{"path":"frontier.md","content":"# Frontier: Where Our Thinking Is Right Now\n\nThis is the only page on this site that moves. Everything else describes something we believe or have built. This page describes where we're still figuring things out — honestly, openly, in the same voice. It's regenerated from the project's frontier tracking every time the thinking shifts; if something here looks confident, check the date, and if something looks unresolved, it is.\n\nThe graph has no permanent bottom. Neither does our understanding of how to build it.\n\n---\n\n## Ontology\n\nThe graph has one node type: Claim. `ClaimBase` reflects that unity in code. The self-similar decomposition principle means nothing is permanently atomic, and the mother-claim pattern means bundled claims should decompose into independently challengeable subclaims with transitive strength flow. The July 2026 instrument layer is live in code — completeness oracle, residue typing with challengeable verdict-claims, disagreement preservation, hinge score, agent-readable surface — and a classified terminus is a fallible fixed point under the current decomposition move-set, never proven exhaustion. The residue map has settled accounting: the compression is the product even for residues that never resolve, typed residues have different afterlives (reframing, analogy-transfer, scope-narrowing, and the analysis-to-attunement handoff are resolution operators beyond decomposition), a well-stated open problem is a Hilbert-grade artifact, and — the rule that makes it a wager rather than a ratchet — hinge-clash, parity and permissivism ceilings count AGAINST convergence rather than as work remaining. Two findings from hand-run extractions are stable: **the crux of a real debate is typically stated on one side and left implicit on the other** — in run 6's legal pair it was unwritten on *both* — which makes implicit-premise recall the pipeline's most important quality axis, and register drives residue type (worldview clash → structural, evidentiary clash → mostly empirical and dissolving). A third is now stable across three consecutive runs: **every taxonomy in the ontology carries exactly the distinctions of the register it was induced from**, and breaks on the next one (claim type on a speaker-attitude hedge, residue type on an is-bedrock terminus, epistemic status on doctrinal inference). The gaps are not one gap — two are missing values, one is a missing dimension — and the standing fix is a confession channel rather than a slot per register, since a closed enum with no \"does not fit\" value cannot report a missing slot and so returns the nearest fit and reports success. **The sacredness brake has stood unratified since Jul 10**, its candidate re-scope being consent-and-timing rather than epistemic exemption. See [self similar decomposition and claim ontology](research/self-similar-decomposition-and-claim-ontology.md), [depth](depth.md), [convergence](convergence.md), [frontier extraction experiment](research/frontier-extraction-experiment.md).\n\nFive live edges. Flattening has **five known mechanisms**, each blind to the others; translation-on-ingest is the only one applied by policy. 2026-08-18: **the instruments measure collapse, not amplification** (over-conviction passes cleanly); the unrubriced quality-score judge was removed rather than calibrated. The third edge is **decomposition itself, under-specified rather than under-built**: critical questions decompose an *inference*, one axis of eight (predicate–argument the highest-value unbuilt one); *decompose as far as possible* has no target, replaced by **decompose along an axis only when someone contests along it**; decomposition **degrades** verification without per-sub-claim evidence alignment, answered by routing evidence through links rather than copies; and category theory contributes twice (**losslessness is a colimit**, **overlapping subtrees are a sheaf gluing condition**). A fourth edge (2026-08-20, from Heylighen): the weighing descent's **contextual exit is ruled first-class and proposable from priors** — the exit is an entrance, mapping the situated structure instead of stopping at \"it depends\"; additivity in strength is on notice; and the evidence-division root question resolved to confirmation holism with an architecture answer — link evidence, never divide it, confess correlation exposure beside the badge (evidence-division-and-the-foundation.md). First decision off the decomposition edge, live: **reported content is always minted as its own claim** — buying a straw-manning detector, though an audit found both of its safeguards satisfied only by accident. A fifth edge opened 2026-08-22: **the ontology is already a theory**, and the test that sorts it is whether the system could hold two users disagreeing about a commitment and keep working — which finds the terminus enum encoding an is/ought split invisibly, and strong particularism unrepresentable *and* scored as a copout. Responses ordered by cost: de-bake, confess, publish — and the same day all three found their places: the terminus enum gained `is_bedrock` (the deep-descent's is-side floor, outside the wager fraction) and a note-required `does_not_fit` confession value, and the is/ought partition itself was **published into the graph** as a claim under live attack from the hedonic-monism node — the ontology's first self-audit through its own machinery.\n\n## Experience\n\nVoice-first disambiguation is the north-star interaction, and every product move is tested against a \"capture immediately, publish deliberately\" loop — meet the user at their level, infer structure rather than demand it, derive provisional structure privately, and confirm lightly before durable graph writes. Beginner-readability is the first-screen contract: show what Deliberus thinks, what earlier material it found, what is still unclear, and the best next step. See [disambiguation ux and voice first contribution](research/disambiguation-ux-and-voice-first-contribution.md), [ux principles](ux-principles.md).\n\nThe clarification layer is shipped — claim-level clarification, concept sense memory, the extraction interpretation checkpoint, the landing preflight across text/URL/upload, `Think with Deliberus` for short authored input, and multi-turn continuation inside one non-public draft. The loop stays \"capture immediately, publish deliberately\". The live UX-governance question is how promotion and lifecycle readiness should become visible across turns without reintroducing jargon, duplication, or cognitive clutter; downstream of that sit voice-to-sense-selection continuity, stronger concept governance, and recursive decomposition beyond the first guidance layer. One design law now governs every gap-displaying surface, and it came from outside this project's own UX work: in the PACE framework a prediction error yields curiosity only after an appraisal that the gap is *closable*, and yields **anxiety-related inhibition** when it is not — so a sorry marker is never neutral disclosure, its sign depends on whether the same surface offers a move. Hence the testable rule **no gap displayed without a move offered** (the completeness oracle's `unsupported_value_premises` is the standing violation), plus a pacing caution, since curiosity tracks rate-of-understanding rather than information volume and therefore thins exactly at the depths most worth reaching. The gradient now has a **venue and a deadline**, decided 2026-08-17: two people side by side, using it live from scratch, *discussing through Deliberus* — a co-present synchronous mode the corpus had no principle for, and the only setting where the one signal that would redirect the build (do people point at claims?) is observable at all, since a handed-over graph confounds it. See [curiosity as growth fuel](research/curiosity-as-growth-fuel.md), [islands of coherence](research/islands-of-coherence.md), [execution plan clarification first roadmap](research/execution-plan-clarification-first-roadmap.md), [disambiguation ux and voice first contribution](research/disambiguation-ux-and-voice-first-contribution.md), [truth graph evidence system](research/truth-graph-evidence-system.md), [ux principles](ux-principles.md).\n\n## Outreach\n\n80,000 Hours alignment is the core positioning; the EA/rationalist world is the most natural beachhead; the differentiated pitch is persistent argument graphs plus bridging signals, not generic AI deliberation. The pitch's backbone is the missing-layer taxonomy — civilization has epistemic infrastructure for conclusions, quality control, agreement-shape, and procedure, and none for reasoning itself, so Deliberus claims an empty slot rather than competing in a class, and its self-measurement reads as \"keeps itself honest\" ([the missing layer](the-missing-layer.md)). High-overlap projects are approached through lightweight empirical collaboration on real material rather than merger language, and mission overlap can now be tested through the live system itself. See [80000 hours ai decision making](research/80000-hours-ai-decision-making.md), [competitive landscape](competitive-landscape.md).\n\nTargeting sharpened along three axes in August. The incentive analysis names the operative question — **when is legibility individually advantageous** — and its eight conditions pick out populations rather than communities: people misrepresented at a cost, those persuading a distrusting audience, parties who must coordinate across a disagreement, anyone already audited; curiosity remains the disposition that cuts across all of them. **The why-now argument gained its other half (2026-08-18)**: machine persuasion now beats expert human persuaders in preregistered experiments, the mechanism is information *throughput*, and an asymmetry made of rate is answered by checking that happens later and per claim — so the pitch is **not faster decisions, but checking that keeps pace with persuasion**. And recruitment gained an empirical axis: a sweep of the founder's fifteen-year correspondence archive found a **continuous 2011–2023 tier of people who already engaged with the idea itself**, so outreach draws from idea-engagement × recency × locality before any demographic axis — the hardest variable (does this person care?) is already answered in their own words. The matching rule's \"political disagreement\" was widened to what it proxies for — a disagreement that is **real, live, and sayable** — across five registers, with within-community governance as the lowest-exposure first register; scan for stake-holders and bridges, never party membership ([islands of coherence](research/islands-of-coherence.md) §5b-ii; named applications live in `.private/` — this page is web-served). See [incentives analysis](research/incentives-analysis.md), [session10 empirical collaboration and atomization frontier](research/session10-empirical-collaboration-and-atomization-frontier.md).\n\n## Sustainability\n\n\"Building IS the cheapest validation\" still governs the short-term. The concrete funding ladder for this project is the x-risk grant ecosystem — fiscal sponsorship via EA Sweden as the structural unlock, SFF and LTFF as the first two grants to target, Open Philanthropy and Vinnova / Horizon Europe as later-stage follow-ons. The self-financing angle remains structured analysis as a service for organizations working with complex corpora. See [sff 2026 and fiscal sponsorship path](research/sff-2026-and-fiscal-sponsorship-path.md) and [governance and financing](research/governance-and-financing.md).\n\nThe play is a parallel slate of non-dilutive grants under EA Sweden fiscal sponsorship, and a verification pass on 2026-08-14 found four of the July entries dead. **LTFF no longer exists** — EA Funds closed it and launched the Transformative AI Fund in its stead, always open, $10k–$150k typical, explicitly early-stage and failure-tolerant, which fits an application whose whole design is *the answer can come out no*. Cosmos and the $10M cooperative-AI call both closed; Foresight has suspended submissions pending a shift to focused RFPs; NLnet reopens September 3 with a November 3 deadline under three new European Open Internet Stack programmes, and is the one route needing no fiscal sponsor — with direct precedent found Aug 20: NLnet already funds argumentation-graph infrastructure (Knowledge Pixels' Nanoarguments, nanopub-based scientific claims — adjacent, not overlapping; differentiation in competitive-landscape.md). SFF's theme-round deadline passed, so that path now runs through the rolling application and its speculation channel. Two channels not previously in view are the sharpest additions, both at the renamed Open Philanthropy: a **career-transition programme whose stated unit is a person rather than a project**, which is the honest description of the situation, and a capacity-building call that names contributing to the discourse as in scope. The standing artifact-first discipline holds — the Manifund gate is two-thirds met and deliberately still shut, because a page that does not teach a funder something is worse than no page.\n\n## Where These Connect\n\nThe non-zero-sum cooperation synthesis connects them: Deliberus lowers the transaction cost of mutual understanding, the shared bottleneck across Coase, Wright, Deutsch, trust economics and cooperative AI, and the **invariant across every scale** from a bundled premise to US-China governance. The identity is **public epistemic infrastructure**, a commons carrying a falsifiable wager: legibility as the floor every session delivers, convergence as the bet the accumulated residue map tests — at once the funder statement, the differentiator, and the falsification frame. Four independent 2026 arrivals point at the same empty reasoning-layer slot, each leaving it unfilled from a different side: LawZero's machine-safety substrate, science's stalled infrastructure attempts, forecasting's ceiling analysis, and the noosphere macro-frame. The **scale ladder** organizes them, from a single claim up to civilization and time — homologous at every rung because the invariant holds, never free to transfer because each rung adds an *adversary*. Dyad and small group are load-bearing now and carry shipped instruments; the institution rung is analogized but unbuilt; the temporal rung got its **first inhabitants 2026-08-22** — dates corpus-wide, a deterministic staleness daemon, human-ratified validity lapse with challengeable grounds, filtered from strength — terminus expiry still open. The adversary rule now has a price: adversaries sort into entropy, interest and strategy, a transfer is free inside a class and a rebuild across one, and by that accounting **no instrument here has yet met an interested party, let alone a strategic one** — 25 sources, all machine- or operator-extracted. The ladder's anti-overreach gate: above small group is lineage rather than roadmap, and a rung is earned only by an instrument that can fail there. **`Opacity Is a Cost, Not a Mystery`** now opens `vision.md` *before* The Founding Wager, since convergence cannot be tested unless the descent is traversable. The architecture bet's stable form: **coherence does not emerge from accumulating locally plausible statements, it must be imposed, and the question is where** — a wiki needs an editor to notice an inconsistency, while attack edges and propagated strength make it computable. See [convergence](convergence.md), [bridging](bridging.md), [the missing layer](the-missing-layer.md), [vision](vision.md), [non zero sum economics and civilizational cooperation](research/non-zero-sum-economics-and-civilizational-cooperation.md).\n\nTwo live forks and one structural inconsistency. **The operator-reflexivity gap** is the sharpest: the institution rung is this project's own, an epistemic constitution is demanded by three docs and written nowhere — the rung where Deliberus would first be caught defecting from its own thesis. **Fork one** — whether to publicly call the residue map disagreement's aleatoric fraction: legible to forecasters, but it miscasts determinate residue as chance. **Fork two** — whether to ratify **\"curiosity is the willingness to pay it\"** as a second conviction term (Kahan supports, PACE partly inverts; not in the ratified `vision.md`). The invariant also became **measurable**: a unit-cost-against-volume claim with a formalism (Romer nonrivalry — value scales with corpus coverage, not user count). Measurement corrected the mechanism — reuse runs through the **concept layer, not the claim layer**, since cosine measures vocabulary while cross-domain kinship is structural. The typed-versus-cheap-structure question **stopped being a fork on 2026-08-14** — resolved by restructuring rather than choosing, since the five framings stack with addressability as the substrate — so what remains is measurement, and the bet is currently **failing at the relationship layer**. **That measurement gained a cheap arm on 2026-08-17**: a machine reader has no affordance to blame, so **the substrate run** (raw sources / flat claims / typed graph) tests addressability and typing separately, needs no human session and is blocked only on quota — under the strict rule that an agent-side pass may never answer a human-side null. The curve is **unmeasurable rather than unproven** (cost tracking was never implemented), and an audit found the mechanism has never been given its conditions: 16 implicit-premise claims in the graph, and decomposition never searches for an existing claim, so terminating in known subclaims is impossible by construction. And the success question now has a researched prior: externalized reasoning wins under four conditions (crisis, fallen articulation cost, institutional carriers, mapping organic practice) — three true today, the fourth ours to keep (fractal-priors-for-success.md). See [agents as a consumer class](research/agents-as-a-consumer-class.md), [lowering the cost](research/lowering-the-cost.md), [curiosity as growth fuel](research/curiosity-as-growth-fuel.md), [fractal scales and temporal frame](research/fractal-scales-and-temporal-frame.md), [red team synthesis 2026 07](research/red-team-synthesis-2026-07.md), [analysis and attunement](analysis-and-attunement.md).\n\n\n## What we're not yet sure about\n\nA short list of genuinely live open questions — the kind we'd answer differently in six months than we would today:\n\n- **Whether the convergence thesis holds all the way down.** Values may converge when decomposed far enough. Or they may bottom out in irreducible parallax. Deliberus is designed so that the system itself can make the difference visible — \"we share more than we thought\" and \"we share less than we hoped\" are both real answers, and both deserve the same graph.\n- **Whether embeddings can stay assistive without flattening meaning.** Embeddings are excellent at discovery and terrible at authority. The product has to keep them in the suggestion layer and never let them decide what a contested word means. That discipline is not automatic; it's a design choice we have to keep making.\n- **How promotion lineage from draft to canonical should actually work.** Rough deliberation material can now live in the same graph as canonical structure under a staged-maturity model. The harder question is how a newcomer-facing surface should show promotion and lifecycle readiness across turns without reintroducing ontology bookkeeping as the first thing a user sees.\n- **How the graph should represent time.** Reasoning strength is computed from the claims underneath a claim, and that computation currently has no idea what year it is. An empirical premise cannot go stale; a classified endpoint cannot expire; so the fraction of disagreement we report as irreducible is a photograph presented as a state. We know the shape of the fix — dates that count, endpoints that can lapse — and we have not yet decided how much of it belongs in the ontology rather than in the way we report the number.\n- **Whether recursive decomposition can be pushed without fracturing.** The self-similar decomposition principle says nothing in the graph is permanently atomic — including the system's own intermediate representations. But at what depth does continued decomposition stop producing insight and start producing noise? We don't know yet; we're finding out by doing it on real material.\n- **How much of this depends on the reader wanting to know.** There is good evidence that giving people a better grip on evidence makes them argue harder for what they already believed, and that the one thing which reverses that is plain curiosity — actually enjoying being surprised. Everything here is built to make understanding someone else cheaper, which changes how easily you can do it and not at all whether you want to. So we may be building something that works beautifully for the curious and hands everyone else a better weapon. We think the honest response is to measure it rather than hope, and we have not built that measurement yet.\n\n---\n\n**See also**: [Attunement](analysis-and-attunement.md) · [Lenses](worldview-lenses.md) · [Bridging](bridging.md) · [Depth](depth.md) · [Convergence](convergence.md) · [Vision](civilizational-vision.md)\n"}