# The Implementation Plan **version** v0.3.27 · **date** 23 August 2026 **status** PROPOSED. Every phase ends at a gate. A phase is not done when the work is done; it is done when its gate passes. **releases** Phase 0 ships as **v0.4.0**, and each later phase takes the next minor version, so the release number says which altitude exists. --- ## The shape Seven phases. Phases 0 and 1 change no book text at all, which is deliberate: the plumbing and the freeze have to be real before any writing depends on them. **Revised 23 August** by [brief 20](../../briefs/20__founder-memo__the-universe-first.md). The first draft had the spine in phase 2 with evidence attached to it afterwards, which is defining the answer before the question is known. The order now inverts: **the universe first, the plot lines from it, the spine from those, and only then the levels.** ``` 0 FREEZE the first edition, hashed and gated no writing 1 PLUMBING the graph format, the generators, the gates no writing 2 THE UNIVERSE every concept, claim, fact and piece of evidence no prose, no spine decomposed from the material we already hold 3 PLOT + SPINE plot lines drawn through the universe; the spine no prose DERIVED from where they converge 4 L1 + L2 the claim, and the compression writing starts 5 L3 the pivot: most connected to the graphs 6 L4 the expansion 7 L5 the prose, which is the book gated, opens after 6 ``` Note what moved and why. **L5 is the prose, not a reference layer** (the memo's instruction 9), and the reference material is not an altitude at all: it is the universe from phase 2, sitting underneath all five. **L3 is where the action is**, so it gets its own phase rather than sharing one. Phases 4 to 7 each produce a **complete, publishable book**. That is the fractal property doing work: there is no phase after which the project has a half-finished artefact. --- ## Phase 0 · Freeze the first edition · **DONE at v0.4.0** **No book text changed. Nothing was deleted.** It was executed as a **move** rather than a copy, per the founder's answer to open question 1: everything that constitutes the first edition now sits under `v1/`, `v2/` exists and is empty, `v1/MANIFEST.json` records 236 files and 10.9 MB, and gate 14 was verified negatively before the phase was called done. Ninety-three redirect stubs hold the former addresses. Two PDF URLs could not be preserved. 1. Copy `content/`, the ladder data and the four review files into `books/01__first-edition/`, with a `MANIFEST.json` recording every path, its SHA-256, the release (v0.3.26) and the commit. 2. Record the hashes of the frozen published pages: `/book/` and the sixteen chapter sources. 3. Write `books/01__first-edition/README.md`: what this is, and the rule that it never changes. 4. Add **gate 14** to `validate.js`. 5. Write the front page described in file 03. Generate its sequence-of-events section from the release table; write the four narrative sections. **Gate 0:** gate 14 passes, and it has been verified negatively by changing one byte in a frozen file and watching the build fail. A gate that has never been seen to fail is a gate nobody has tested. --- ## Phase 1 · The plumbing **Still no book text.** 1. `graph/registry.json`: the edge vocabulary, seeded from the existing edge registry. 2. `graph/nodes.json`, `graph/edges.json`: empty but schema-valid. 3. `gen_graph.py`: validate, resolve, compute. Emits `graph.json`. 4. `gen_levels.py`: project units to pages, both forms. 5. `gen_checks.py`: the five existing checks, ported, plus `theme-has-converging-evidence`. 6. `gen_grammar.py`: the edge set chapter from the registry (ADR-5). 7. Gates 15 to 21 in `validate.js`. 8. One graph module (ADR-7), and the four existing graph pages repointed at it. **Gate 1:** the whole chain runs green over an empty book. An empty book that builds is worth more than a full book that does not, because everything after this depends on the chain being trustworthy. --- ## Phase 2 · The universe **No prose, and no spine either.** This is the phase the first draft of this plan skipped, and skipping it is what let a spine be declared before anything was known. The material already exists and the work is decomposition rather than research: | Source | What comes out of it | |---|---| | 21 carried documents, 63,038 words | concepts, claims, the arguments made and their evidence | | 6 vault analyses and the capability scale | facts with numbers, and demonstrations | | the concept map, 24 concepts | definitions, also-called, near-but-not, existing typed edges | | the first edition, 17 units | its claims, its findings, its 8 open contradictions | | the Wardley material in the Send corpus | the sixth spine candidate's missing evidence | 1. Decompose into the six families of ADR-3. Every node typed, every claim carrying its evidence state, every evidence edge resolving to a carried source with a hash. 2. Do **not** decide what belongs at which altitude. That is phase 4 onward. 3. Build the evidence view and the concept view, and look at them. This is the first moment the second book can show something the first could not. 4. Re-test the audit's hypotheses (file 02) against what the universe actually contains, and record which ones moved. That number is a measurement of how much the universe knew that the guess did not. **Gate 2, the stopping rule.** Building a universe has one obvious failure mode, which is that it never ends. The rule is computable and is checked rather than felt: > **Every candidate plot line can be traced end to end through the universe without a gap, and > every spine candidate has converging evidence or is explicitly labelled an opinion.** The universe is not trying to be complete. It is trying to be **sufficient for the story we are choosing to tell**, which is a property of the plot lines rather than of the material. --- ## Phase 3 · The plot lines, and the spine derived from them **Still no prose. This is the phase that decides whether the method works.** 1. Draw plot lines (ADR-9): named, ordered walks from a question to a resolution, each with its turn. This is the whodunit: the sequence, the pace, the punchlines. 2. Look at where they converge. **The spine is what several plot lines pass through**, not what was written down in file 04 three days earlier. 3. Compare the derived spine against the five candidates. Some will survive, some will move down an altitude, and at least one should be something nobody listed. If all five survive unchanged, the derivation is not doing any work and the check is too weak. 4. Run `theme-has-converging-evidence` over the derived spine. A statement without converging evidence is demoted, re-evidenced, or labelled an opinion. All three are legitimate; asserting it anyway is not. 5. Run the pacing check (ADR-10) over each plot line, before any of them becomes a level. **Gate 3:** the derived spine passes the evidence check, the plot lines pass the pacing check, and **the founder recognises the spine as the book he intended**. That last one is not computable and is stated anyway, because it is the real test and it names who decides. --- ## Phase 4 · Level 1 and level 2 Writing starts. Roughly 150 words, then roughly 1,500. 1. L1: one paragraph, projecting the spine. It must contain the title's own words, which is the test the first book failed (ladder finding 4: compressed honestly, its top paragraph had no room for "fractal"). 2. L2: five sections, one per spine statement, each ending at a node L3 will expand. 3. Both forms: hyperlinked and reading. 4. First release of the second book. It is complete at its altitude. **Gate 4:** every L2 node declares what it compresses; those targets do not exist yet, so the gate is deferred to phase 4 and recorded as a **known open edge set**, published as such. A book that says which of its edges do not yet resolve is more honest than one that waits. --- ## Phase 5 · Level 3 The grammar, the method, the worked examples, the honesty table. Roughly 8,000 words. 1. Carry chapter 3 (the rules) and chapter 9 (2FA), placing them at nodes first. 2. Generate the edge set chapter from the registry. 3. Generate the honesty table from the claim nodes' evidence states. 4. Write the method chapter: this is new, and it is the retrospective's material. The book explaining how it was built is not an appendix in a graph-first book; it is the argument. 5. Close the L2 to L3 descent edges. **Gate 15 now applies in full.** **Gate 5:** gates 15, 16, 18 and 19 all pass. Every L2 node has children. The honesty table is generated and therefore cannot be wrong in the two ways the first book's was. --- ## Phase 6 · Level 4 The objections, the history, the failures, the tensions. Roughly 25,000 words. 1. Against schema-first, with its concession stated early (memo 1.6). 2. The fractal argument in full. 3. The evidence estate: the vault analyses distilled, which is what the case-study programme was for. 4. Lineage, participant disclosure, what this loses. 5. Wardley maps: no longer a decision to make here. The verdict was reversed on 23 August and the material is a spine candidate, so its altitude is settled in phase 3 with the rest. **Gate 6:** the whole descent resolves L1 to L4. Every claim has an evidence state. The inventory of facts, assertions and opinions is derived, not typed. --- ## Phase 7 · Level 5 **Gated closed until phase 6 passes.** Memo 2.11: level 5 arrives once the graphs are agreed and mapped. The concept register with full definitions, the path atlas, the carried sources, the measurement tables, the complete findings register. **Gate 7:** the machine surface is complete. `graph.json` publishes every node, edge, evidence hash, peak with its formula, and check with its rule. An agent can answer any question about what the book claims without parsing a sentence. --- ## What runs alongside, continuously - **The review workflow.** The second book is reviewed the same way: reviews become `r005` onward in the same register, and the decisions page spans both editions already. - **The estate.** `/vaults/`, `/docs/` and `/documents/` keep moving and keep being cited. - **The release discipline.** Version bump, versions row, gates green, commit subject, tag. --- ## The two ways this plan fails **It gets to phase 4 and starts writing prose faster than it declares nodes.** This is the likely failure and it will not announce itself, because the prose will be good. The gate that catches it is 20 (every unit carries a provenance block) combined with 15 (every node has a parent), and the discipline that prevents it is Rule 1. If a session finds itself writing a paragraph before the node exists, it has already failed and should stop. **Phase 3 produces an uncomfortable answer and it gets softened.** If a spine statement has no converging evidence, the temptation is to add evidence until it does, which is the graph equivalent of p-hacking. The defence is that the evidence edges are authored, published and dated, so an edge added to rescue a statement is visible as one. It should still be named as a risk, because the person adding it will be the person who wants the statement. --- This document is released under the Creative Commons Attribution 4.0 International licence (CC BY 4.0).