebody from finance, graphs of graphs must become spreadsheets of spreadsheets, because their world really nests them. Corrections are the training (better meaning, missing nodes, manual overrides — what pre-training is to an LLM). The LLM layer-warming picture, redone with determinism: our lines are exact, so facts and hypotheses can anchor to them. And some layers may one day need an LLM — acceptable, because graph in and graph out are both kept as evidence.	.md
36	Operators as first-class folders: tune each one individually	The file-explorer treatment, applied to the engine's own building blocks: every operator in a dedicated folder with its code, its md, its html workbench and its json; schemas as json files; examples and sample data; official operator data marked standard, authored or derived (from another operator or transformation); a reusable UI to execute, test, debug and visualise input, transformation and output; and visual representations of architecture and execution, js-drawn and ascii-drawn.	.md
37	Keep zooming: the code itself gets the graph treatment	Graphs of graphs, one zoom further in: the operator scripts are small but context-heavy, so give them the Bret Victor treatment — group the parts of the code visually, explain each part on a right pane (what it does, what the variables do, the inputs and outputs of the inner bits), and draw the architecture as flowcharts and fluxograms, for an audience that knows JS. Apply the main WCLM's visual linking of components, data and ids to every operator — each one a mini app with its own js and css for experiments. Try it on the json files too. And frame all of it as the lab: PoC ways to see, run, visualise and debug in the small space, then promote the ones that worked into the main WCLM UI and workflows.	.md
38	Three books from this estate: the book-writing pack	Something a little bit different: not a build round but a writing round. Three books from this repo and the work so far, each in its own fresh session, read on tomorrow's long flight — the universe/atlas volume for 'Fractal Semantic Graphs: Meaning Through Connectivity', the book itself (title locked, full editorial freedom elsewhere), and 'Creating a Book Using Fractal Semantic Graphs' for other authors: the story of book 2's experimentations, PoCs with screenshots of the evolution, and the agentic workflow — an expansion of the v0.4 retrospective. Delivered as one dev-pack-style pack with three entry prompts, since most of the briefing is shared.	.md
39	Park the engine, ship the books, open the review era	Recorded after reading the books on the flight, and it opens with a finding: the making-of book's alignment is stronger by orders of magnitude than the FSG book's, which is the power of good source material. Then the turn: a proper non-functional pass over the code, the tests and the harness; the WCLM and its operators PARKED, because the experiment needed a better target; the two books shipped all the way to Leanpub, because shipping is one of the most important things you can do; the site refactored as the home of two books; and then the real work — the review process itself, run as change control with human AND agentic reviewers (plan, map, define, review, approve, implement, approve the implementation). Plus per-book versioning: the repo's version moves on every push, a book's version moves only when its content changes.	.md
40	The agentic team, and who the making-of book is for	The first memo of the v0.6 era, and it opens with a verdict rather than a complaint: the making-of book's content, voicing and pacing are right, and what follows is a change of FRAME, not a rewrite. Then the machinery: an agentic team as folders, one per role, each with a role.md, actions, briefs, debriefs and a work environment, so personas can be spun up in parallel and kept isolated — not for throughput but for judgement, because “agents advocating for certain things, who have specific centres of gravity” give better decisions than one generalist. Seven roles named, every one of them customised to THIS estate rather than generic. The first question put to the team is the book's own title, which the founder judges wrong: the making-of does not use the fractal element, it is about the workflows. And underneath it, the two audiences in Wardley's terms — the villager who wants to write a book with productised tools, and the explorer who wants to see how the sausage was made and change it — plus the thesis the book may really be about: the scaffolding matters more than the code, and you build the ship while preparing to fly.	.md
41	Refactor the structure, and formalise the workflows	Three things at once. The publishing order flips — the making-of book goes FIRST, because it is more mature, has a market now, and carries the workflow ideas, while the graph book “needs the technology advancements that we're going to do in the making-a-book”. A restructure is proposed around books at the top level, with v1 and v2 renamed v0.1 and v0.2 as versions of the graph book. And the beginning of a compatibility model for technology shared between books. The agent's reading agrees with the principle and disagrees with the restructure on measured grounds: v1 and v2 are timestamps, not types, and 172 of the 200 frozen files sit outside the book folder.	.md
42	The craftsman makes the tools, and the book should lead with them	The making-of book's thesis, stated outright: you cannot separate the creator from the technology, workflows and scaffolding they use. Brian May built his own guitar because craftsmen always did. The founder is never a passive consumer — the power of generative AI here is not writing words but building the environment that makes the writing possible. With the cost named honestly (“you still need really good engineering”) and a limit placed on the method (“I don't think it's realistic to say that somebody without programming experience could provide the prompts I do”). Then the instruction: part one of the book becomes a feature tour — the art of the possible FIRST, the history second, because that is the order that motivates a reader. Plus a new requirement: pages designed for agents to consume.	.md
43	Everything is a graph, so put the book through the same machine	The philosophy stated at full strength — everything is a graph, the unit is whatever you choose, even an air gap is representable, and a graph's maturity is how easily it links, changes and transforms. Nature is the model, and the point is aggregation blindness: an atom does not know it is made into molecules. Then the instruction: take the decomposition built for one pilot document and run the whole book through it, so JSON becomes the source of truth and everything else is a transformation. The agent probed that claim before answering — the machinery does run over all 17 chapters (165 sections, 819 blocks, 26,118 words) and three strains are already visible, including a level the generator does not have. Ends with the constraint the WCLM taught: no technology without a named output.	.md
44	The Leanpub story, the title, and the determinism boundary	Answers three things at once and adds a concept. The Leanpub story the agent asked for turns out sharper than expected: the earlier workflow was already good — markdown, GitHub, reuse of his own blogs — and what failed was not the writing but everything after it, because “you almost start to be locked by the first version of the content, because making changes becomes quite painful”. The title is decided: Creating a Book Using Agentic Workflows. Restructure before publishing. And the concept the founder says he has not explained well — source materials versus projections of them, two-way doors against one-way ones, prose being a one-way door because a model wrote it, and the rule that follows: keep the sources in the deterministic layer so the outputs can be destroyed and rebuilt without losing the core.	.md
45	What a workflow is, and the measure of success he judges maturity by	The definition, stated plainly: a workflow is “how you operate”, the steps you actually take, and opening a word processor is one. What changed in 2026 is that a workflow can now change its own tooling, visualisation, comprehension and context, and the purpose of changing them is to “allow the author to stay in the zone”. Villager and explorer are phases rather than kinds of person: experience changes how deep you can reach into a workflow and how fast you read, not the steps you take. And the metric he uses for maturity, which inverts how a busy week reads: “can I do a process without making any changes to the workflow?” The memo also sets the phased experiment this document came out of, and asks for a diff between two versions of a book that is computed from the graphs rather than the markdown.	.md
46	The reader who followed the link, and the front page that lost him	A LinkedIn comment made the failure concrete: a reader had never heard of fractal semantic graphs, went looking, and was pointed at this site. The founder’s objection is one sentence — “that reader will be lost” — and the measurement bears it out: the phrase appeared four times on the front page, every one of them inside a book title or a filename, and nothing defined it. The page also had no image, no diagram and nothing to click, while arguing that meaning is visible through connectivity. The instruction is a split rather than a rewrite: move the archive to its own page whole, and let the front page answer the question people actually arrive with.	.md
47	A sibling site's agent reads this book and reports the claim backwards	The first reading of this work by anyone outside the estate, and the first brief here that did not come from the founder. sgit.ai published its own Fractal Semantic Graphs page, read chapter 6 while writing it, and sent back a correction: what survives a zoom is the GRAMMAR, and the ontology is meant to change at every altitude, so a system whose types and verbs are identical all the way down is a hierarchy rather than a fractal. That is right and it was actioned. Two things the brief could not know are recorded with it: the book was quoting the corpus rather than contradicting it, and had simply never carried the corpus's own definition of “rules”; and this estate had already made the same correction in another book on 23 August and never propagated it. Four of the brief's seven items ask for frozen first-edition pages to be rewritten, which is this site's failure to make the freeze discoverable rather than the brief's mistake.	.md

For an agent

The raw markdown under /briefs/ is the source of truth and each page here renders one client-side. Inside each file, the block quote is the founder verbatim, including transcription artefacts, and everything after it is the agent's reading. Do not attribute the second half to the founder. Where a memo corrects an earlier document, the correction is recorded beside the original rather than replacing it.