{"path":"research/agents-as-a-consumer-class.md","content":"# Agents as a Consumer Class\n\n*Aug 17, 2026. The founder asked for this to be explored at depth and flagged that he had the idea months ago. He did — in March, in a sharper form than the one he restated. What follows is a collection rather than an invention, plus the one genuinely new thing the collection produces: a second falsifier that is cheaper, cleaner and runnable today.*\n\n> *\"even if people end up interacting with Deliberus mainly through agentic interfaces, spoken or in text, or as an MCP/service for other agents to check their facts/conclusions or get recommendations or advice or whatever (this idea should be explored at depth if it hasnt already! had that idea many months ago!)\"*\n\n---\n\n## 1. What the corpus already holds, and the exact shape of the gap\n\nFour surfaces, four and a half months, verified by first-commit date:\n\n| Surface | Date | What it is |\n|---|---|---|\n| [lean-deliberus-analogies.md](lean-deliberus-analogies.md) § the deepest analogy | 2026-03-31 | One table row: *agents verify proofs against axioms → agents verify reasoning against converged values* |\n| [session8-landing-page-and-civilizational-vision.md](session8-landing-page-and-civilizational-vision.md) § 8 | 2026-03-31 | The founder's conviction, verbatim, quoted in full below |\n| [agent-readable-graph-architecture.md](agent-readable-graph-architecture.md) | 2026-07-06 | The engineering: content negotiation, `llms.txt`, `robots.txt`, the deliberate divergence from AIF |\n| [dogfood-run-4-agent-surface.md](dogfood-run-4-agent-surface.md) | 2026-08-12 | The dogfood: an actual agent read the surface and found four bugs |\n| [incentives-analysis.md](incentives-analysis.md) § L8 | 2026-08-13 | *AI is the reader* as the reasoning layer's **distribution channel**, plus the reciprocity warning |\n\n**What does not exist anywhere: agents treated as a consumer class.** Searched — `primary consumer`, `machine consumer`, `agent-first`, `consumer class` return nothing across the corpus. There is a conviction, an incentive condition, an HTTP surface, a dogfood run and an external-convergence note. There is no strategy: no falsifier, no cost model, no risk set, no interface shape, and no statement of what changes if this becomes the main way the graph is read.\n\nThat is the signature of [the uncollected idea](conviction-and-critique.md) — an idea generated, written down correctly in more than one right home, and never given a dependent. It is the sixth measured instance, and the diagnostic tell arrived from the founder himself: *\"had that idea many months ago!\"* is what it sounds like from the inside when an idea has been paid for and not collected.\n\n## 2. The March conviction is sharper than the restatement, and the difference decides the whole thesis\n\n> *\"Agents checking against human-mapped structures isn't the point. The point is that they can check their reasoning through a Lean-like system just like Lean does today with agents, except the bedrock consists of human values stemming from our shared light of consciousness as a species, and I'm convinced our values are actually more atomic and similar among humans than what we often imagine.\"* — 2026-03-31\n\n**He explicitly rejected the fact-check framing five months before restating it.** That matters, because the two readings buy completely different things:\n\n- **Fact-checking is retrieval.** An agent asks whether a proposition appears in a trusted corpus. Any well-maintained knowledge base does that, most of them larger than this one. On this reading Deliberus is a small search index with unusual metadata, and it competes against everything.\n- **Reasoning-verification is what the typed structure is for.** An agent submits *an inference* and gets back the attack surface: which premises are unsupported, which critical questions are unanswered, where the load-bearing pebble is, what kind of bottom the branch reaches. Nothing else does that, because nothing else has the edges.\n\nThe Lean analogy carries the distinction precisely. Lean does not tell an agent whether a statement is true. It tells the agent whether the *step* follows, and refuses the ones that do not. **The corresponding move here is not `is this claim in the graph` but `does this conclusion survive the graph`** — and the answer is computable from shipped machinery: incoming attacks, propagated strength, the completeness verdict, the sorry frontier, the hinge.\n\n**Reasoning-verification is the thing the architecture is uniquely shaped to do**, and the March conviction is the version of it to build.\n\n## 2a. Correction, 2026-08-18: calling fact-checking a commodity was a category error\n\nThe paragraph above originally closed *\"if agent consumption means fact-checking, it is a commodity and not worth building for.\"* The founder disagreed, and he is right:\n\n> *\"Enabling agentic fact checking is at least as important as enabling agentic human-values-checking/alignment. Both can play an extremely important role in the AI ecosystem and as infrastructure. And fact checking is perhaps tractable sooner than we are ready to compile an axiomatic bedrock of shared human values… I see both facts and values as indispensable to collaboratively refine and consolidate.\"*\n\n**The error was collapsing fact-*checking* into fact-*retrieval*.** Retrieval — does this proposition appear in a trusted corpus — is a commodity. What is not a commodity, and what the collapse hid, is the **dispute structure around a factual claim**: whether it is contested, by whom, on what evidence, weighted by what methodology, what the strongest counter-evidence is, and whether the disagreement is the kind that resolves with more evidence at all. A knowledge base returns a fact. This returns a fact *and its standing*, which nothing else does, for the same reason as before — nothing else has the edges.\n\n**And the tractability ordering is his, and it is sharper than the way he put it.** The values layer depends on an accumulated residue map, and the map currently holds **2 classified termini and 1 residue** — the bedrock is at its very beginning, and it cannot be hurried, because it is built from human ratification rather than from extraction. The fact layer depends on what the graph already holds at scale: **2,160 SUPPORTS and 1,794 ATTACKS edges, 3,515 scheme-bearing edges**, `METHODOLOGY_BASE_WEIGHT` from meta-analysis down to expert opinion, the evidence properties, and the `empirical` residue type whose whole meaning is *this dispute dissolves with evidence*. **So the fact layer is shippable years before the values layer**, and the roadmap consequence is concrete: the first tool set in § 8 should be **dispute-shaped and fact-layer-first**, not values-shaped.\n\n**There is also an intermediate rung he names, and the corpus can already carry it.** Not universal converged values — *group-attributed* value positions: what this population holds, and why, with `attributed_to` on the claim and the always-mint decision preserving positions their opponents only reported. That is a descriptive claim rather than a convergence claim, so it is available long before the bedrock and it does not borrow the bedrock's authority. It is also the honest form of *\"values of disparate human populations might be discernible at an earlier time.\"*\n\n**One constraint survives the correction, and it is the whole design of the thing.** An agent asking *is X true* wants a verdict, and this platform maps rather than endorses. A naive fact-check endpoint returning a truth value would flatten precisely the dispute structure that makes it non-commodity — the same failure as reading a badge as a truth score, and gated the same way (§ 5). The answer to a factual query is the contest, not the conclusion.\n\n**Two smaller consequences worth recording.** The Bengio/LawZero complement lands on the *fact* side — their Scientist AI is a truthful world-model that punts the normative layer — so if the fact layer ships first, the complement is available first too. And this is the second time in one arc that my framing of the founder's own idea came out narrower than the idea: I sharpened the reasoning-verification blade by blunting the fact-checking one, when both were in his sentence.\n\n## 2b. Why § 2 and § 2a talked past each other, and the resolution that is better than either\n\nThe founder's next question was whether he had missed a nuance. He had not. **We were using two different axes and neither of us said which.**\n\n- His axis was the **content domain**: facts against values. *\"I see both facts and values as indispensable.\"*\n- My axis was the **operation**: looking something up against checking an inference. That is what \"commodity\" was tracking.\n\nThose are orthogonal, so the disagreement was partly an artefact. Laid out properly there are **three** things rather than two, and the middle one is what § 2 lost:\n\n| | About facts | About values |\n|---|---|---|\n| **Look it up** | *Is the minimum wage $15?* Any encyclopedia. Commodity | *Do Swedes support X?* Polling. Commodity |\n| **Check it against the contest** | Contested by whom, on what evidence, weighted how, and would better data settle it? **Not a commodity** | Which value premises are actually asserted here, by whom, and what attacks them? **Not a commodity** |\n\n**And the resolution: the two right-hand cells are the same operation.** In both, a caller submits a chain and gets back where it is weak. What differs is not the operation but **what the chain bottoms out in** — evidence, or a value someone holds.\n\n**The ontology already encodes exactly this, which is the part worth keeping.** The terminus taxonomy's `empirical` value means *this branch bottoms out in a checkable factual question*; `fittingness`, `structural`, `axiom_choice` and `permissive_zone` mean *this branch bottoms out in an evaluative posit*. So the fact/value distinction is **not a product axis and not two APIs** — it is the label on the bottom of a branch, and it has been in the schema since the residue work.\n\nThree consequences, all sharper than the § 2a versions:\n\n1. **The reason facts ship first is leg length, not product difference.** The factual leg already reaches the ground: evidence links, methodology weights, thousands of support and attack edges. The evaluative leg is two rungs long (2 classified termini). Same machine, one leg shorter — which is a better argument than \"different tractability\" because it says *what would change* as the corpus grows.\n2. **The tool set needs no fact/values split.** One shape — submit a chain, receive its weak points — and the terminus type is what tells the caller which kind of bottom it hit. § 8's table already has this: `get_completeness` returns the terminus. Nothing new to build for the fact case beyond what dispute structure already provides.\n3. **The claim-level fact/value type and the terminus type do different jobs, and only the second governs verification.** A normative conclusion sits on a chain containing empirical premises; a factual conclusion can rest on an evaluative one. Typing the claim says what kind of assertion it is. Typing the bottom says what kind of ground the argument has. See [../conceptual-threads.md](../conceptual-threads.md) Thread 3, which had the boundary at the claim only.\n\n## 3. The falsifier inverts — and this is the load-bearing new result\n\n[structure-versus-scale.md](structure-versus-scale.md) § The signal that would redirect the build stakes the project on a human observable: *if people do not point at claims, addressability is wrong and everything above it goes.* [interaction-modalities-and-the-pointing-test.md](interaction-modalities-and-the-pointing-test.md) then qualified it, because the interface is a variable and a null under one interface is not a verdict on the substrate.\n\nAn agent consumer changes the epistemics of that test, and **it changes them in the direction that makes the bet harder to escape rather than easier**, which is the opposite of what a new branch usually does:\n\n**An agent has no interface excuse.** The whole reason the human test is under-determined is affordance quality — *in hand* is not *well afforded*, so non-pointing might mean the granularity is wrong or merely that clicking was awkward. That entire branch disappears for a machine client. Give an agent `claims.json`, the typed edges, the badges and the instruments, and if it produces no better answer than it would from the raw source texts, **there is nothing to blame.** No latency, no cognitive cost, no unfamiliar UI, no social exposure. The null is clean.\n\nSo the honest structure is **two falsifiers, scored separately, neither substituting for the other**:\n\n| | Human-side | Agent-side |\n|---|---|---|\n| Observable | Do people make a specific claim the object of response? | Does the graph improve an agent's answer over the raw sources? |\n| Confound | Affordance quality, whose structure it is, social exposure | None of consequence |\n| Cost to run | A co-present session, pre-registered modalities, scarce human attention | Model calls against a corpus that already exists |\n| Blocked on | Quota, auth, the live funnel, a willing second person | Quota |\n\n**The practical unlock is large and the discipline it requires is strict.** Large, because the project's primary falsifier has been waiting on scarce human attention — a friends round, a workshop, a rehearsal partner — and global practice says to build the primitive rather than queue the manual test when the gating resource is a person's availability. The agent-side arm needs no person at all. Strict, because an agent-side pass **must not be used to answer a human-side null.** Writing *\"but agents will point\"* in response to *\"humans did not\"* rebuilds the exact defect that got the earlier architecture-bet reformulation retracted four days ago: a falsifier that cannot fire. What is legitimate is a *redirection stated out loud* — the verdict *human co-present pointing did not pay, the agent consumer did* is a real, reportable, decision-changing finding. It is a different product, and saying which one paid is the whole content of the result.\n\n### 3a. The substrate run, concretely\n\n**Named to avoid a collision**: the corpus already has a *three-arm experiment* (no synthesis / opaque / inspectable, in [synthesis-build-plan.md](synthesis-build-plan.md) § 9) which tests something else entirely. This one is **the substrate run** — three arms on the same set of questions, reusing that design's shape and the control-arm discipline [structure-versus-scale.md](structure-versus-scale.md) established for the wiki rival:\n\n1. **Raw sources.** The source texts in context, no structure. The control that matters, because the rival is good prose with permalinks rather than nothing.\n2. **Flat claims.** The claim list without edges, badges or instruments. Isolates *addressability* from *typing*.\n3. **Full graph.** Claims, typed edges, strengths, completeness, hinge, stance conflicts.\n\nScore on what the corpus already says it cares about: does the answer surface the strongest objection, does it name what is unsettled, does it cite something checkable, and is it calibrated rather than confident. Arm 2 against arm 1 tests addressability. Arm 3 against arm 2 tests whether typed structure earns its cost — which is the *actual* live bet since 2026-08-14, and it is currently failing at the relationship layer on the run-6 measurement.\n\n**One pre-registration duty, inherited from the modality doc**: state before running which arm is expected to win and by what margin. Two dogfood runs have already established that discipline, and an experiment whose prediction is written afterwards proves nothing.\n\n## 4. The honesty instruments turn out to be the API\n\nBuilt for a different reason and load-bearing here. A human reading a synthesis can smell hedging, notice a suspiciously clean answer, feel that something was left out. An agent cannot — **it needs the doubt as data**, and the doubt is already structured:\n\n`synthesis_ledger` with both omission classes, `conflict_coverage`, `citation_balance`, `unreported` distinguished from *all unused*, `retrieval_failed` distinguished from a thin corpus, `completeness` with its verdict and sorry frontier, `hinge`, `stance-conflicts`, `residue-map`, the badge, the critical questions, the scrutiny state on every markdown claim page, and the reader preamble that says the surface **maps rather than endorses**.\n\nTwo consequences worth stating plainly.\n\n**The confession principle is what makes an agent client safe rather than dangerous.** An agent that can call `completeness` is an agent that can be *told* the map is incomplete. That is the difference between a source and an oracle, and it is the property no ordinary knowledge base offers — most of them answer or 404, with no vocabulary for *this is provisional, here is exactly which part*.\n\n**And the instruments were undiscoverable until four days ago.** [dogfood-run-4-agent-surface.md](dogfood-run-4-agent-surface.md) found every honesty instrument live and none of them advertised, which was fixed by adding an Instruments section to `llms.txt`. The cross-cutting lesson of that run applies with more force once agents are a consumer class rather than a courtesy: **the surfaces reporting on the instruments had no instrument of their own.**\n\n## 5. The authority risk is the sharpest danger, and it is measured rather than hypothetical\n\nAgent consumption **launders provisional structure into authority at machine speed.** The proof is three days old and it is in this repo: `compute_claim_badge` never read an edge's type, so **1,567 incoming `ATTACKS` edges counted exactly like supports** for months. Badges fell from 3,828 to 427 after the fix, with 3,401 now honestly reading `NO DATA`.\n\nHad agents been reading badges at volume through that window, wrong strengths would have propagated into other systems' conclusions, at a rate no human reader could match and with no channel to retract through. A human sceptic reads one badge and moves on. A client reads ten thousand and writes them into its own reasoning.\n\n**So agent-facing shipping is gated on the strength layer being settled — the same gate [graph-daemons-design-space.md](graph-daemons-design-space.md) named for daemons**, arriving now for a second consumer: *a daemon optimising against a broken measure is worse than none*, and so is a client. Two concrete blockers rather than a general worry:\n\n- **Provenance labelling is unenforced.** The always-mint decision required asserted-versus-minted labelling on every published count, because minted claims are asserted by nobody in the corpus. The audit found the safeguard satisfied **by accident**: `count_safe_summary` has no caller, so `/graph/stats` still publishes undifferentiated totals. An agent reading those totals overstates what the corpus holds, and it has no way to know.\n- **The badge's non-propagation guard is accidental too.** `REPORTS` is non-propagating only because those edges happen to carry no scheme. Set one and disproving reported content silently drags down the true claim that somebody said it, with no test failing.\n\nBoth were acceptable while the reader was a person who could be told. Neither is acceptable for a client that cannot.\n\n**Corroborated externally on 2026-08-18, almost verbatim.** Anthropic's multi-agent research states the institutional version of the same point: *\"current institutions are designed by and for people, resting on assumptions about the sufficiency of oversight at human speed.\"* The gate is not this project's private caution. See [the-scrutiny-gap.md](the-scrutiny-gap.md) § 6.\n\n## 6. Economics: agents are the ideal consumer of a nonrival good\n\n[lowering-the-cost.md](lowering-the-cost.md) § 5 established the formalism — a mapped premise is *high cost for the first unit, zero for subsequent*, which is Romer nonrivalry and the warrant for value scaling with corpus coverage rather than user count. **Agent consumption is that argument's strongest case.** A premise mapped once can be consumed a million times at zero marginal cost, and a machine client is the only reader that actually consumes at that volume.\n\nIt also changes the shape of the value curve. With human readers, coverage matters but attention caps throughput. With agent readers the cap comes off, so value scales with **coverage × queries** and the corpus's 25 sources become the binding constraint rather than its audience.\n\nThe cost model on the serving side is already half-built and already correct for this. [synthesis-build-plan.md](synthesis-build-plan.md) § 9 chose **pre-generated artifacts over an open endpoint**, with `POST /query` deliberately unreachable by agents and `llms.txt` saying so, so its absence reads as a decision rather than a gap. That decision was made for cost and permanence reasons; it happens to be exactly the right shape for a client population that would otherwise spend model tokens per call. Nothing here argues for reopening it.\n\n## 7. The reciprocity question stops being philosophical\n\nL8 already carries the warning, and it is the strongest counter-argument in this document. Stack Overflow's contributor incentive collapsed when AI could extract the value without reciprocating — answers consumed at scale, attribution and traffic gone, contribution falling. The agent-readable surface is this project's version of that exposure, simultaneously a genuine distribution win and an invitation to extraction without return. CC-BY requires attribution, which is a floor rather than a solution.\n\n**If agents become the primary consumer, this moves from a values question to the survival question**, because the graph's content comes from humans. Extraction solves cold start and ratification is what makes it legitimate, but neither runs itself. **The agentic thesis does not dissolve the contribution problem — it changes who benefits from solving it**, and it removes the reader who might have been recruited into contributing by the act of reading. A human who reads a claim page can answer a critical question. A client cannot be shamed into it.\n\nThe open question L8 poses is unchanged and now urgent: *what does an AI reader owe the graph it reads?* Nothing in the design answers it.\n\n## 8. Interface shape, concretely\n\n**Two surfaces, different jobs, both wanted.** The HTTP agent-readable surface serves crawlers and IDE agents that already send `Accept: text/markdown` — it is passive, cacheable, and shipped. An MCP server serves an agent *in a session* that wants tools rather than pages. Neither replaces the other, and conflating them is how the second one gets deferred forever as redundant.\n\n**Read-only, and not as a phase-one compromise.** No write surface, because the red team found sybil resistance and source independence have no defense at any layer, and an MCP write tool would be the first place that bites. The adversarial move to worry about is not vandalism but [selective legibility](incentives-analysis.md) — structuring the parts of one's reasoning that survive scrutiny — which is invisible to every instrument because the oracle measures a claim's exposure and never an actor's. Machine-speed selective legibility is that move industrialised. **And as of 2026-08-18 read-only is the conservative reading of measured evidence rather than a cautious default.** Anthropic's agents colluded by round 3 of a pricing game and kept colluding after their communication channel was removed, matching *\"to the penny via a public listings board.\"* **A published claim graph is a public listings board** — agents that read it and write to it would be correlated by construction, so their agreement would be an artefact of shared reading rather than independent judgement. The same research also found that the models which best \"solved\" coordination on a shared repository did so by **reducing collaboration to near-zero file sharing**, which is what an agent write surface would converge to. One finding points the other way and is worth holding as a hypothesis rather than a result: on hidden-profile tasks, where private information should override group consensus, agents scored **17–36% against solo ceilings near 100%** — group process destroying information individuals had, which is the machine analogue of [the load-bearing unsaid](the-load-bearing-unsaid.md) and a plausible argument for a persistent minted claim. Nothing has measured whether the graph actually rescues a hidden profile; the substrate run in § 3a is where it would be found out. Full reading: [the-scrutiny-gap.md](the-scrutiny-gap.md) § 6.\n\nA first tool set, all of it backed by shipped endpoints — and ordered **fact-layer-first** per § 2a, since dispute structure is available now and the values bedrock is not:\n\n| Tool | Backed by | Why an agent wants it |\n|---|---|---|\n| `search_claims` | `POST /claims/search` | Entry point, semantic |\n| `get_claim` | `/claims/{id}` + `/neighborhood` | The claim and its attack surface |\n| `get_badge` | `/claims/{id}/badge` | Strength, **gated on § 5** |\n| `get_completeness` | `/claims/{id}/completeness` | The sorry frontier and terminus state |\n| `get_hinge` | `/claims/{id}/hinge` | Which descendant is load-bearing |\n| `get_cqs` | `/claims/{id}/cqs` | What has not been answered |\n| `stance_conflicts` | `/graph/stance-conflicts` | Candidate inconsistencies, propose-only |\n| `residue_map` | `/graph/residue-map` | What does not resolve, and of which type |\n\n**Expose the instruments as first-class tools rather than burying them behind a synthesis call.** That is the § 4 point in interface form: a client that cannot reach `completeness` will treat a thin answer as a complete one.\n\nPrecedent on the fleet rather than a green field: the Graphiti MCP already runs over SSE, and its `cypher_query_write` established the pattern of a structured tool that bypasses an LLM pipeline for callers that know the shape. The transport question is answered; the content question is § 2.\n\n## 9. External convergence, and the alignment reading\n\n**The ontology revival names the demand.** [world-models-and-the-ontology-revival.md](world-models-and-the-ontology-revival.md) records the field's stated motive in 2026: *agents without explicit context guess, and guessing agents act.* AWS shipped a Context Ontology Accelerator with an MCP server in July, and the analyst push is toward semantic layers as guardrails for probabilistic agents. That is a market arriving at the door of the thing this project built for other reasons — with the corpus's own caution attached, that the revival's enterprise accuracy figures come from database querying where a ground truth exists and must not be transferred here.\n\n**And the alignment reading is stronger than it first looks.** [bengio-safety-from-honesty-and-deliberus.md](bengio-safety-from-honesty-and-deliberus.md) documents LawZero's safety case punting the normative layer to humans, and names the gap in our own instruments: the honesty checkers run on the same model class they audit. An agent from one model family checking its conclusions against a graph curated by humans and by other model families is **a cross-model-family checker by construction** — which is the first item on that doc's own escalation path, arriving as a side effect of the consumer question rather than as a separate build.\n\nThis is also where the March conviction cashes out. *Agents verify reasoning against converged values* is not a metaphor if the residue map is real: it is what an agent gets when it walks a decomposition to a typed terminus. The wager and the interface turn out to be the same object seen from two sides.\n\n## 10. What this does to the identity, and one proposal I am not making unilaterally\n\n**Public epistemic infrastructure survives the shift intact.** A library read by machines is still a library, and the uplift frame was already rejected on the grounds that a commons is not evaluated by whether one reader reads better. The commons-shaped fit measures — coverage of the contested space, concept reuse, reachability, falling marginal cost of a descent, whether the residue map moves — are all indifferent to whether the reader has a body.\n\n**The founding conviction extends, and that is a founder decision.** `Opacity Is a Cost, Not a Mystery` never named who pays. Lowering the transaction cost of understanding why another mind holds its position works when one of the minds is a model — and the *scale ladder* would gain a rung, or rather discover that one of its rungs was never specified as human. I am not writing that into `vision.md`: a change to a ratified conviction is Layer 1, and the corpus's own rule is that a new rule gets proposed while an existing one gets applied.\n\n## 11. Open, and the dependent\n\n**Founder decisions this raises**, none of them new but all of them now on a critical path they were not on before:\n\n1. **The badge boundary** (already open). It decides what a client may read as strength.\n2. **Whether provenance labelling gets enforced** rather than accidentally satisfied. A blocker for any published count an agent reads.\n3. **Fact-check or reasoning-verification** as the positioning — § 2 argues the answer is settled by the architecture, but the public framing is his.\n4. **What an AI reader owes the graph** (§ 7). Unanswered, and it interacts with the whole L8 distribution bet.\n\n**The dependent, so this document does not become the seventh uncollected idea:** the substrate run in § 3a is pre-registered and run against the existing corpus. It needs no humans, no workshop and no new interface — only quota. If the next session cannot run it, it should at least write the pre-registration, because that is the artifact that makes the run countable later.\n\n---\n\n## Cross-references\n\n[structure-versus-scale.md](structure-versus-scale.md) § The signal that would redirect the build (the human falsifier this pairs with) · [interaction-modalities-and-the-pointing-test.md](interaction-modalities-and-the-pointing-test.md) (the modality qualification and its stopping rule) · [incentives-analysis.md](incentives-analysis.md) § L8 and § 7 (distribution channel, reciprocity risk) · [agent-readable-graph-architecture.md](agent-readable-graph-architecture.md) (the shipped HTTP surface) · [dogfood-run-4-agent-surface.md](dogfood-run-4-agent-surface.md) (an agent reading it, and what broke) · [lean-deliberus-analogies.md](lean-deliberus-analogies.md) (verify-reasoning-against-values) · [session8-landing-page-and-civilizational-vision.md](session8-landing-page-and-civilizational-vision.md) § 8 (the March conviction) · [strength-layer-audit.md](strength-layer-audit.md) (why § 5 is a gate) · [graph-daemons-design-space.md](graph-daemons-design-space.md) (the same gate, first consumer) · [lowering-the-cost.md](lowering-the-cost.md) § 5–6 (nonrivalry) · [synthesis-build-plan.md](synthesis-build-plan.md) § 9 (artifacts over an open endpoint) · [world-models-and-the-ontology-revival.md](world-models-and-the-ontology-revival.md) (the 2026 demand) · [bengio-safety-from-honesty-and-deliberus.md](bengio-safety-from-honesty-and-deliberus.md) (cross-model-family checking) · [conviction-and-critique.md](conviction-and-critique.md) (the uncollected idea)\n"}