Every Choice Was Right at the Time
Good choices, built on past the end of their reasons
Prepared for John Kershaw | November 2026
John Kershaw, who owns Harbour, put it this way on 14 November: he assumes every choice was right at the time, and the mistake was to keep building on a choice after it had outlived its intention. The cure he pictured is to map the code, find what is held in the wrong shape, have a person rule, and rebuild; that, he said, is the pattern built into Harbour's loop, and it is fractal. The record bears most of this out. Harbour's mess is better read as good choices built on past the end of their reasons than as bad choices, and that shape recurs from a single line of code up to the loop of agents meant to clear it. The record also separates two acts: tidying inside a choice has bought percentages and often added code, while the large removals on record each took a choice away, and each retirement this essay traced to its start began with a word from John. The cure repeats too, at the scale of a task and of an area. But the record falls short of the framing in places. Taken as a finding rather than an assumption, "every" fails: some choices were wrong the day they were made. Nothing shows the clearing is easy. The hundredfold John imagines for the front end is unmeasured, and nothing on record comes near it. As built, the loop puts clauses (single lines of a ticket's text) to the person and settles choices (the standing decisions about what the code is for) itself. And the loop's own scaffolding ran up the same debt inside two weeks; turning the cure on the loop is still a plan.
Harbour is a web app that reads a team's issue tracker, writes a prompt for an AI coding agent from each ticket, sends the agent off and checks what comes back. Agents built almost all of it, and Harbour's own tickets go through the same pipeline [H-method].
One day in November
The day began with a stall. Harbour runs work in lanes, each a run of related tickets in one area of the code driven by one supervising agent session. Overnight the lanes had stopped, waiting on rulings, and the questions were not John's kind. (A ruling, in this essay, is John's answer to a question about the work. The loop asks by putting a card in front of him with a suggested answer; he also rules in comments and in conversation.) One question came from a measurement written into a ticket's Done-when, its finish condition, so a reading meant to inform had become a gate. The area notes, a prose file per area of the code where agents record lessons and a "Waiting" list of known work, had grown from about 11 KB to 89 KB in six days and were acting as a second backlog. John: "a rule added by an agent, becomes a rule that blocks or causes more work to be done unnecessarily." [J] [LOG]
A fresh read of the history of the steering text, the prompts, rules and notes agents read before they work, found that changes to the process which added a step or a rule outnumbered those which removed one about six to one, and that the text had shrunk mainly when code took over its job. John: "How do we get out of Groundhog Day?" The survey session, the long-running agent session that drives this work with John and keeps a log, answered with removal and, in the same reply, proposed three new rules. John: "I think we're seeing the issue." Then: "extra scaffolding and temporary structures are fine. We just need to plan their removal." [J] [LOG]
From there the talk widened: to keeping the notes as checked data rather than prose ("simply saving the notes as TS, so they get checks"), to how everything Harbour holds, "code, comments, docs, notes, papers, tasks, logs", should be kept in a shape, and to tangled code: "if everything is in its right place, no tangles." [J] A companion essay of the same day, Nothing fights its shape, takes up that second question [H-shape]; this one takes up the three statements that followed.
The day produced three papers (steering text [H-steer], beyond the code [H-beyond], lanes across the seam [H-lanes]). Then John said three things [J]:
"I imagine Harbour looks at the code, maps it out, then figures out what's holding complexity or mess simply because a square peg is trying to go in a round hole. Eg in hindsight our UI could be 100x simpler, and get the same results. The right shape for the right choices, and that's the Theseus method in full effect."
"I assume every choice was right at the time. The mistake was continuing to build on it when it had outlived it's original intention. But, prior to the past few weeks, that wasn't a natural part of Harbour's process. Now we understand how it happens, and for the first time we have a mechanism to clear that tech debt easily. It requires humans to rule (eg switching to a react front end, or htnx, or ejs), but that's exactly right."
"That's also the exact pattern we've built in our loop. It's fractal."
The Theseus method is Harbour's way of renewing its code while it runs: find each decision scattered across the code, decide which area owns it, gather it there, then rebuild it from a fresh design, with every change to production code judged by a replay that runs the old code and the new on the same inputs and compares what an outsider would see [H-method].
Right at the time
"I assume" is the right word, because it changes the question. "What is wrong with this code?" invites a list. "What was this for, and is that still true?" has an answer, and the record usually holds it.
The idea is as old as the debt metaphor. Ward Cunningham, who coined it in 1992, named the pitfall "the failure to consolidate" (C92), and later said what he had not meant: "I'm never in favor of writing code poorly, but I am in favor of writing code to reflect your current understanding of a problem even if that understanding is partial" (C09). In its first sense, debt is yesterday's understanding still standing. Martin Fowler, on eBay: "The right architecture to support 1996-ebay isn't going to be the right architecture for 2006-ebay" (F14).
Harbour's front end is the worked case. Its first commit, on 4 January 2026, was a read-only viewer of one issue tracker's projects as a collapsible tree: twelve files and 1,112 lines outside the lock file, 912 of them in six code and style files, with one browser script of 188 lines. Its instruction file for agents, CLAUDE.md, carried one line among the design principles: "Keep it minimal - no frameworks, no build step" [H-first]. No reason was written beside it. None was needed.
Other choices have their reasons on record. The code that reads issue trackers took Linear's wire format as its shape when Linear was the only tracker [H-theseus]; prompt sentences went in where something had failed [H-steer].
Taken as a finding rather than an assumption, though, "every" is too strong. Some choices were wrong when made. On 8 November Harbour built a settings switch from a brief that had turned a passing answer of John's into a product mechanism; meeting it, John said: "technically it's correct, but it's also obviously wrong and a mistake." The paired area read, a blind before-and-after reading of a sample of an area's files, was written into two lanes' finish conditions, and the session's own review later found: "The kit had no power to detect a change, so making it a gate guaranteed a ruling." [LOG] A consolidation in October put a shared helper in another area and so re-tied a two-way dependency between areas five days after a deletion had untied it [H-theseus].
So "every" is a stance, not a finding, but a useful one. G. K. Chesterton's reformer, finding a fence across a road, says "I don't see the use of this; let us clear it away," and is told to "Go away and think." When he can say what the use was, "I may allow you to destroy it" (CH). The assumption sends you to find the use. John's framing adds the step Chesterton leaves out: asking whether the use has ended.
Building on it
On 14 November the front end had 44 browser scripts and 35,236 lines, about 187 times the founding script [H-first].
Much of it stands because of the founding line, each piece a sensible answer given the line. Research in June called the hand-written components, built from template strings, "the deliberate, build-step-free answer to 'are we inventing our own framework?'" (H-482). Reading the rule as "no shared module between server and browser" produced eleven pieces of server logic copied by hand into browser scripts. An October paper calls such copies twins, the same logic kept in two places, and named the founding line their "largest single producer" (H-coherence).
This is John's second claim, and the best supported. Once a choice is fixed, every later choice is judged against it rather than against the need it once served. Each is locally right, and the misfit compounds. The complexity sits not in the square peg or the round hole but in everything built to hold one in the other. The code that reads issue trackers (Harbour calls each adapter a provider) shows it too: "In practice the shape is Linear's GraphQL wire format, and the other providers rebuild it in their own normalisers" (H-theseus). So do the area notes, which were meant to save minting a ticket for every small debt: within six days five kinds of writer were appending to them, a validator and a quota policed them, and one repository's Waiting list went from 30 entries to 118 [H-beyond] [LOG].
None of this was for want of noticing. The front end's rule was questioned three times and kept each time. On 16 March a plan accepted a copy that "will be embedded/duplicated in the client script since there's no build step" (H-swim). On 15 June research weighed "ejs / htmx / React" and said no; the direction "was decided with the owner", the rule "was reaffirmed for now", and htmx "has real merit" but was "Deferred to Step 2" [H-482]. A later review: "Step 2 was never ticketed." (H-debrief) On 8 October the twins paper said: "Browsers load ES modules without a build step, so the rule does not require the copies; the reading of it does." [H-coherence]
"For now" had no date and no sign to ask again, which is typical. None of 27 notable process changes since June named a condition for retiring it, and the one sunset written into a process step was never read (H-land). The steering-text paper gives three reasons temporary structure stays: "Nothing notices that its job is done, nobody owns removing it, and sometimes nothing could remove it even if asked." [H-steer] The front end adds a fourth: its reason was never written down, so nothing could have checked it. Continuing to build is not an error someone makes. It is the default, because a choice arrives with a beginning and no end.
Tidying inside a choice, and retiring it
"100x" makes sense once two acts are told apart: tidying inside a choice, and retiring it. The record separates them sharply.
Inside a choice, tidying has bought percentages and often added code. Of 57 commits whose titles say cut, remove, simplify, fold or retire and that touched the steering text, 37 left it larger, and trimming single rules "has never moved the total by as much as a kilobyte" [H-steer]. The October charter, a four-day programme to "simplify the code, unify the UI, and update the pages" in John's words (H-charter), ended with production code 9,725 lines larger; byte-identical moves took one server file from 3,782 lines to 882 while the directory around it grew from 26,955 to 28,559 (H-retro).
That is size. By other measures, tidying does better. Harbour's Periodicals are scheduled independent reviews of the codebase, and the two run on 14 November read the tidy lanes of 13 and 14 November. The Stability Review counts the tidy programmes at "net +781 production lines" and calls the pattern "replacement churn", but its headline is that the wave "is converging on the things it aimed at, and churning hard to get there". The blind reads it counts, a fresh reader scoring each change from its code alone, came out 20 better, 2 same and 0 worse, as the loop recorded them [H-stab]. The Drift & Coherence review reads the window as "better", with seven earlier findings resolved [H-drift]. Only the paired area reads found "no detectable difference", from samples that mostly missed the changed files (H-reads). None of them shows the code smaller.
Rebuilding inside a choice does better still on shape. By size, three design-first rebuilds cut code "by up to 39% in a file" (H-ladder). By complexity, roughly the number of paths through a function, ten rebuilds over three rounds in October cut the most complex function in each file by between 29% and 88% [H-method]. The largest multiple came when the scattered decisions were gathered first. The endpoints that start agent runs held 24 shared decisions; a rebuild on its own took their worst function from 122 to 63, gathering the decisions into one place took it to 29, and a design-first rebuild of that place took it to 13 [H-method]. The gathering also cut the places holding those decisions from 95 to 26 and took 317 lines out of production code (H-consolidate). That is a multiple, on one function's shape, and it came the way the others did: copies were deleted.
Retiring a choice is a different order of thing. Until October a model wrote the prompt for each stage of the pipeline, working from a 106 KB "meta-prompt" and the stage definitions, and from June to 7 October 53 of the 63 pull requests that changed the stage definitions (84%) had to change the meta-prompt too. On 4 and 5 October that path was deleted: "the 106 KB meta-prompt is gone, the stage definitions shrank from 132 KB to 110 KB, and an 8 KB selector took its place" [H-coherence]. The first deletion took a net 1,012 lines out of 27 application files (H-paths). Deleting the thing a passage described accounts for 74% of all the real shrinkage in the steering text that agents carry on every run [H-steer]. The charter's clearest win was a deletion too: one superseded page removed, 7,681 lines [H-retro]. Two-way dependencies between areas, which Harbour calls tangles, tell the same story: 24 born from January to mid-October, 5 dissolved, and "All 5 dissolutions took something away." Six of the 24 births came from consolidations [H-theseus].
What these share is that something was taken away, and retiring a choice takes the most, because what was built to serve it goes with it. So the kind of gain John names is in the record. Its size is not. The meta-prompt went whole, but the selector is a different text with a narrower job, and the stage definitions stayed: code still builds prompts from them. Across the text that made prompts, then, about 238 KB became about 118 KB, roughly two to one. (The survey log's figure, a routing prompt of 106.9 KB becoming 10.1 KB [LOG], compares the deleted text with the routing prompt that replaced it and leaves the stage definitions out, which is why it reads as about ten to one.) Nothing on record comes near a hundred, and for the front end nothing has been measured. The 187 is how far the code has grown, not how much could go: the product now has more than a dozen named views (H-views).
Nor has the record blamed the front end's platform: in June it blamed adoption of the component layer, in October the reading of the rule. On 13 November, the day before the quotations above, John wrote that "the UI issues trace back to one of my very first rules: no build step… it's the biggest target of the Theseus method." Yet the ticket keeps the rule: "There's still no build step and no framework: the rule stays, read as allowing the browser to import, as browsers do natively" (H-4446). Three browser scripts became native modules the same day [H-first]. The record's best present reading is that the front end's square peg was a reading of a rule. Whether another platform would be a hundred times simpler is open: nobody has built one page both ways.
Not yet a habit
The record agrees that retiring was not routine. By count, about six process changes added a step, a rule or a scheduled wake-up for every one that removed one [H-land]; by size, 6.3 bytes of steering text went in for every byte that came out [H-steer]; no tangle dissolved between 20 January and 6 October [H-theseus]. The retirements that did happen were events, each begun by a person from outside the ordinary flow. John re-scoped the meta-prompt's ticket twice on 4 October, and a long-running operator session did the work directly rather than through the pipeline [H-3300]. Share links, whose removal dissolved two tangles, went because they "aren't what John wants"; the charter's page went on John's word that it "has probably been superseded" [H-removals].
What is new is the looking. The Theseus method made rebuilding a routine with a judge in mid-October [H-method], and since 8 November code sends every fresh ticket through a look-first sequence before any building (H-survey). That round, look first, size the move, build, tidy what is left, is what John calls the loop. "For the first time" is fair for the looking, not for the removing.
Who rules
Each retirement this essay traced to its start began with a word from John, on a card or not: the meta-prompt, share links, the charter's page. Small ones too. The settings switch went about three quarters of an hour after John's call, leaving its area 137 lines smaller; the paired reads' fixed sample went on a word from John in conversation, which the log gives as "John: drop it." [LOG] These are the cases the essay went looking for, not a sample. Nothing here counts the retirements that began some other way, or the words from John that began none.
The Theseus method asks people "for very little", and can settle ownership by rule, because its moves leave behaviour unchanged [H-method]. Retiring a choice changes what the thing is for, which is not written in the code, so it needs a person. An agent will not do it. The survey session's fresh review: "Text gets its authority from where it sits, and the only way out of it is John." [LOG] To an agent, a standing choice reads as a constraint. And asked to remove, agents add: the session's three new rules, and the paper that counted six additions to the process for every removal, whose own recommendation became a new rule [LOG].
Here the record cuts against the framing: as built, the loop has its questions the wrong way round. Most of its rulings are about clauses, because of the cut rule. Since 7 November code has added a line to every stage that chooses or builds, making any clause of a ticket's text that is not delivered in full a cut, to be raised as a ruling [H-survey]. It went in for a good reason: a measurement had found that review catches a cut someone named and misses one nobody did [LOG]. In the two tidy lanes the survey log records fifteen rulings in about a day and, by this essay's count of its own labels, marks about ten as questions the lane could have settled, such as engineering trade-offs and clauses an agent had written. Both lanes sat idle overnight on answers. Meanwhile the choice to take a larger move goes ahead under a standing policy, "err larger", with no ruling [H-survey]. Of 45 rulings withdrawn when their ticket closed, about 15 "were scope or design calls the run settled itself", a few of them small product calls that shipped: "What was lost was John's say, not the information." [LOG] Clauses reach the person; choices do not.
A person's rulings outlive their reasons too. The front end's founding rule was John's. In October a comment of his that prototypes "may be over specific" became a policy, "No prototypes, screenshots or design review pages go to builders", which the charter's retro named as the cause of pages that did not match their designs [H-retro]. Since 7 November Harbour has sealed every filed ticket description as the person's words, including text that planners and other agents drafted [LOG]. A ruling needs what any choice needs: its reason kept with it, and a way to mark it over. Michael Nygard's decision records keep a Context and a Status that can become "superseded", so that "The motivation behind previous decisions is visible for everyone, present and future" (N11).
The same shape at every size
John's "fractal" is a claim about the loop: "That's also the exact pattern we've built in our loop." The pattern is the one in his first statement: map the code, find what is held in the wrong shape, have a person rule, rebuild. Read that way, the cure is meant to repeat at every scale. The record shows it at two, and at a third it is a picture.
At the scale of a task, the Survey, the loop's first step, asks before any building "why is this difficult?" and offers moves in sizes, from "Just drill" to "Theseus the area. Free the seam, consolidate, rebuild." [H-survey] It maps and finds the misfit. The person's step is only half there: a smaller move than asked goes to John as a ruling, while a larger one goes ahead under standing policy [H-survey]. At the scale of an area, the Theseus method maps the scattered decisions, decides where each belongs, gathers it there and rebuilds, and it produced the largest multiple on record, 122 to 13 on one function [H-method]. At the scale of the whole product, John's own picture, Harbour mapping all of its code and finding the square pegs, is the same steps again. Nobody has run it yet.
The debt the cure is for repeats as well. Each row below is a choice made for a reason, and what was built on it after the reason ran out or was left behind.
| Scale | The choice, and its reason | What stayed or was built on it |
|---|---|---|
| A line | A marker kept "until LIN-4015 retires the tone names" | The ticket is Done; the marker is still there [H-steer] |
| An area | Linear's wire format as the shape of an issue, when Linear was the only tracker | Four other providers rebuild it; three helpers translate back [H-theseus] |
| The platform | "no frameworks, no build step", for a 188-line script | A component layer and eleven hand-made copies of server logic [H-first] [H-coherence] |
| The process | Area notes, so that small debts need not each become a ticket | A second backlog: about 11 KB to 89 KB in six days, five writers, a validator, a quota [LOG] [H-beyond] |
| The loop | The cut rule, so that no cut to the person's ask goes unnamed | With every description sealed as the person's words, clauses agents wrote became locks only John could open; lanes idle overnight [LOG] [H-survey] |
| The person | A comment made for a reason | A policy applied past it [H-retro] |
The survey session's reply to Groundhog Day, three new rules offered with a call for removal, is not in the table. It was not a choice outliving its reason; it is the habit that keeps producing them. Whoever acts next, an agent on a ticket, a session running lanes or a person answering a card, meets what is there as given, answers the local problem by adding, and has neither the reason nor the standing to take anything away. The steering-text paper gathers outside evidence that people overlook subtraction [H-steer].
So the debt shows at every scale this essay looked at, the loop included, and the cure at two. The cure has not yet been turned on the loop itself. The loop's own machinery ran up the debt in under two weeks. Its plan had warned: "A new step is a new gate, and gates are where prose collects." [H-survey] On 12 November the blind read became a gate in code that a ticket cannot close without [H-survey], a new floor of about $1.02 on every task [H-stab]. Some of this went within the hour of John's word. Most still stands as this is written: the notes' quota, the blind-read gate, and the cut rule [H-survey]. The agreed next step is to take them one at a time, "stays, goes, or stays with a planned removal, in John's words" [LOG]. That would be the loop turned on itself. It is a plan.
What it asks for
In practice the framing asks for little that is new: a standing choice kept with its reason and the sign that it is over; a looking step that asks about the choices a task stands on, not only the code; and the choice, not the clause, put to the person.
Where the framing runs ahead of the record:
- "I assume every choice was right." As an assumption it does its work, sending the reader to find each choice's use. As a finding it would fail: some choices were wrong when made.
- "Easily." Quick to carry out once ruled, but not shown easy to find or to prove. The largest rebuild is "about 930 lines over two files" [H-ladder], and the Theseus method's own test, later changes landing smaller, "has not had time to show" [H-method].
- "100x." Unmeasured. Nothing on record comes near it; the figures are in the section on tidying and retiring.
- "It requires humans to rule." True of every retirement this essay traced to its start, though it chose those cases; and as built the loop asks the person about clauses and settles choices itself.
- "It's fractal." The cure runs at the scale of a task and of an area; at the scale of the loop it is a plan.
- The evidence. One codebase, one owner, a few weeks; cases chosen, not sampled; and much of the loop's evidence is the survey session's own log about itself.
Annotated reading list
Markers in the text refer to the entries below. Each says what the source supports here and what it does not. Harbour's record was read at LinearViewer 993f618bc unless another sha is given.
Harbour's record
- [J] LIN-4486, the ticket's description ("As asked", 14 November 2026). Supports John's quotations and the steps of the session, quoted as written. The ticket's own selection from one day's conversation, not a transcript.
- [LOG] The survey session's log: the file
docs/plans/survey/log.mdon the branchclaude/cool-rubin-6acj5j, read at commita593d0ea. It is not on main. Lines 3060–3071 and 3109 (the settings switch), 2890–2893 (why the cut rule exists), 3612–3614 (withdrawn rulings), 3649–3835 and 3997 (the tidy lanes' rulings), 3840–3856 (the stalled lanes, the notes meant to replace ticket-minting, the gate the cut rule locked, "John: drop it"), 3879–3900 (the fresh review), 3903–3925 (the notes: growth, the validator), 3927–3957 (the history of the steering text: six process changes adding to one removing, the five writers, the seal, the routing prompt's 106.9 to 10.1 KB at line 3945, the six-to-one paper's recommendation becoming a rule at line 3936), 3959–3977 (Groundhog Day, the plan). Supports the session's events, John's words as recorded, and the counts of rulings. One session's record of its own work, written by that session; its clock headings can be off by an hour or more, so timings here are approximate. - [H-steer]
steering-text.md, "Why does the text that steers Harbour's agents keep growing, and what has ever made it shrink?" (14 November; version 1, unchecked). Supports 6.3 bytes in per byte out, 37 of 57 cut-titled commits growing the text, single-rule trims never moving a kilobyte, deletion as 74% of the carried text's real shrinkage, a marker waiting on a Done ticket, the three reasons temporary structure stays, and its reading of the outside literature on subtraction. Bytes measure size, not reading load; it does not show that the growth made the work worse. - [H-beyond]
beyond-the-code.md, "What do agents need to know about a codebase beyond the code…" (14 November; version 1, unchecked). Supports the Waiting list going from 30 to 118 in six days, the notes as a backlog no work view shows. It does not show whether notes pay for themselves. - [H-lanes]
lanes-across-the-seam.md(14 November; version 2 in the repository, and its independent check,lanes-across-the-seam-check.md, covers version 1 only). Cited only as the third paper of the day, and as the keeper of LIN-4446's text. - [H-shape]
nothing-fights-its-shape.md, "Nothing fights its shape" (LIN-4485, 14 November; version 1, read by an independent reviewer before merge, its revisions not re-checked). Cited only as the companion essay on how the kinds of information around a codebase should be held; nothing here rests on its findings. - [H-4446] LIN-4446, description ("As asked", 13 November), as kept in
lanes-across-the-seam-evidence/issues/LIN-4446.json. Supports John's words on his "very first rules" and the ruling that the rule stays, re-read. A ticket description drafted by the survey session quoting John. - [H-first] The repository itself. First commit
a82b8758(4 January 2026): twelve files;CLAUDE.mdline 36 carries the rule;public/app.js188 lines; 1,112 lines in the eleven files outsidepackage-lock.json, 912 of them in the six code and style files. At993f618bc: 44 files and 35,236 lines inpublic/*.js; three browser scripts converted to native modules on 13 November (0cf05ce8,b8e2a0aa,6348e578). Line counts read withwc -l. They measure size, not what a simpler front end would need. - [H-482]
docs/lin-482-ui-simplification-research.md, lines 175–180 and 238–251 (15 June). Supports the framework question asked and answered no, the rule reaffirmed "for now", and htmx deferred. A research note, with the owner's decision recorded by an agent. - [H-debrief]
docs/reviews/v1-passage-debrief-2026-10-10/backlog-questions.md, line 96. Supports "Step 2 was never ticketed." It does not say why. - [H-swim]
plans/swim-view.md, lines 150–151 (commit0a6f6cec, 16 March). Supports the copy accepted "since there's no build step". - [H-coherence]
coherence-as-it-grows.md, lines 75–77 and 92. Supports the founding line as the largest producer of twins (eleven), "the rule does not require the copies; the reading of it does", and the meta-prompt figures (53 of 63; 106 KB gone; stage definitions 132 KB to 110 KB; an 8 KB selector). A count of twins by two readers; it does not test what a framework would do. - [H-paths]
one-job-many-paths.md, lines 149–159. Supports the size of the meta-prompt path's first deletion: 236 lines added and 1,248 removed across 27 application files. In its own words "a reproducible change-size measurement", not proof that agents' work improved. - [H-3300] LIN-3300, description and comments of 4 October 2026, read through Harbour's tracker. Supports John's two re-scopes and the operator session doing the work outside the pipeline. The comments are an agent's record of what John said.
- [H-removals] LIN-3325 (description, 6 October 2026) and LIN-3912 (description, 24 October 2026), read through Harbour's tracker. Support who began the removal of share links and of the charter's superseded page. Each is a ticket written by an agent reporting John's wish; LIN-3912 quotes him.
- [H-theseus]
theseus.md, "Can Harbour be rebuilt plank by plank, and which plank first?" (15 October), lines 60–135 and 171–197. Supports 24 tangles born and 5 dissolved, none dissolved between 20 January and 6 October, all 5 by taking something away, 6 births from consolidations, the re-knot five days on, and Linear's wire format as the providers' shape. Measured to 13 October; version 1, unchecked. - [H-method]
theseus-method.md, "The Theseus Method" (essay, 17 October), lines 35, 46, 62, 70–90 and 136. Supports the moves (find, own, free the seam, consolidate, rebuild), judging by replay, agents having built almost all of Harbour through its own pipeline, ten rebuilds cutting the most complex function by 29% to 88%, the dispatch entrance's worst function from 122 to 63 to 29 to 13, people asked "for very little", ownership by rule, and "has not had time to show". Three rounds on one codebase; complexity is one measure of shape, not of size. - [H-ladder]
rebuild-ladder.md. Supports "about 930 lines over two files" and "up to 39% in a file". One run per arm. - [H-consolidate]
consolidate-then-rebuild.md, lines 14–24 and 51. Supports the places holding the entrance's decisions going from 95 to 26, and 317 production lines removed by the gathering. One run per target; the 95 to 26 is a hand count. - [H-land]
how-process-changes-land.md, lines 14–24 and 126–160. Supports about six process changes adding a step, rule or wake to one removing one (a count of changes, not bytes; interval about 2:1 to 21:1, resting on four removals), none of 27 changes with a retirement condition, and the one sunset never read. - [H-charter]
how-the-charter-got-complicated.md, line 83. Supports the charter's goal in John's words. - [H-retro]
docs/reviews/charter-retro-2026-10-26/README.md, lines 7, 24, 39–41, 45 and 54. Supports production code +9,725 lines, the moves (3,782 to 882; 26,955 to 28,559), the no-prototypes policy named as cause, and the deleted page's 7,681 lines. A retro written from the record by an agent; some of its inputs are marked not checked, none of the figures used here. - [H-stab]
docs/reviews/stability-review-2026-11-14.md, lines 102–119, 134, 168, 186–195 and 307. Supports "converging on the things it aimed at", net +781, "replacement churn", the blind reads' 20 better, 2 same, 0 worse, and the $1.02 floor. A window of 1.4 days; the blind-read verdicts are the loop's own, as the review itself notes. - [H-drift]
docs/reviews/drift-coherence-review-2026-11-14.md, line 11. Supports "better" and seven earlier findings resolved. It follows named duplicated decisions, not size. - [H-reads]
comparison-302edafc3-vs-4b5001ddf.mdandcomparison-49776653d-vs-e8e92d50a.mdindocs/plans/survey/area-read-reports/, line 3 of each. Supports "no detectable difference", and 3 of 14 sampled files changed in one lane. The reports say this is not evidence that the work failed. - [H-survey]
docs/plans/survey.md, lines 84–90, 98–101, 112 and 171;lib/code-routes.jslines 60–68;lib/size-of-move.jslines 10–13;lib/run-closeout-state.jsline 80 (the blind-read gate, merged aseb14b7ffon 12 November);lib/prompts/stage-kinds.jslines 26–29 (the cut rule, from11dc1d7b, 7 November, added to the stages that choose or build);docs/areas/README.mdlines 67–72 (the notes' quota). Supports the Survey's question and menu, the look-first sequence run by code, "err larger" as standing policy, smaller moves as rulings, the plan's own warnings about gates, and the three structures still standing. The code shows that they exist, not whether they should. - [H-views]
docs/architecture/views.md, line 5. Supports more than a dozen named views in three tiers: six first-class and seven experimental by name, plus a flagged tier.
Outside sources
- [C92] Cunningham, W. (1992). The WyCash Portfolio Management System. OOPSLA '92 Experience Report. http://c2.com/doc/oopsla92.html. From 1992. Supports debt as first-time code that is fine if repaid, and "the failure to consolidate". About one product built in Smalltalk; says nothing about agents or prompts.
- [C09] Cunningham, W. (2009). Debt Metaphor (video; transcript on the c2 wiki, "Ward Explains Debt Metaphor"). https://c2.com/wiki/remodel/pages/WardExplainsDebtMetaphor. Supports debt as code that reflects an earlier, partial understanding, not poor code. A recollection seventeen years on.
- [F14] Fowler, M. (2014). Sacrificial Architecture. https://martinfowler.com/bliki/SacrificialArchitecture.html. Supports replaced architectures as right for their time. Its examples are whole systems outgrown by scale; it does not say how to tell when the time has come.
- [CH] Chesterton, G. K. (1929). The Thing, "The Drift from Domesticity". Read as quoted in https://en.wikipedia.org/wiki/G._K._Chesterton (section "Chesterton's fence"), not in the book. From 1929, about social reform; an analogy only. It says to learn a fence's use before removing it, and nothing about what to do once the use has ended.
- [N11] Nygard, M. (2011). Documenting Architecture Decisions. https://cognitect.com/blog/2011/11/15/documenting-architecture-decisions. Supports keeping a decision's context with it and marking it superseded. A practice for teams of people; there is no evidence here that agents would keep such records current.
Next
These are options for John, not decisions. All three are lines in proposals.md.
- How much of the front end exists to serve the no-build rule, and not the product? Classify the browser scripts, the render modules and the tests that police them at one commit by whether each would exist under native modules as now ruled, under a template engine, and under htmx; then rebuild one page each way, design-first and judged by replay (the old page and the new run on the same inputs, with everything a user could see compared), and report the sizes. It would put a number where "100x" stands.
- Does retiring a choice shrink the code more, and for longer, than tidying inside it? Set the retirements on record (the meta-prompt, share links, the charter's deleted page, the settings switch) beside tidies of similar size: net lines at merge and thirty days on, and the size of later changes in the same files.
- Which questions reach John? Sort the rulings raised since 8 November into clauses and choices, and list the choices settled with no ruling. If the split is what the survey log suggests, the fix is in where the cut rule points, not in more rulings.
The steering-text paper already proposes registering each temporary structure with a condition code can check [H-steer]. A register of standing choices, each with its reason and the sign that it is over, is the same idea one level up.