{"path":"research/what-he-held-in-the-group-2011-2013.md","content":"# What he held in the group, 2011 to 2013\n\n*His own 100 posts and 532 comments in the Facebook group Collaborative Argumentation Analysis, read\nwhole on 2026-10-22 from the archive scraped 2026-03-27 (`data/fb_group_archive/`), on his word that day:\n\"Wow... Yes!!!! Do it. Gold mine I'm sure.\" Every quotation below is his. Other members are named only\nwhere `docs/history.md` already names them, and otherwise paraphrased as \"a member\".*\n\n*The other members' side of these threads, and where every idea in the group stands today, built, reversed or still a delta: `what-the-group-held-2011-2017.md`.*\n\n**How to read the dates.** When this note was written the archive showed only an age per post (\"14y\"), so\nthe year ranges below place a post within a year: *2011–12* means between 27 March 2011 and 27 March 2012.\nEvery post has carried its exact header date since the evening of the same day (§ 4); a comment still\ncarries no time.\n\n**What the reading found, in one paragraph.** Most of what the project holds today was said by him in\nthis group in 2011 and 2012, often in a sentence: strength judged on the arrow and not on the claim, one\nargument carrying many wordings, concepts defined by the statements that use them, automation first and\nmicrotasks for what is left, one claim in the centre with a few inferences around it, private spaces\nbeside a public commons, science as a rough diamond to be sharpened by making every inference checkable.\nThree positions he held then he has since reversed, each by a recorded ruling (§ 2): that the inference\nneeds no statement of its own, that evidence should stay outside the system as a reference, and that a\nladder of validators should decide what is valid. And the archive's March copy was less sound than the\ncount suggests: 49 posts carried text from Facebook's 2025–26 feed in place of or in front of the real\npost, six of them under his name; all were re-read from the live group the same day (§ 4).\n\n---\n\n## 1. What he held, by theme\n\n### 1.1 Where strength is judged: on the arrow, not the claim (2011–12)\n\nAsked by a member whether participants should vote on the arguments that connect a premise to a\nconclusion rather than on the conclusions: *\"That is actually what I'm planning. I've been adamant in\nthis respect when people suggest voting on the claims themselves.\"* He then dropped relevance and took\nit back within the year: *\"I've decided to skip the relevance metric. It seemed redundant, since the\nvalidity metric should encapsulate all we need anyway. If an argument is valid, it is automatically\nrelevant. Right?\"* (2011-12-18); a member argued that validity belongs to a subclaim and relevance to its\narrow, and on 2011-12-30 Alex Vance conceded *\"relevant = valid\"* for the senses they were using. The\nsketch with both columns came first, not after: on 2011-07-17 he posted *\"Voting results, for the aspects\nvalidity and relevance, are displayed as columns on either side of the arguments and their co-premises\"*,\nand on 2011-07-29 he wrote that voting on relevance *\"can be augmented or even replaced by arguing about them\non a meta level. So that a claim that \"claim X is relevant to Y\" is auto-created and the aggregated\nprobability for the claim is used as the relevance score for link X->Y.\"* (Dates and the July post from\nthe group page he saved on 2012-01-24, which keeps comments the archive lost; corrected 2026-10-22.) On the pair of words: *\"Should soundness be\nkept separate from relevance (as a dimension of argument appraisal), or is that asking too much of\nusers? … Sorry, meant validity - as soundness implies validity of premises, not just the current\nargument.\"*\n\n**Today.** The two factors on a link, the supporting claim's own strength and the link's relevance,\nruled 2026-09-28 (ledger Q33, built, live 2026-09-30), are this 2011 design; the July 2011 post is also\nthe relevance claim reified on its link (AM5) and Q33's *argued by people, never voted*, in his words.\nWhat shipped differs in one respect he was *\"adamant\"* about: the agree/disagree vote sits on the claim,\nsingle-axis, and the computed badge is on the claim too, though it is computed from the links and the\nanswers to their questions.\n\n### 1.2 One argument, many wordings (2011–12)\n\n*\"Thinking about ways to limit duplication... Every argument should be mergeable with any other, provided\nsufficient pressure from users. They will then have their wordings collected in one argument, and the\ndifferent wordings will be ranked by number of votes. The most popular wording will be visible when\nbrowsing the system.\"* And on the losing wordings: *\"Yes, they'll just be one click away, so that you\ncan always vote on other phrasings.\"*\n\n**Today.** Cluster-carries-the-strength (the merge doc credits it to him, from about 2008; this is its\nfirst dated public statement), built 2026-10-20 and switched off; sameness is decided before any relation\n(ruled 2026-09-13); a claim's text is written once, a rephrase is a new claim and the old one stays\n(ruled 2026-10-17, ledger AL10). The 2011 \"one click away\" is the 2026 rule that nothing is overwritten.\n\n### 1.3 An open wiki, with history, and a queue for the needy (2011–12)\n\n*\"Rewording claims, changing pretty much anything - open wiki model (just logging every change, all\nreversible) better than requiring a voting threshold to achieve a kind of inertia/resilience to\ndestructive changes?\"* … *\"I guess version history is unavoidable. Question is how much inertia will be\nappropriate.\"* … *\"If no inertia or very little, or maybe in any case, the system could provide a queue\nview of changes to be scanned by users with too much free time on their hands, to aggregate \"needy\" data\nfor inspection and adjustment.\"* Later the same year: *\"I have had revision history for arguments in my\nnotes since more than a year back.\"*\n\n**Today.** Contributions are added, never overwritten (ratified 2026-08-27); edge supersession and the\n`needs-help` feed are the queue; *never bottlenecked* (ledger AS1, ruled 2026-10-20) and its extension\nto kept links (AS2, ruled 2026-10-22) answer the inertia question his way: proceed, tell, let them undo.\n\n### 1.4 Who judges: the validator ladder, and the fallacy told back (2011–12)\n\n*\"Normal users can say that an argument is logically valid. \"Validators\" can mark arguments as valid,\nbased on user votes. Validators cannot change each others' judgments of validity; that's left to another\nclass of users called \"inspectors\" (or whatever). Inspectors can't change each other's judgments - only\nadmins can.\"* Promotion by agreement: *\"When a normal user's validity votes overlap with the judgments of\nvalidators to a certain extent (9/10 maybe) after a certain period of time, they will achieve the\nvalidator title.\"* And the aim: *\"So there would be a power hierarchy, but we will do all in our power\nto make sure logic is the sole selection pressure.\"* Also: *\"As soon as anything about the content/form\nof an argument changes, users are notified to re-evaluate their votes.\"* On feedback: *\"as a validator,\nyou could mark an argument as invalid, but you receive extra points for also indicating what type of\nfallacy has been committed. (Knowing the fallacy is important, the app can then notify the original\nauthor of the argument and show them visually and very pedagogically exactly in what way they were\nwrong.)\"*\n\n**Today.** Reversed, by his own rulings: there is no ratify act and bare judgments are telemetry\n(2026-08-31); no expert gate, since myside bias is independent of ability; what earns the right to judge\nby measured agreement is a machine pass, not a ranked user (the trust ladder, proposed). The fallacy told\nback survives as the critical questions a scheme carries, and as the rule that no gap is displayed\nwithout a move offered. The corpus graph already records this reversal against the Simplenote notes of\nthe same period; the group confirms the ladder was his, in his words.\n\n### 1.5 A concept is defined by the statements that use it (2011–12)\n\nAgainst a philosopher's objection that regulating language is a non-starter: *\"The solution I have\nthought up is not by any means complete. But I think of it as a way to chart and navigate the complexity\nof language, not work against it. For each concept, you would simply see truthy statements that use the\nconcept, and those statements more or less comprise the \"definition\" of that concept. An organic,\nexhaustive and nuanced way to reinvent the dictionary.\"* Then: *\"Statements use concepts/meanings. A\nconcept therefore is defined by its (truthy) statements. So you just correlate meanings which have\noverlapping truthy statements: those are probably close to synonyms.\"* With his own caveats: *\"to\ncalibrate the thresholds for creation/removal of new meanings/concepts so that it works in a balanced\nway … if duplicates arise, they are merged when close enough in meaning and not otherwise (this is a huge\nproblem though)\"* and *\"I know I should be wary of confirmation bias here. I guess I should do a\nsmall-scale test soon, to see if some of my thinking can be verified.\"*\n\n**Today.** *A use is a vote on the meaning* (ledger AM7, ruled 2026-10-19), the concept layer, and\n`store-the-uses-read-the-meaning.md`. The ledger row already cites his January 2011 Simplenote words;\nthese are the same position argued in public within the year, with the threshold problem he named still\nthe open one (sense granularity, § 2.3 of `core-periphery-and-the-concept-layer.md`).\n\n### 1.6 Many arguments for one conclusion: the squirrel (2011–12)\n\n*\"Just to clarify I mean simply to argue that \"The squirrel is dead.\" from \"The squirrel smells of\ndecomposition.\" along with \"Dead squirrels smell of decomposition.\" as well as perhaps from \"The squirrel\nhas not eaten anything for a whole week.\" & \"Squirrels need to eat to stay alive.\" There can of course be\nmany reasons/arguments for a conclusion, I guess what I'm wondering is if they have to be\ndeconstructed/reduced so that what remains is one most proximal set of premises on which the inference\nis made. Leaning towards no. But I'm sleepy.\"*\n\n**Today.** Q38, ruled 2026-10-22: within one argument the weakest premise governs; across arguments the\nbest one counts until their independence is read, then independent arguments add. The squirrel is the\nworked case of `philosophy-of-the-unwritten-step.md` and `the-squirrel-and-the-strength-calculus.md`.\n\n### 1.7 Does the inference get a statement of its own? (2012) — reversed\n\nA member asked whether the Socrates syllogism has four parts, the fourth being *3 is true because of 1\nand 2*. His answer then: *\"I'd say it is (and perhaps should be only) implicit that the conclusion\nfollows from the premises.\"* … *\"The relation already is represented by a link (edge, in graph\nterminology) between the statements (conclusion and premises), so I don't think it needs to have its own\nstatement.\"* He wanted one tag instead: *\"What we need, though, is to tag arguments as either deductive\nor inductive, to know when the argument is actually deductive and this rule therefore is assumed.\nInductive arguments, to remind everyone, are just truthy, not true/false.\"* And on size: *\"Can one always\nbreak up polysyllogisms like that? It should be praxis within the app to try to do this whenever\npossible, I think, to avoid unneccessary complexity...\"*\n\n**Today.** The opposite, by ruling: the link carries a relevance claim, a machine's sentence written\nunder a link is a claim the moment it is written (AG12, ruled 2026-10-10), the rule under an application\nis a reason (corrected 2026-10-21), and a step is written to its smallest size, one stated rule applied\nonce (AG14, ruled 2026-10-10). The polysyllogism sentence is the floor's first statement; the\ninference-needs-no-statement sentence is a position he left.\n\n### 1.8 Evidence stays outside, referred to (2011–12) — reversed\n\nSaid twice, under his feature map and when turning to induction: *\"it might be better to just have users\nrefer (in their claims) to evidence that exists outside of the system and to argue about that, so as to\navoid having to convert every relevant piece of evidence for any argument to data that can be stored in\nour database.\"* With: *\"I definitely think we'll need to do some example maps in order to elucidate this\nquestion. It's just too hard to think up a good way to handle probabilistic data in the context of\nnatural-language arguments without a scaffold for how it's going to look like...\"*\n\n**Today.** His decision of 2026-03-29 goes the other way: evidence decomposes into constituent parts and\n*\"that PDF should be decomposed as far as possible\"* (`scheme-bounded-decomposition-and-evidence-as-subgraph.md`).\nWhat counts as evidence and how it enters is Arc Q, open.\n\n### 1.9 Axioms, belief, and a filtered view that says it is filtered (2011–12)\n\n*\"The current thinking is to let users vote on what is an axiom, the claims with most votes show up in a\nqueue for validators … who can then deem them to be axioms or not. Once acknowledged to be an axiom,\npeople can indicate they believe it or not.\"* Default: *\"assume the median of the distribution of above\nmentioned belief indications.\"* And the rule that outlived the mechanism: *\"If the user's view is\nallowed to be filtered, it should be visually made obvious that it is filtered somehow and one should\nnever be far away from deactivating it or readjusting the belief indications which shape the filter.\"*\nOn the bar: *\"Setting the bar high is good, but we're still going to have to deal with actual users to\npopulate the site, users who are only human...\"*\n\n**Today.** The worldview lenses (`docs/worldview-lenses.md`); *say where the reader is* (ux P25, ruled\n2026-09-22: the compartment is unmistakable on every page); no horizon is set, every kind is read\n(2026-09-15). The axiom vote is gone with No Copout Axioms.\n\n### 1.10 Automation first, microtasks for what is left (2011–12)\n\nUnder his feature map: *\"The general strategy is to leverage automation as far as possible, then move on\nto crowdsourcing; designing microtasks for users to do the work that still is beyond computers.\"* When a\nmember said users would never post arguments cut that fine: *\"Yes, designing microtasks will be key, just\nneed to design macrotasks first to be able to split them up.\"* On what the system is: *\"It's not about\nreplicating the function of human brains (as with Bayesian networks et al), it's more about guiding and\nharnessing multiple human brains working together, to represent in the form of interconnected syllogisms\nwhat we know and how we know it to be true, and continually refine and expand this data.\"* On method:\n*\"we are just going to build this to harness natural language + crowdsourcing + logic guidelines + a good\nUI. We're not trying to do machine learning on any advanced level, at least, hrm, not yet... Maybe some\nnatural language processing.\"*\n\n**Today.** The machine passes and the microtasks they leave to people (`microtasks-in-broad-daylight.md`),\nthe harness conviction (ratified 2026-08-30), and the April position that premises are *found or created*\n(corrected 2026-10-07 from Claude's narrower *never generating*). The 2011 order, automation then crowd,\nis the order the passes run in.\n\n### 1.11 One claim in the centre, a few inferences around it (2011–12)\n\n*\"we'll design one particular view for looking at a single argument, one view for maybe 3 arguments at\nonce, one view for 7 or so, one view for even more than that. These will be done in somewhat different\nways for displaying the supporting args for a conclusion versus the args that a conclusion supports.\"* To\na member who feared the graphs would be huge: *\"you wouldn't show the whole graph at once. Just maybe one\nor three inferences, and you could zoom out to show more sometimes, but mainly just browse around, always\nshowing only a small portion of the argumentation.\"* On a member's phrasing, one point in the centre and\nthe map outwards from it: *\"That's what I'm imagining as well.\"* And earlier, on the forces: *\"The problem\nis I want it to look somewhat similar on each render (to not confuse the user with different node\nplacements).\"*\n\n**Today.** The map view (2026-09-21) and the wheel view (2026-10-20): the claim on a line, what it bears\non above and what bears on it below, three rings, a click travels. The two-direction split he named is\nthe two halves.\n\n### 1.12 A public commons and private spaces (2011–12)\n\n*\"In any case, there will be a large public workspace, somewhat like Wikipedia, where anyone can join.\nWhat we'd be selling is the ability to have private argospheres, hosting for those, and support +\nadministration.\"* And: *\"we'll probably have to make sure it can be deployed for exclusive use on private\nservers … or as an option, you could have a private space for your arguments on our servers.\"*\n\n**Today.** Blobs (built 2026-09-19): a separate store per community, the commons silent and a compartment\nmarked. The word changed; the shape did not.\n\n### 1.13 An answer service, and a map for any URL (2011–12)\n\n*\"Could make an answer service to show perhaps the top 3-5 pros and cons for a claim/question -\ndetermined through collaborative argument mapping behind the scenes.\"* And, after a conversation with\nTobias Wallin: *\"a URL service like bitly … to sit between social networks and any info sources … and show\nany argument discernable in the content of the link as an argument map. So instead of hoping people come\nto me, I'd allow anyone to start creating overlayed argmaps on top of the rest of the internet.\"* On\nhypothes.is: *\"This is not that. We can complement each other.\"*\n\n**Today.** Extraction from a URL, Think with Deliberus and the truth graph's answer layer, and the\nagent-readable surface. The overlay on the rest of the internet is not built.\n\n### 1.14 Science as a rough diamond (2012)\n\nOn a member's remark that trust in science rests on being able to check the work: *\"Deliberus makes the\nprocess more granular, maps out the entire reasoning process, documents every single inferential step.\nEven all the fallacies & errors found and corrected during the mapping of the issue, all tweaks and\nrewordings will be there in the version history. You can browse the final map/s and check up on every\nthought/inference and verify its adherence to the rules of logic - this is true for the smallest of\nassumptions, for every little piece of the puzzle. Science is a subset of consensus tech, and in its\ncurrent form it's a very rough diamond. It's about time we sharpened it a bit.\"* And against the worry\nthat this starves intuition: *\"I don't think this approach precludes or limits human creativity and\nintuition. It is a complement … we'll have all the more mental energy to take a holistic, creative\nperspective.\"*\n\n**Today.** `the-missing-layer.md` and the science lineage (*extends science as public error-correction*);\n*checking that keeps pace with persuasion* (2026-08-18). The 2012 sentence is the public one-liner's\nancestor.\n\n### 1.15 Smaller positions, each with its home\n\n- **The founding question (2010–11):** *\"Can the function of thought experiments be preserved/obsoleted\n  by allowing hypothetical statements to be inputted, interlinked and arranged as logical inferences,\n  then visualized?\"* → an *if*-sentence is itself a link (AM5, ruled 2026-10-19).\n- **What he wanted it for (2011–12):** *\"In every meeting, every (ever so slightly) argumentative\n  conversation, every far-reaching decision … I long for a higher level of abstraction and definition\n  for the methodology of deliberation, and I wish that somehow it could just be overlayed on our\n  biological senses to make it second nature to people in the developed world.\"* → `docs/vision.md`.\n- **The hive mind (2011–12):** *\"above all I want to raise the sanity water line in society and help us\n  morph into a hyperrational hive mind.\"* → narrowed by his ruling of 2026-10-10 (ledger S8): members\n  give up working everything out alone and keep their own position; a hive-mind reading is excluded.\n- **Gamification (2011–12, then 2012):** first *\"game design will be integral to making the\n  argumentation environment engaging\"*, then *\"I don't know about gaming as the perfect model, you risk\n  hollowing out intrinsic motivations by overdoing gamification.\"* → the engagement gradient; curiosity\n  as the precondition on the wager.\n- **Karma by subject (2011–12):** *\"good contributions to arguments within a certain subject (tag) gives\n  the user more power to affect change within that subject.\"* → not built; the contributor-rationality\n  score was killed on evidence (no general bias-proneness factor).\n- **Many narrow views over one store (2011–12):** *\"This is a possible long term, narrow application of\n  the more general aim of efficient collaborative argument mapping. We'll have to get that up and running\n  and working good first, then see how the massive amounts of data generated from that can be visualized\n  and applied to specific needs.\"* → the agent-readable surfaces; the wager that one graph serves many.\n- **Deductive first, inductive later (2011–12):** *\"Starting to move on to inductive arguments after\n  thinking about deductive ones too much for a while...\"* and, on aggregation: *\"How should the strength\n  of arguments aggregate? How does an added weak argument affect existing arguments' strength? How to\n  weigh pros against cons, algorithmically?\"* → the strength layer's whole history; Q38; the open\n  independence question.\n- **Pitching by use (2012):** *\"If you're talking to a CEO, mention more balanced and rational\n  decision-making … For someone working in jurisprudence, emphasize massively distributed analyses of law\n  text or even court documents perhaps. If you're talking to someone who won't have any use for it,\n  don't.\"* → `incentives-analysis.md`: eight conditions where legibility pays.\n- **The name (2011–12):** *\"I like the input, but I really like \"Deliberus\". It's short, friendly, gives\n  a hint that it is about deliberation...\"* and, to the worry that people hear *deliver us*: *\"well...\n  isn't that a good thing?\"*\n\n### 1.16 The project's own history, in his words\n\n- Open source or startup (2011–12): *\"I'm torn and have been for a while now.\"*; *\"Currently I lean\n  toward going that path though, since I want to be able to work on this full-time, and actually I have\n  an angel investor who supports the idea, if I want to go that route.\"* The privacy audit's anchor, an\n  earlier agent's *\"angel investor secured\"*, overstated a real sentence: support for the idea, not money\n  secured.\n- Why the alpha stopped (2012): *\"I stopped working on it last year because it seemed wise to consider\n  other technologies than Rails and MySQL, and to design the system before implementing it, so as not to\n  paint ourselves in any corners.\"*; *\"the data model has not followed my realizations for the last year,\n  which is kinda part of why I stopped coding on it.\"* A graph database was weighed in 2011 and set aside\n  on a friend's advice; the project runs on one today.\n- Why it was slow (2012): *\"The slow progress so far has been mostly because I have not awarded it enough\n  time (need to prioritize and focus better in my life, I say yes way too often).\"*; *\"Armchair\n  philosophy only goes so far.\"*\n- The format problem, which this corpus answers (2012–13): *\"It's a problem that everything just vanishes\n  in the linear history. We need categorization at least. The wiki was a good try. I think the latency\n  actually made me stop using it.\"* Earlier: *\"I've DLed all of the group content, and am just trying to\n  summarize everything concisely and then publish as gdocs...\"* (a 2012 download that may sit in the 2014\n  Google Docs export on the keeper's drives; not looked for).\n- The Pirate Party (2012): *\"They seem interested in trying Deliberus for development of their party\n  program. … Their ~90 active forum members would be a nice userbase to start testing with when a usable\n  alpha is ready.\"*\n- Leverage Research (2012–13): *\"I'm interning at Leverage Research and working on consensus tech.\"*\n- The startup (2013): *\"I'm doing this as a startup together with Nils Janse. We're putting together an\n  advisory board\"*; on the group: *\"It's kinda why I created this group but then way too many people got\n  added I guess\"*; and *\"Swedish splash page is up since a little while back.\"*\n- After it (27 November 2014; the post was re-read from the live group the same day this note was written):\n  he had left the startup the year before for lack of motivation, remained interested in *\"deliberative\n  politics, argument mapping and massively parallel collaborative reasoning/decision-making\"*, and pointed\n  the group to a member's open-source, non-profit initiative in the same spirit.\n\n---\n\n## 2. The record's claims about his early words, checked against the group\n\n| Claim in the record | What the group holds |\n|---|---|\n| Ledger AM7: his January 2011 position that a concept is defined by the statements that use it (the Simplenote note; corrected 2026-10-22) | Confirmed, and argued at length in public in 2011–12 (§ 1.5), threshold problem included |\n| `philosophy-of-the-unwritten-step.md`: the squirrel example is his, 2011 | Confirmed, verbatim (§ 1.6), with his *\"Leaning towards no\"* |\n| The merge doc: cluster-carries-the-strength is his, from about 2008 | The group holds its first dated statement, 2011–12 (§ 1.2); nothing earlier is in the group |\n| Corpus-graph reader finding (2026-10-10): the 2011 validator ladder was reversed by *there is no ratify act* | Confirmed; the ladder is his in his words (§ 1.4) |\n| `structuring-gradient-lineage.md`: P20's architecture is his from December 2011 and he held the opposite on structured input in January 2012 | The group has the companion: *\"designing microtasks will be key, just need to design macrotasks first\"* (§ 1.10); the December 2011 line is in the group too, in comments the archive lost: *\"try to automate (because code has no biases); fail rather early and fall back to crowdsourcing\"* (2011-12-06) and *\"at every step in the user's process for input and editing, we'll try to integrate duplication-sensing and error-pruning\"* (2011-12-14), on the page saved 2012-01-24 (corrected 2026-10-22) |\n| `history.md`: the hybrid-AGI line is Ivan Phillips's | Not in the group; it is a December 2011 thread preserved in a chat log (checked 2026-10-22) |\n| `docs/history.md` § 2013: the startup with Nils Janse, the Swedish site | Confirmed (§ 1.16) |\n| CLAUDE.md privacy audit: *\"angel investor secured\"* was invented | Half: his 2011–12 sentence says an angel investor *\"supports the idea\"*; *secured* was the invention |\n| The 2026-10-07 correction: his March 2026 sentence said *finding or creating* premises | Consistent: in 2011 the machine was to *\"leverage automation as far as possible\"* (§ 1.10) |\n| `self-similar-decomposition-and-claim-ontology.md`, the 2026-03-29 ruling that evidence decomposes as far as possible | He held the opposite in 2011–12 (§ 1.8); the 2026 ruling is a change, not a continuity |\n\nNothing in the group contradicts a ruled row. Two rows should carry the group as a second source: AM7\n(§ 1.5) and Q38 (§ 1.6); and `history.md` should say that the inference-as-link position of 2012 is one he\nlater reversed (§ 1.7), since a reader mining the early record for the ontology would otherwise take it as\nstanding.\n\n---\n\n## 3. What the group does not hold\n\nNo post of his states the convergence wager, the two poles, or the horizon kinds; the nearest is *\"the\nsolution is fairly known in this context: find truth\"* (2011–12) and the sanity-waterline line. No post\nstates the inversion principle in P20's words. The 2011 design decisions he refers to as being in his\nnotes (revision history, the semantics module, UI sketches) are pointed at, not stated: *\"There are lots\nof notes in various places (physical and digital)\"*. The group is the public half of a record whose\nprivate half is the Simplenote notes, the chat logs and the 2014 Google export. One limit on every\nabsence in this section: the archive keeps about ten comments per post, and for the first year about 166\nof his comments are only on the page he saved on 2012-01-24 (`what-the-group-held-2011-2017.md` § 5, The\narchive is capped); for posts after that date nothing restores them.\n\n---\n\n## 4. The archive, read against the live page, and repaired\n\n**Feed text in place of posts, found and fixed 2026-10-22.** The March scrape's text grab fell back to\nthe whole page when it could not find the post's dialog, and read Facebook's home feed behind it. Read\nagainst the live group, 49 posts carried 2025–26 feed text (a model leak, a video-model shutdown,\npodcast blurbs, the same nine-word headline on ten posts of 2013) in place of or in front of the post,\nsix under his name, one entire (post `765070753541549`, 27 November 2014, on leaving the startup); 64\nhad their opening text twice, the feed's preview beside the full text; and short posts whose text sits\noutside Facebook's usual text container were stored empty. Keyword and shared-paragraph checks found\n34 of the 49; the rest came out of the doubling and of a check against the group feed's own preview\nof each post (`2026-03-27_post_index.json`), which no feed text can match. The same day the scraper was\nmade to read inside the post's dialog only (a page with no post dialog yields nothing, never the\nfeed), to drop a preview that is a prefix of the full text, to take a post's text whole when its\ncontainer has no inner text node, and to recover the header's date (Facebook draws it one character\nper span with decoys stacked at one position; the characters in flow, read left to right, give it).\n132 posts were re-read from their permalinks in the keeper's own signed-in session into\n`2026-10-22_archive.json`, with their dates, reactions and surviving links; the other 274 are\ncarried forward byte for byte with the day they were read. 46 posts have no text of their own on\nthe live page either, shares of content since removed or photos; 3 keep the shared link under\n`attachments`. The one comment this note first counted as feed text (a media link one member left under\nthree posts) is on all three live pages, so it stands. The index (`mannaminne --fbgroup`) and the\npeople register read the new file, take a post's header date where it has one, and anchor a\ncarried-forward post's relative age on the day it was read.\n\n**The rest of the archive, the same evening.** The other 274 posts were re-read between 15:28 and 18:18 by\n`scripts/fb_group_backfill.py launch` (chunks of sixty, a checkpoint every five posts, one retry per page), so\nevery one of the 406 posts carries its header date, 395 their reactions and 111 their attachments; the\ncomment count rose from 1,956 to 1,959, his from 532 to 534. A comment still shows only a relative age (*11y*) with the absolute\ntime behind a hover, which the scraper does not read. The remaining gaps: `TODO.md` § THE FACEBOOK GROUP\nARCHIVE.\n\n---\n\n*Homes this note points at: `docs/history.md` (the collaborators and the timeline),\n`docs/research/simplenote-archive-analysis.md` (the private half of 2011–12),\n`docs/research/structuring-gradient-lineage.md` (the December 2011 line and the January 2012 reversal),\n`docs/research/philosophy-of-the-unwritten-step.md` (the squirrel), the ruling ledger rows AM7, Q33, Q38,\nAG12, AG14, AM5, AL10, AS1, AS2 and S8. The archive itself: `data/fb_group_archive/`; searchable with\n`mannaminne --fbgroup <words>` since 2026-10-22.*\n"}