{"path":"research/the-wiser-civilization-as-successor-carlsmith-and-faggella.md","content":"# The wiser civilization as the successor: Joe Carlsmith and Daniel Faggella on *The Trajectory*, read against the corpus\n\n**Date**: 2026-11-08 · **Type**: a podcast episode read against the corpus · **Status**: analysis. Every\nposition below that is not quoted from a ruling is Claude's and unruled. No ledger rows were opened:\nthe connections are readings of two public positions, not design positions; the two proposals in § 6\nwait on the founder's word and are filed here only.\n\n**Source**: *The Trajectory*, Worthy Successor series, episode 15, *\"Joe Carlsmith - A Wiser,\nAI-Powered Civilization is the 'Successor'\"*, published 2025-11-14, 1 h 52 min. Host Daniel Faggella;\nguest Joe Carlsmith, then a senior advisor at Open Philanthropy, speaking in his personal capacity\n(he had announced his move to Anthropic on 2025-11-03). Faggella's episode article: danfaggella.com/carlsmith1.\nThe transcript used here is YouTube's caption track, saved with timestamps in\n`sources/podcasts/trajectory-015-carlsmith-wiser-civilization-2025-11-14.md`; quotations below are from it, with\nmisheard names restored (the captioner writes *Carl Smith*, *Felli*, *Yakowski*, *finesis*, *canatus*).\nTimestamps are the caption's. A local run with speaker separation finished the same day and came out in\nSwedish (KB-Whisper is trained on Swedish), so it is not a source for quotations; who said a line is\nsettled by the captions' turn marks and the audio.\n\n**The founder's instruction**: *\"Transcribe this and read it. It connects to my aims with Deliberus in a\nfew places. Find them and reflect deeply on how Dan and Joe would like Deliberus. I should reach out\nsoon and tell them about it when it's matured a bit more.\"*\n\n**Read depth**: the whole episode from the captions (22,900 words), Faggella's episode article in full,\nand Carlsmith's essay *On restraining AI development for the sake of safety*, already read in full for\nthe AI-risk intake (2026-10-03; `ai-risk-for-mathematicians-and-the-coordination-game.md`). Faggella's\nearlier Trajectory episode with Eörs Szathmáry was read in September\n(`major-transitions-heredity-and-evolvability.md`). The corpus documents named below were opened this\nsession in the sections cited.\n\n**Who wrote this reading**: a language model. Neither Carlsmith nor Faggella has been told their\nconversation was read here; this page agrees with both more than it disagrees, and says where it\ndisagrees.\n\n---\n\n## 0. In short\n\n**What they say.** Carlsmith's best picture of a *worthy successor* is not an AI but *\"a civilization\nthat develops along the lines that we have most acquaintance with and most reason to trust\"*, with AI\nintegrated as instruction-following tools that *\"feed into the processes in civilization that I trust\"*\n(01:19–01:21). Three criteria: a civilization *\"good at the truth\"*, with *\"philosophical competence\"*\nin reflecting on its own values (01:02); material abundance and *\"political maturity\"*, since *\"the\nstate space of political systems has also gone unexplored\"* (01:06–01:07); and depth, art, love and a\nspirituality defined as *\"wanting contact with reality with a capital R, with truth with a capital T\"*,\n*\"no false gods\"* (01:08–01:14). Faggella agrees on continuity (*\"you can't drop a pin in mind-space and\nexpect value to unfold\"*), weights persistence higher (Spinoza's *conatus* as amoral and necessary),\nand doubts the breathing room: *\"I just don't think we're going to get to catch our breath\"* (01:49).\n\n**What it is to this project.** ⚠ Corrected 2026-11-09 (§ 8). The first version said Carlsmith's *worthy\nsuccessor is a civilization with AI integrated* was the 2011 *\"argument map + human society = hybrid AGI\"*\nformulation said by an alignment philosopher. The two sentences share one claim, that the unit worth caring\nabout is the civilization-level process and not a standalone machine mind, and differ in subject: Phillips's\nanswers whether a shared argument map makes a society a general intelligence (so that Deliberus warrants the\ncare AGI does, the question he was answering); Carlsmith's answers which successor is worthy, and puts machine\nAI inside it as instruction-following tools. He names no argument map, graph or debate structure anywhere in\nthe episode. What he names is the need: *\"good at the truth\"* (01:02), *\"good mechanisms of collective\ndecision-making, how do we solicit people's input in the right way\"* (01:07), *\"truth-related collective\nepistemology in a good way … I'm hoping that we can get the AI to help us with that\"* (01:28), with the\nmechanism left open. Deliberus is one proposed mechanism for that need; it is neither his picture nor the\n2011 sentence. The closer match is Faggella's *Worthy Successor* series, a hand-made Deliberus: twenty-two\nguests' positions on one question, closed each time with *agreement, disagreement, and difference*. Carlsmith names the gap Deliberus is\nbuilt for in his own words: *\"how do you do truth-related collective epistemology in a good way … I'm\nhoping that we can get the AI to help us with that\"* (01:28), and on ethics *\"we have notions like\nreflective equilibrium … and I think we don't currently know all that much beyond that\"* (01:05).\nFifteen meeting points are listed in § 2, with the corpus home of each.\n\n---\n\n## 1. What each of them holds\n\n### 1.1 Carlsmith\n\n- **A good future is recognised by its path, not its end state.** *\"I'm not actually sure that I could\n  recognize a good future from my current perspective … I think more about traits and virtues that are\n  on the path rather than the end state\"* (02:55). Our minds are *\"a radically impoverished mechanism\n  of access and understanding\"* to value, present and future (04:49); yet value beyond our conception\n  is *\"conceivable to us along some trajectory of development\"* (00:00, repeated 01:00).\n- **Metaethics, held fallibly.** *\"My best guess metaethical position is that facts about value hold in\n  virtue, at least in part, of facts about human psychology, current human psychology\"* (15:22), so\n  what makes a future value valuable is *\"its relationship to some process of reflection and growth\n  that has its origins in our current psychology\"*. The content may be *\"radically alien\"*; the\n  grounding is not. Hence the paperclipper's descendants, however advanced and however much they grow,\n  do not disclose value *\"relative to us\"*: *\"I am not such as to endorse even on reflection or even\n  upon growing any old thing\"* (21:16–24:00).\n- **Deep atheism versus allegiance to evolution.** He places much of AI-safety thought at one pole\n  (*\"radical alienation from any kind of trust in nature and natural processes\"*, 26:39) and Teilhard at\n  the other, and refuses the inference *\"evolution led to all the stuff I like, so I should trust where\n  it goes\"*: *\"the answer is importantly no to that\"* (31:00). Power-focused processes, evolution and\n  capitalism among them, *\"are insufficient for stewarding the thing that we care about\"* (00:00, 55:00).\n  A *\"paniply of unworthy successors\"* can coordinate and trade perfectly well and still be a bad future\n  for anyone whose values were shut out (49:31–53:00).\n- **The successor is a civilization.** Asked what traits to carry forward, he moves the question:\n  *\"less like what are the individual traits of our AI systems and something more like what processes\n  in civilization do we want to make sure remain suitably empowered\"* (01:00:30). The AI systems need\n  not be conscious, loving or spiritual to help with consciousness, love or spirituality (01:25:42).\n  They should be *\"something in a broadly similar manner to the assistants that we have today … that\n  follow your instructions, do so safely\"*, superintelligent, *\"very good at truth … very good at\n  philosophy\"* (01:26:43).\n- **Evaluation, of AIs and of civilization, with AI help.** *\"How do we ever know whether civilization\n  is going well?\"* (01:23:05); a mature civilization *\"would not be like, well, GDP is going up\"*\n  (01:24:01); *\"you need AIs themselves to be involved in the evaluation, and I think that would be\n  true of our civilizational evaluation as well\"*.\n- **Safety first, upside second, both real.** Minimum standard: *\"avoid obvious irreversible\n  catastrophe\"*: loss of control, *\"AI-powered authoritarianism and lock-in\"*, bio-extinction (01:29).\n  Then *\"channel the best AI labor we have … into the good processes in civilization … and do that\n  early\"* (01:30). A prudent and coordinated civilization *\"would not accept the current level of\n  risk\"*; the hard part is the international regime (01:35–01:38). The tools built under commercial\n  incentive will serve the upside if safety holds: *\"people want tools that will help them do stuff they\n  care about better\"* (01:32).\n\n### 1.2 Faggella\n\n- **The flame.** What dolphins, apes and humans carry: sentience and self-creation (*autopoiesis*).\n  Two imperatives: *\"don't let the flame go out\"* and *\"do whatever helps the flame of value\n  proliferate\"* (01:31). Throwing the baton to an unknown mind and hoping it catches is *\"dumb\"*; so is\n  *\"let it rip\"* accelerationism, *\"patently against everything I have written for 14 years\"* (35:30).\n- **Spinoza's three words.** *Conatus* (the impetus to persist), *potentia* (the powers that let a thing\n  persist, which expand), *potestas* (the collective bounding of individual powers so the whole has\n  more). *\"You could win a lot of soccer games with an AK-47, but it's just not in the rules\"*; the\n  labs today have *\"no soccer game\"* and no *\"international banking rules\"* (43:36, 01:32).\n- **Persistence is amoral, not immoral.** *\"Value depends upon the thing persisting\"*; he weights\n  survival higher than Carlsmith and expects conatus to remain the substrate of any future mind, while\n  Carlsmith leaves open another core impetus (01:43–01:46).\n- **The method of the series.** *\"My purpose is to get the goddamn ideas on the table\"* (10:15). He\n  restates the guest's position and asks *\"am I picking up what you're putting down?\"* (13:34, and\n  again at 23:30, 37:00, 01:22); he keeps *\"a draft list of what I suspect to be worthy successor\n  traits and I beckon others to give me theirs\"* (56:29); and he closes every episode with *\"areas of\n  agreement, disagreement, and some of the really unique ideas\"* (01:39:49), here with a third category,\n  *\"I call this difference\"*: a divergence of intuition that is *\"not a disagreement with Joe\"*\n  (01:47:24).\n- **The breathing-room doubt.** Carlsmith's successor echoes Toby Ord's *long reflection*; Faggella\n  finds it *\"a delightful thing to hope for and not a naive thing to strive for\"* but doubts the\n  *\"forces of destruction and transformation\"* (the AGI race, immersive AI, AI relationships, perhaps a\n  civilizational dumbing-down) will permit the plateau (01:48–01:50). So he puts weight in parallel on\n  making sure whatever follows us carries sentience and autopoiesis.\n\n---\n\n## 2. Where the episode meets the corpus\n\nFifteen places. Each gives the passage, the corpus home, and what the meeting is (a confirming\ninstance, a sharpening, or a challenge). Readings are Claude's unless a ruling is quoted.\n\n1. **The successor is a civilization with AI integrated: one claim shared with the 2011 sentence, not the\n   same position** (corrected 2026-11-09, § 8). Carlsmith, 01:16: *\"It's less like here's an AI … that I'm now\n   happy to hand the future to. It's much more like what is an ongoing blossoming of the processes at work in\n   human civilization that I trust, with AI integrated into that.\"* The corpus's 2011 sentence is Ivan\n   Phillips's answer to the founder's question whether Deliberus *\"warrants the same kinds of considerations\n   that AGI does\"*: *\"A system of (argument map + human society) might be a hybrid AGI\"* (`vision.md`;\n   provenance in `simplenote-archive-analysis.md` § 3). Phillips's sentence is about what a shared map makes\n   of a society; Carlsmith's is about which successor to want, with the machine a tool inside it and no map,\n   graph or argument structure named (the transcript's only *graph* is Faggella's word for Bostrom's\n   state-space picture, its only *argument* is *\"makes exactly the same argument\"*). By the keeper's keys the\n   pair reads *overlaps* (one shared claim: the civilization-level process is the unit that matters), not\n   *same*; the first version wrote *same* on a shared locus and a shared word, which is the topic-overlap\n   error the keeper is built to refuse (`three-minds-and-the-leap.md` § 5a). What does carry over is the\n   aim: the founder's 2012 draft (*\"some algorithm for humanity to reason and make decisions together\"*) and\n   the root purpose in `what-deliberus-is-for.md` § 3 (*humanity can think and decide together well enough\n   to come through*) say as an aim what Carlsmith asks for at 01:07. **An overlap**, and a statement of the\n   need from a field that usually names a machine as the successor (Faggella expected him to be *\"more\n   anthropocentric\"* and was surprised).\n\n2. **\"Good at the truth\" and \"collective epistemology in a good way\": the missing layer named.**\n   Carlsmith, 01:02: *\"we need to have processes in play in civilization that can see clearly what's\n   going on … a huge thing that AI should do is to help us with that\"*; 01:28: *\"truth-related\n   collective epistemology in a good way … I'm hoping that we can get the AI to help us with that.\"*\n   `the-missing-layer.md` says civilization has infrastructure for conclusions, quality control,\n   agreement-shape and procedure, and none for reasoning itself; the harness conviction\n   (`the-harness-conviction.md`) says machine reasoning's quality is a function of harness and\n   grounding, and Deliberus is the harness. **A confirming instance with a sharpening**: Carlsmith\n   wants the AI to *help* collective epistemology and does not say how; the corpus's answer is a\n   record where reasons are addressable and the machine proposes and never asserts. He is naming the\n   slot the project claims.\n\n3. **\"Philosophical competence as a civilization\": the gap the descent fills.** Carlsmith, 01:02–01:06:\n   *\"what sort of skill and subtlety and depth do you bring to processes of reflecting and growing\"*;\n   on ethics, *\"we have notions like reflective equilibrium, we have notions like you shouldn't be\n   directly contradictory, you shouldn't be acting on the basis of false beliefs about the world, and I\n   think we don't currently know all that much beyond that as to what it is to do ethics well, but we\n   need to do it anyway … I would want a civilization that has directed quite a lot of serious effort\n   towards that goal.\"* The corpus's descent is a concrete proposal for exactly this: the critical\n   questions of each kind of argument (`schemes.py`), the separability floor (`depth.md`), the horizon\n   kinds that say what a disagreement currently bottoms in (`what-belongs-in-the-ontology.md` § 6c), and\n   the palette of five levels on which value decomposes, down to its two floors, the phenomenal and the\n   normative (`decomposing-value.md`). **A confirming instance and a\n   stranger test passed**: he names the three tools he has (equilibrium, consistency, true beliefs);\n   Deliberus's descent uses all three and adds the addressable record and the typed horizon. One\n   register he names, *\"philosophy that is neither ethical nor straightforwardly factual, like\n   metaphysics\"*, is a register the horizon enum's is/ought split does not hold, which\n   `what-belongs-in-the-ontology.md` § 2 already found from the inside.\n\n4. **His metaethics is the convergence wager's grounding, and the paperclipper caveat is the\n   \"bedrock of human values\" line.** Carlsmith, 15:22–24:00 (§ 1.1). `convergence.md` places the\n   wager closest to coherent extrapolated volition, *\"what humanity would want if we knew more, thought\n   faster\"*, with a mechanism; the founder's 2026-03-31 line on agents is *\"the bedrock consists of\n   human values\"* (`agents-as-a-consumer-class.md`), and BF1 (ruled 2026-11-03) says a record of\n   reasoning reaches a model's stated reasoning and never its policy. **A confirming instance**: his\n   *\"some process I would endorse via which at the end of it I would see and recognize the value\"* is\n   what the graph is built to record: the process, attributed, with the horizon marking where endorsed\n   reflection has reached. His caveat that *not any advanced mind's disclosure counts* is why the\n   corpus keeps the agent surface read-only and the bedrock human. **One difference worth keeping**:\n   he grounds value in *current human psychology plus endorsed growth*; the corpus's line on good hearts\n   (BD28) and its line on the theorems behind the wager (BD27) are claims about *where people's reasons\n   meet*, not about what grounds value. The wager is compatible with his anti-realism and with the\n   realism of his friends, and says so (`good-hearts-and-permissivism.md` § 2). One caution he would\n   share: Codex's reading of 2026-10-28 narrows what the cited theorems establish (Aumann needs a common\n   prior over the whole state space, Blackwell and Dubins need absolute continuity), and that narrowing\n   stands in the corpus unanswered (`imagined-graph-extraction-and-checkable-reasoning-2026-10-28.md` § 7).\n\n5. **\"How do we ever know whether civilization is going well?\": the instruments, with a cost of\n   now.** Carlsmith, 01:23–01:25: civilizational metrics for *\"how are we doing on truth, how are we\n   doing on philosophy\"*, richer than GDP, with AI in the evaluation. The corpus's instruments measure\n   the *graph*: the horizon map's wager fraction (`convergence.md`), disagreement preservation, the\n   confession principle, counts labelled by who stands behind them. **A sharpening**: a graph whose\n   coverage is wide would be one civilizational metric of the kind he asks for (what fraction of a\n   civilization's live disagreements bottom in facts, in words, in weights). Today it measures a few\n   dozen sources, so this is a cost of now (BD18), not a limit of the design. **A caution he shares\n   without saying so**: the machine grading its own homework; the corpus removed `self_eval`'s score\n   in August for that reason, and the harness conviction's ceiling (coherent absence, frame lock-in)\n   is the part of his *\"AIs involved in the evaluation\"* that a machine cannot supply.\n\n6. **AI as instruction-following tools, not conscious, feeding trusted processes: propose, never\n   assert.** Carlsmith, 01:21–01:27 (§ 1.1). The corpus: *propose, never assert* (canon); microtasks\n   as one small judgment each; a read-only agent surface; `the-disempowerment-objection.md`, where the\n   founder answers the same worry (*\"AI in decision-making means giving up human control\"*) with *an\n   AI that decides and an AI that lays every reason open so people decide are opposite answers to the\n   same risk*. **A confirming instance**: Carlsmith's tool picture is the project's machine role stated\n   from the alignment side. His point that the AI need not have the trait to help the trait flourish is\n   the harness conviction's claim that the remedy is independent of the model's architecture (*nobody\n   upgraded the brain, they built the notation*).\n\n7. **Faggella's method is Deliberus by hand.** *\"Get the goddamn ideas on the table\"* (10:15); the\n   habit of restating the guest's position and asking for confirmation (13:34 and later); the trait list he *\"beckons others to give me theirs\"*\n   (56:29); the closing taxonomy *agreement / disagreement / difference* (01:39–01:48). Three homes.\n   The restating and asking is **consent**: a person ruling on a rendering of their own position, the\n   one act no majority can win (`what-human-judgment-is-for.md`, blessed end-state 2026-08-31). The\n   trait lists across twenty-two guests are attributed claims on one question, which is what a source\n   page and the always-mint ruling hold. And his *difference* category, a divergence of intuition that\n   is not a disagreement, is a **horizon kind read by ear**: the persistence question (§ 2.11) bottoms\n   partly in weighting and partly in a word. **A sharpening**: he does by voice, once per episode and\n   lost to the listener's memory, what the graph does persistently and addressably. This is the\n   strongest practical connection in the episode.\n\n8. **Twenty minutes on \"are we disagreeing or using different terms\": sameness before relation, and\n   the concept layer.** From *deep atheism* (26:39) to *\"it's possible we just don't disagree then\"*\n   (34:11) to Carlsmith's *\"I would actually want to be drawing a bunch of distinctions between things\n   like survival, self-interest in a narrow sense, power\"* (about 46:00), the two spend a fifth of the\n   episode finding out whether Faggella's conatus/potentia/potestas and Carlsmith's power/selection/\n   survival name the same things. The ruling that sameness is decided before any relation is written\n   (2026-09-13, built as the keeper), the concept layer's core and periphery\n   (`core-periphery-and-the-concept-layer.md`), and the opacity conviction's split between a conflict of confusion\n   (dissolved by understanding) and a conflict of interest (only clarified) are all exhibited live. **A\n   confirming instance of the thesis on its own ground**: two careful thinkers needed twenty minutes\n   and still ended at *\"maybe I'm just using different terms\"*; a map would have shown the two\n   vocabularies against one set of claims and left the residue, Carlsmith's distrust of selection for power\n   as a *normative standard*, as the one thing in dispute.\n\n9. **\"Strata of value we can't articulate\": the confession slot in a taxonomy of values.** Faggella,\n   06:40–10:00, against Yudkowsky's *\"kindness and fun\"* as rendered by Faggella and Deutsch's *\"we are\n   Turing machines\"*: may there be *\"entire strata and categories of value for which you and I have no\n   ability to articulate\"*. The three-runs law (every taxonomy breaks on the next register) and the\n   `does_not_fit` value, the slot an enum keeps for what fits none of its members (`taxonomy-gaps-and-the-closed-enum.md`) are this demand applied to the project's\n   own enums; the finding in `decomposing-value.md` § 2b that *the primitive count is the wrong axis* is the\n   founder's own version of Faggella's point. **A confirming instance.**\n\n10. **\"Radically impoverished\", value \"conceivable along some trajectory\": nothing is permanently\n    terminal.** Carlsmith, 04:49 and 21:16 (the *treasure chests* we should not leave unopened). The\n    canon line *Nothing is permanently terminal* (ruled 2026-09-15, the explorers sentence added\n    2026-10-26) says a horizon is where shared mapping currently ends, moving as the corpus grows, and\n    *\"explorers reaching beyond the current horizon to uncover new ones is a beautiful image\"* (the\n    founder's words). **A confirming instance**, with a difference in scale: Carlsmith's horizon is the\n    species' current psychology; the corpus's is a corpus's current arrangement. Both refuse to call the\n    present edge the end, and both say the edge moves only along a trajectory one can endorse.\n\n11. **Conatus and autopoiesis: Heylighen's bottom, and the Szathmáry episode.** Faggella's flame is\n    sentience plus autopoiesis; his substrate for any future mind is conatus. The corpus's naturalistic\n    bottom for a descent is Heylighen's autopoietic valence (`heylighen-value-complexity-and-deliberus.md`:\n    values originate in self-maintenance; *the AI danger boundary is selection, not intelligence*),\n    which is Carlsmith's distrust of selection for power stated as biology, and Faggella's *conatus* as\n    the layer under value. The reuse flywheel is argued in assembly-theory terms as *raising the\n    assembly ceiling of arguments that can persist and replicate* (`assembly-theory-and-the-reuse-mechanism.md`):\n    persistence of arguments, not of agents. Faggella's own earlier episode with Szathmáry supplied\n    the corpus with *\"a higher level organization always suppresses the evolutionary potential at the\n    lower level\"* and his closing on red lines versus green arrows\n    (`major-transitions-heredity-and-evolvability.md` § 9). **A sharpening**: the persistence question\n    he and Carlsmith leave as a *difference* is decomposable, and part of it is a word: Faggella's\n    *amoral* and the *immoral* he hears in the AI-safety milieu are two senses of *moral*, which the\n    concept layer would hold as senses with holders.\n\n12. **\"No soccer game for the labs\": the coordination rung, and what a public map can serve.**\n    Faggella, 43:36 and 01:32; Carlsmith, 01:35–01:38 (*\"the appropriate sort of regulatory regime that\n    doesn't allow private actors to risk destroying the lives of every single human\"*; the international\n    piece is the hard one). Carlsmith's essay already supplies the corpus's first bottleneck,\n    *\"widespread political will is the first and most central bottleneck to serious capability\n    restraint\"*, and `shared-sensemaking-what-deliberus-holds-and-owes.md` § 1 says what a public map\n    can serve is that will, not the racers' bilateral assurance, which runs on classified information no\n    public map holds. BF5 (ruled 2026-11-03): a record binds nobody by what it shows; it coordinates\n    where a party has committed in advance to be bound by it. **A confirming instance of the corpus's\n    modest claim**: neither of them expects a map to make the soccer game; both expect public will and\n    a regime. The corpus says the same, and says which half the map reaches.\n\n13. **Spirituality as contact with reality, \"no false gods\": attunement that contains analysis.**\n    Carlsmith, 01:08–01:14: spirituality is *\"the world comes forth as more meaningful, as more alive,\n    as more beautiful, as more there and more real\"*, and *\"any sort of compromises on truth are in my\n    head compromises on the central spiritual impulse\"*. The corpus's core dialectic\n    (`analysis-and-attunement.md`) holds that the two poles may not eat each other, and the sacred\n    re-scope (ratified 2026-08-25, `parity-and-the-zone-interior.md`) says *no stop signs*: analysis may\n    go anywhere and owes attunement in proportion to depth. Ginges and Atran's result that on sacred\n    values argument produces outrage and symbolic recognition produces flexibility\n    (`does-descending-change-the-descender.md` § 9) is the empirical side of his *\"harmonized with and\n    in dialogue with the stuff about truth\"*. **A confirming instance**, and one the founder will\n    recognise from his own lenses: Carlsmith's word is *disclosure*, the register of Vervaeke and of the\n    phenomenologists already in the corpus (`reflection-and-the-graph.md`).\n\n14. **\"The state space of political systems has gone unexplored … how do we solicit people's input\n    in the right way\": the root purpose.** Carlsmith, 01:06–01:07, adding *\"even holding human nature\n    fixed, I bet we can do way better than we do now\"*. The root purpose (`what-deliberus-is-for.md`\n    § 3) and the 2012 draft are this ambition; the compartments (blobs) with their own written rules and\n    the advice-process rung (`blobs-community-sandboxes-and-the-advice-process.md`) are its smallest\n    live test. **A confirming instance.**\n\n15. **\"You need AIs themselves involved in the evaluation\": the evals loop, and the second-model\n    caveat.** Carlsmith, 01:24. The corpus's standing work and error-rate discipline\n    (`proving-a-pass-reliable.md`, `the-evals-loop-and-our-machine-passes.md`) run exactly this on the\n    machine's own output, with the canon caveat *a second model is not a second observation* (R7) and\n    the harness ceiling. **A confirming instance with a limit he does not state**: his picture has the\n    AI evaluating the AI; the corpus's measured instance is that correlated errors rise with accuracy,\n    so the independent reading must come from a person or a different family.\n\n---\n\n## 3. How Carlsmith would read Deliberus\n\n**What he would recognise at once.** His first criterion built as infrastructure rather than hoped\nfor; his tool picture of AI (instruction-following, non-agentic, not needing the trait to serve it)\nas the machine's role; *propose, never assert*; a wager held the way he holds his metaethics (*\"which\nI'm not sure it is, to be clear\"*); and a horizon that says where endorsed reflection has reached\nwithout calling it the end. The episode's own best sentence for the pitch is his: *\"how do you do\ntruth-related collective epistemology in a good way.\"*\n\n**Where he would press, in order of weight.**\n\n1. **Safety first.** He would place Deliberus on the upside side of his two-part program and ask why\n   it matters before catastrophe is avoided. The corpus's answer is his own: the first bottleneck to\n   restraint is political will, a public picture serves exactly that, and nothing else in the project's\n   claims reaches the racers' assurance (§ 2.12). Deliberus is not a safety case; it is infrastructure\n   for the public half of the coordination he says is hardest.\n2. **Steering.** An AI that helps people reflect on values can nudge them before anyone measures it.\n   He did not raise this in the episode; his essays on giving AIs safe motivations suggest he would. The\n   corpus has the DeepMind facilitation result (N=879; preferred, no better consensus, allocations\n   moved) and the design rule it produced: show structure, ask questions, never restate their words\n   (`live-election-test-design.md`; CLAUDE.md § Standing Epistemic Threat Model, entry 2). He would\n   want the numbers of the instrument that measures whether a rendering preserves the disagreement in its source (`disagreement_preservation.py`), and the honest statement that every\n   source so far is machine- or operator-extracted.\n3. **Over-commitment in the ontology.** His *\"we don't know much beyond reflective equilibrium\"* could\n   cut the other way: does a typed horizon enum encode a metaethics? The corpus's answer is the sorting\n   test (a commitment belongs in the graph if two users can disagree about it and the system keeps\n   working, `what-belongs-in-the-ontology.md`), the three responses (de-bake, confess, publish), and\n   the ruling that a horizon kind is *read, never set*. He would also find the is/ought split in the\n   enum, which the corpus found first.\n4. **What the graph reaches in a model.** Now at Anthropic, he would ask whether a record of reasoning\n   is an alignment instrument. BF1 (ruled 2026-11-03) says no: a check against the record reaches a\n   model's stated reasoning and never its policy; the project claims the layer that says *what*\n   alignment is to, in a form that holds disagreement, and never a way to know whether a model is\n   aligned (`alignment-infrastructure-red-team.md`). That is the humble claim he would accept and the\n   overclaim he would catch if the March text were still up (it is not).\n5. **His axis from deep atheism to allegiance to evolution.** He would ask whether the wager trusts nature. It does not trust selection;\n   it bets that people's reasons, decomposed, meet more than their conclusions do, and it counts hinge\n   and parity ceilings and permissivism against itself (`convergence.md`). That is a position between\n   his two poles, and he said in the episode that he lives there too (*\"I'm sympathetic to both of\n   these impulses\"*, 31:30).\n\n**What he would like most, if a guess is allowed**: that the map shows the reasons and not a verdict;\nhis whole episode resists verdict-bearing machines and asks for machines that make human reflection\ndeeper.\n\n## 4. How Faggella would read Deliberus\n\n**What he would recognise at once.** A tool for the thing his series is: positions on one question\nfrom many minds, kept, attributed, with the agreements, disagreements and differences drawn rather\nthan spoken into an outro. The consent act he performs in every interview (*\"am I picking up what\nyou're putting down?\"*) as a built mechanism. His three-way closing as horizon kinds. A home for his\n*draft list of worthy successor traits* beside every guest's, which is what he says he wants (*\"I\nbeckon others to give me theirs\"*). And at the bottom, autopoiesis honoured (§ 2.11): the corpus's\nnaturalistic floor is the thing he calls the flame.\n\n**Where he would press.**\n\n1. **Breathing room.** *\"I just don't think we're going to get to catch our breath.\"* If the long\n   reflection never comes, infrastructure for it is a beautiful thing that is never used. Two answers,\n   both already in the corpus. First, the near-term register is not reflection but checking: the\n   promise is *checking that keeps pace with persuasion* (`the-scrutiny-gap.md`), which is wanted most\n   exactly in the turbulence he predicts. Second, the success prior's fifth consequence: absence of\n   traction in the flat phase is not evidence against approaching the threshold, and the surviving\n   strategy is cheap persistence (`fractal-priors-for-success.md`). His doubt is the strongest outside\n   statement the corpus has of the view that no plateau comes; it belongs beside the signposts\n   (`signposts-for-the-whole-project.md`) as the assumption it tests.\n2. **What does it do for the arms race?** He has just come from the UN General Assembly and Shanghai.\n   The corpus's answer is modest and he would respect it for that: a shared assessment is necessary and\n   not sufficient; it serves political will; it does not replace verification or the regime (§ 2.12;\n   the same note's warning that sensemaking without a coupling to action *changes no outcomes*,\n   `zizek-schmachtenberger-connections.md` § 3.4, is his own worry stated earlier).\n3. **His persistence weighting.** He would want his *conatus is amoral* position mapped against the\n   milieu's *immoral*, and would find that part of it is definitional and part a weighting (§ 2.11).\n   He would enjoy that, since he said in the outro that this is the one place he and Carlsmith part.\n\n**What he would like most**: that the thing he does by voice can be done by a graph, and that it\nwould make the Worthy Successor series' twenty-two positions one object a newcomer can walk.\n\n---\n\n## 5. What the episode adds to the corpus, and what it challenges\n\n- **Adds a statement of the need from the alignment side, and one method match.** Carlsmith names the\n  need (collective epistemology, soliciting input, mechanisms of collective decision-making) and the\n  machine's role (tools that help, never minds that decide), and shares with the 2011 sentence only the\n  locus (§ 2.1, corrected 2026-11-09). The first version counted *four independent arrivals* at one picture\n  (Wright's noosphere, Heylighen's global brain, Szathmáry's transition, Carlsmith); they share the locus\n  and differ in claim (a collective mind, a self-organising global brain, a transition with its price, a\n  worthy successor), so the count is withdrawn. Faggella's series is the method match (§ 2.7). Nothing new\n  enters the canon.\n- **Challenge 1, the breathing room (§ 4.1).** A standing doubt about the premise that a deliberation\n  layer gets its moment before the transition. Candidate signpost, unruled: *the register that matters\n  in turbulence is checking, not reflection; if the project's uptake comes only from reflective settings,\n  Faggella's doubt is being confirmed.*\n- **Challenge 2, civilizational metrics (§ 2.5).** Carlsmith asks for measures of how a civilization is\n  doing on truth and philosophy. The corpus's instruments measure a graph. The honest line is that a\n  wide graph would be such a measure and today's is not; stating it keeps the project from the\n  overclaim he would catch.\n- **Enrichment 1.** Faggella's *difference* category (a divergence of intuition that is not a\n  disagreement) is a lay name for a weighting horizon; worth remembering when the horizon kinds are\n  shown to strangers.\n- **Enrichment 2.** The twenty-minute vocabulary reconciliation (§ 2.8) is a clean, public, two-expert\n  specimen of confusion-conflict dissolving under decomposition, with a small real residue left. It is\n  usable as an example on the pages that explain sameness and the concept layer.\n- **Enrichment 3.** Carlsmith's *\"philosophy neither ethical nor straightforwardly factual\"* joins\n  `what-belongs-in-the-ontology.md` § 2's finding about the horizon enum from outside.\n\n---\n\n## 6. What to show them, and two proposals (unruled, filed here only)\n\n**Proposal A: map the Worthy Successor series as a dogfood cluster.** Twenty-two episodes, each with a\nguest's criteria for a worthy successor, Faggella's own list, and his stated agreements, disagreements\nand differences. Every criterion a claim attributed to its holder; the contested words (*worthy*,\n*successor*, *value*, *consciousness*, *the flame*) in the concept layer with their senses and holders;\nthe cross-guest links drawn by the keeper and read by a person first. Substance test: posthuman ethics\nis a substantive debate, not argumentation theory, so the Pipeline Testing Principle (substance over meta) is met. Precedent\nfor the source choice: the credit-run candidates of 2026-11-05 (he chooses the sources). Cost of now:\nextraction runs from the keeper's machine only, and the keeper's held-out pairs (the nine real pairs no\nprompt example shows) are read first. The\ndemonstration for Faggella would be his own series walkable as one map.\n\n**Proposal B: this episode as a hand-made map.** Carlsmith's three criteria, Faggella's two imperatives,\ntheir agreement (continuity), their disagreement (persistence, amoral against immoral), their difference\n(a levelled-up civilization as the successor, against no breathing room before the transition), with each side's reasons written to the separability\nfloor as the institute map was. Žižek's parallax, already in the corpus (`zizek-schmachtenberger-connections.md`\n§ 1.3), is the name for what Faggella calls a *difference*: a divergence no added information removes, which\nthe map marks as such rather than resolving. Smaller, and it shows both of them their own conversation with the\nresidue typed.\n\n**The outreach itself.** Telling people about Deliberus is the founder's and has no gate; asking them to\ntry it waits on the eight parts (ledger AV2, and AV4 ruled 2026-11-02: *the eight parts gate the public\nrung alone*). His word today: *\"I should reach out soon and tell them about it when it's matured a bit\nmore.\"* The relationship facts and the pitch angles per person are held privately.\n\n---\n\n## 7. Sources\n\n- *The Trajectory*, Worthy Successor ep. 15, Joe Carlsmith with Daniel Faggella, 2025-11-14. Audio on\n  Buzzsprout; video youtu.be/bsTwkg6eYCs; caption transcript in\n  `sources/podcasts/trajectory-015-carlsmith-wiser-civilization-2025-11-14.md`.\n- Daniel Faggella, episode article, danfaggella.com/carlsmith1 (14 November 2025, with his criteria\n  list and *Concluding Notes*); *A Worthy Successor*, danfaggella.com/worthy; *Our Final Imperatives*,\n  danfaggella.com/imperatives (the latter two not read for this note).\n- Joe Carlsmith, *On restraining AI development for the sake of safety*, 19 March 2026 (read in full\n  2026-10-03, `ai-risk-for-mathematicians-and-the-coordination-game.md`).\n- Corpus homes cited in § 2: `vision.md`, `convergence.md`, `the-missing-layer.md`,\n  `analysis-and-attunement.md`, `depth.md`, `what-deliberus-is-for.md`, `the-harness-conviction.md`,\n  `what-human-judgment-is-for.md`, `what-belongs-in-the-ontology.md`, `decomposing-value.md`,\n  `heylighen-value-complexity-and-deliberus.md`, `assembly-theory-and-the-reuse-mechanism.md`,\n  `major-transitions-heredity-and-evolvability.md`, `shared-sensemaking-what-deliberus-holds-and-owes.md`,\n  `alignment-infrastructure-red-team.md`, `the-disempowerment-objection.md`,\n  `core-periphery-and-the-concept-layer.md`, `taxonomy-gaps-and-the-closed-enum.md`,\n  `parity-and-the-zone-interior.md`, `does-descending-change-the-descender.md`,\n  `fractal-priors-for-success.md`, `the-scrutiny-gap.md`, `good-hearts-and-permissivism.md`,\n  `proving-a-pass-reliable.md`, `the-evals-loop-and-our-machine-passes.md`,\n  `blobs-community-sandboxes-and-the-advice-process.md`, `live-election-test-design.md`,\n  `signposts-for-the-whole-project.md`, `wright-god-test-noosphere-and-deliberus.md`.\n\n---\n\n## 8. Correction, 2026-11-09: one shared claim was written as the same position\n\nHis words, the morning after the walkthrough: *\"That elder passage and the initial part of the podcast\nreflections, they feel a bit silly treating the \"hybrid AGI\" idea from more than a decade ago as the\nequivalent of modern AI done safely (which is (an intermediate step) what Joe wants iirc?) or as the\nequivalent of Deliberus? Does Joe mention argument graphs? He DOES mention \"getting good at truth\" and\ncollective sensemaking or whatnot later. I'm just curious. Maybe I'm missing something. But we need to be\nmaking relevant distinctions and have discernment throughout all this.\"*\n\nHe was not missing anything. Checked against the caption transcript (23,345 words): Carlsmith mentions no\nargument map, graph, debate structure or deliberation tool; the word *graph* occurs once, Faggella's\n*\"graphic\"* for Bostrom's state-space picture (00:20), and *argument* once, *\"makes exactly the same\nargument\"* (00:10). His recollection of Carlsmith's order is right: safety first (*\"avoid obvious\nirreversible catastrophe\"*, 01:28), then *\"take … the best AI labor we have\"* into the good processes; the\nwiser civilization with safe instruction-following AI inside it is the end picture, and safe AI the\nintermediate step (§ 1.1).\n\nThree things were run together in § 0, § 2.1, § 5 and in the walkthrough's part 1 (lines *\"Our vision page\nhas carried that sentence since December 2011 … He reached it from the other side\"* and the elder's two\nwitnesses):\n\n| | Subject | Claim |\n|---|---|---|\n| Phillips, 2011 | what a shared argument map makes of a society | it might be a general intelligence, so Deliberus warrants AGI-grade care (his answer to the founder's question) |\n| Carlsmith, 2025 | which successor is worthy | a civilization continued along the processes we trust, with machine AI as tools inside it; the need for collective epistemology named, mechanism open |\n| Deliberus | a mechanism | a graph of attributed reasons, machine proposing and never asserting |\n\nBy the keeper's keys (the sameness microtask, `deliberus/keeper.py`) the first two read *overlaps*: one\nshared claim, that the civilization-level process is the unit that matters, and not a standalone machine\nmind. The note wrote *same*, on a shared locus and the shared word *AI*, which is the keeper's topic-overlap\nerror done in prose (`three-minds-and-the-leap.md` § 5a). The elder's *two witnesses* therefore testified to\ntwo matters, and its own caveat (*\"witnesses who read the same books are one witness twice\"*) was not run:\nPhillips wrote among the same rationalist milieu Carlsmith's work comes from, so independence was asserted,\nnot checked.\n\nRepaired this morning: § 0, § 2.1, § 5 here; the see-also in `wright-god-test-noosphere-and-deliberus.md`;\nthe kept walkthrough's header (its spoken lines left as spoken, the audio unchanged on his word). What\nstands: § 2.2 (he names the slot, not the mechanism), § 2.6 (the machine's role), § 2.7 (Faggella's\nmethod), § 3 and § 4.\n\n**Candidate rule, unruled, filed here only.** A reading note that calls an outside passage a *confirming\ninstance* of a corpus position writes the two sentences side by side and names the key (*same*, *contains*,\n*overlaps*, related by topic only), as the keeper must before it writes a link; a walkthrough may call two\nsources two witnesses to one matter only on *same*, with their independence checked. The fit note the\nelder's lines already carry (*applies · breaks*) is the place the check belongs, and it was written and not\nrun.\n\n## From the Topic Index\n\n*The `CLAUDE.md` Topic Index entry that pointed here, word for word as it stood on 2026-11-15, before the index was trimmed to the warnings only it carries. The index keeps a short entry; the full text lives here.*\n\n- **\"Does an alignment philosopher want what Deliberus builds?\" / Carlsmith and Faggella on The Trajectory / the worthy successor as a civilization with AI integrated / \"collective epistemology in a good way\" / how Dan and Joe would read Deliberus** → `docs/research/the-wiser-civilization-as-successor-carlsmith-and-faggella.md` (2026-11-08; no ledger rows). Carlsmith's worthy successor is a civilization with AI as instruction-following tools feeding the processes it trusts: one claim shared with the 2011 hybrid AGI sentence (the civilization-level process as the unit), not the same position, and no argument map named (corrected 2026-11-09, its correction section); he names the gap the project is built for (*\"how do you do truth-related collective epistemology in a good way\"*, and on ethics *\"we don't currently know all that much beyond\"* reflective equilibrium). Faggella's series is Deliberus by hand: positions on one question, the guest's rendering checked back (consent), closed with agreement, disagreement and *difference* (a weighting horizon by ear). Fifteen meeting points with homes; where each would press: Carlsmith on safety first and steering, Faggella on breathing room (*\"I just don't think we're going to get to catch our breath\"*). **Do not say the corpus's instruments measure a civilization**: they measure a graph, and a wide graph would be the metric he asks for (a cost of now). **A reading note that calls an outside passage a confirming instance names its key** (same, contains, overlaps, related by topic only), with the two sentences side by side: the first version wrote *same* on a shared locus and a shared word (candidate rule, unruled).\n"}