{"path":"research/the-analogy-daemon.md","content":"# The Analogy Daemon\n\n**Date**: 2026-08-30 · **Type**: founder vision (verbatim) + design mapping onto the daemon\ntaxonomy · **Status**: designed on paper here, nothing built; every design mapping below is a\nsession proposal awaiting the founder's pacing.\n\n---\n\n## 1. The founder's statement, verbatim\n\nImmediately after the fractal-priors-as-mother-claims mapping\n([convictions-across-scales.md](convictions-across-scales.md), 2026-08-30 update):\n\n> \"We could also have daemons buzzing around Deliberus in the near future whose job it is to\n> discover and evaluate analogy claims. To identify kinship across domains and debates and research\n> and reflect on fractal priors. Perhaps a first LLM pass could simply note what prior it would set\n> on the claim, and then a separate pass could investigate fractal priors and the evidence\n> for/against them etc etc. Document all this carefully, I love it, think this will be very helpful\n> and near-crucial to be able to reason about everything in the knowledge/argument graph in a way\n> that isn't autistic in some sense, too detail-focused you know, seeing the broad strokes and\n> intuiting what belongs or doesnt...\"\n\n## 2. What this is, in the corpus's own vocabulary\n\n**The graph's gestalt organ.** The graph is hyper-analytic by construction — atomic claims, typed\nedges, local checks — and the founder is naming the missing faculty: broad-strokes pattern\nrecognition, *intuiting what belongs*. In his own intellectual-lens vocabulary this is the\nright-hemisphere corrective (McGilchrist) to a left-hemisphere substrate, and it is the\n**analysis–attunement dialectic applied to the system's own cognition**: the analytic pole\ndecomposes; this daemon's job is to *see across* — kinship between debates, recurring shapes,\npriors that patterns justify. The critical design move that keeps it honest: **the intuition is\nmade inspectable by decomposition** — a daemon's \"this belongs with that\" arrives as an\nanalogy-expressing mother claim with per-instance subclaims and its counter-analogy questions open\n(the fractal-prior representation), so the broad stroke is never an oracle pronouncement, always a\nchallengeable structure.\n\n## 3. The two-pass shape, and why the separation is load-bearing\n\nThe founder's sketch has two passes, and they map onto two different daemon jobs with different\nrisk profiles:\n\n| Pass | Job | What it touches | Tier (daemon taxonomy) |\n|---|---|---|---|\n| **Prior pass** | \"note what prior it would set on the claim\" — a proposed base score with stated grounds | the strength layer's inputs | **propose-only, never direct** — a wrong prior is a house-choice multiplier: determinism repeats it on every claim it touches ([the-residual-error-taxonomy.md](the-residual-error-taxonomy.md) class 4) |\n| **Fractal-prior pass** | investigate the pattern behind the prior — assemble instance subclaims across scales/domains, find counter-lenses, run the four-check transfer test | mints analogy mother claims + instance claims + attack edges | **propose-only minting** — the roaming-philosopher tier (see § 4) |\n\nThe separation is not workflow decoration — it is **segregation of duties inside the daemon**: the\npass that *proposes a number* and the pass that *audits the pattern behind the number* must be\nseparate calls (ideally separate model families, per the two-family jury), or the prior-setter\ngrades its own homework. The four-check transfer test (Recurrence · Mechanism identity · Adversary\n· Substrate condition) is the second pass's checklist, and it is already ruled as the cross-scale\nenrichment of the shipped analogy scheme's critical questions — so the daemon's output slots into\nexisting ontology with **nothing new to build at the schema layer**.\n\n## 4. This is the first sanctioned roaming philosopher — name the crossing\n\nThe daemon design space ([graph-daemons-design-space.md](graph-daemons-design-space.md))\ndeliberately quarantined *\"roaming philosophers\"* from the six defensive daemons: **asking what\nnobody asked changes the equilibrium**, and the enzyme constraint says a daemon may only accelerate\na change a human would have made anyway. The analogy daemon crosses that line **on purpose and with\nthe founder's explicit commission** — discovering kinship nobody asked about is its whole job. The\ncontainment that makes the crossing safe is the one the taxonomy already prescribes for\nequilibrium-touching work:\n\n- **Propose-only, always.** Every analogy claim, instance subclaim, prior suggestion and attack\n  edge enters as machine-proposed structure with non-human provenance, ratifiable and rejectable.\n  Nothing it produces propagates strength until ratified (and the standing-of-untouched-machine-\n  structure question — founder-open — governs what its unratified output may ever mean).\n- **Queue, never ping.** Its discoveries accumulate as a browsable proposal queue; interrupts stay\n  reserved for the taxonomy's named cases.\n- **Pause on rejection-rate.** The taxonomy's own dead-man switch: a rising proposal-rejection rate\n  pauses the daemon rather than letting it flood the queue with plausible-wrong kinship.\n- **Preconditions unchanged**: the proposal layer, non-human provenance, an undo path, a spend\n  gate, and the strength layer settled first — a kinship daemon optimising against a broken\n  measure would manufacture exactly the false sameness the concept layer exists to prevent.\n\n## 5. Why this answers a measured gap rather than adding a feature\n\nThree findings this daemon is the response to, all measured:\n\n1. **Cross-domain kinship is structural and cosine cannot see it.** The Israel/Palestine run added\n   8 similarity edges, all within its own pair, none to the other sources — vocabulary similarity\n   misses structural kinship by construction. The daemon asserts kinship as a challengeable claim\n   instead of hoping a metric detects it.\n2. **The corpus's own fractal priors are uncollected** *(⚠ storing THEM is held — the founder's 2026-08-30 hold on self-hosted priors and Deliberus-beliefs-in-Deliberus applies here, pending the open reviews; the daemon over user content is not what the hold names).* Seventeen success-prior lenses and the\n   whole convictions-across-scales sweep live as research prose; not one is in the graph. The\n   daemon's fractal-prior pass is the mechanism that would keep such patterns *collected* —\n   including running them against new claims as they arrive.\n3. **Assembly theory's own criticism, answered**: the assembly index \"only takes into\n   consideration identical copies rather than the full spectrum of causal operations\" — which is\n   run 6's finding that reuse lookup must match structurally, never by identity. An analogy daemon\n   IS the structural-match lookup, running continuously.\n\nAnd one warning inherited at full strength: **an analogy is the highest-leverage vehicle for\nplausible-wrong** (a sloppy analogy passing as inference is on the founder's own slop list, with\nthe scheme's critical questions as the deletion mechanism). A daemon that mints analogies at scale\nmints the failure mode at scale too — which is why the counter-analogy question (*is there a\nsimilar case where the claimed property fails?*) is not optional output but part of the minting\ncontract: **no analogy claim without its strongest counter-candidate attached.**\n\n## 6. Relation to the harness conviction\n\nThis daemon is the harness conviction pointed at the graph itself: the same move that makes a\nmachine reader steely (ground every judgment in an addressable, checkable record) applied to the\nsystem's own intuition. \"Seeing the broad strokes\" without the harness is vibes; with it, every\nbroad stroke decomposes into instances anyone can attack. The founder's phrase *near-crucial to be\nable to reason about everything in the graph* places it on the roadmap between the strength-layer\nhonesty work (its precondition) and the agent-readable surface (its consumer: an agent asking\n*what is this claim's prior, and what pattern justifies it?* gets an inspectable answer).\n\n---\n\nCross-references: [graph-daemons-design-space.md](graph-daemons-design-space.md) (the taxonomy this\nextends; the roaming-philosopher quarantine this deliberately crosses) ·\n[convictions-across-scales.md](convictions-across-scales.md) (the four-check test as analogy-scheme\nCQ enrichment; the mother-claim representation) ·\n[the-harness-conviction.md](the-harness-conviction.md) ·\n[fractal-priors-for-success.md](fractal-priors-for-success.md) (the lens inventory a first run\nwould store) · [the-residual-error-taxonomy.md](the-residual-error-taxonomy.md) (house-choice\nmultipliers; same-hand bias) · [scheme-set-exhaustiveness.md](scheme-set-exhaustiveness.md) (the\nanalogy scheme and its siblings)\n\n**Semantics note (2026-08-31, from the IRIT harvest)**: the daemon's evaluate-analogies pass has a ready-made principled semantics line — Amgoud's analogical-argument work (Choquet-integral aggregation for interacting features; the 2026 premise-sensitive semantics where attacker *dispersion across premises* matters; the four philosopher considerations Relevance/Number/Diversity/Disanalogy, which map closely onto the four-check transfer test). Per-scheme aggregation is licensed by the same school: analogy edges may deserve counting-flavored semantics while other schemes keep quality-precedence. Detail: [the-irit-harvest.md](the-irit-harvest.md) § 3.\n\n**Lineage + two proposals (2026-09-02, from a founder-shared thread)**: this design is Hofstadter's Copycat re-described — codelets in scout/tester/builder roles are the two-pass shape, the Slipnet is the concept layer, and Copycat's **temperature** (the system's own settledness measure, governing where it explores) is the transferable piece the daemon lacks: a region-level temperature to roam by, proposed unruled. Also proposed: a human-seeded channel (a person's *\"this reminds me of X\"* is the idiosyncratic juxtaposition; the daemon develops and tests it). The post's own governing analogy, run through the four checks, comes out Conditional. Full read: [perceiving-possibility-and-the-analogy-organ.md](perceiving-possibility-and-the-analogy-organ.md).\n\n## 7. The human-seeded channel — founder-endorsed 2026-09-02 (*\"Very useful, document this idea.\"*)\n\n**The idea in one line**: a person supplies the idiosyncratic juxtaposition — *\"this reminds me of\nX\"* — and the daemon does the expensive part: develops it into an analogy mother claim with\nper-instance subclaims, finds the strongest counter-analogy, and runs the four-check transfer test\n(Recurrence · Mechanism identity · Adversary · Substrate condition). Perceiving possibility stays\nwith the positioned human; the harness does the checking.\n\n**Why it is a design element and not a courtesy.** Rama Ganesan's objection on the Herrán thread\n(*\"even though AI has the vast library of online data available to it, it does not make those\nunexpected connections\"*) is the creativity-shaped form of the derivability finding\n([derivability-of-missing-considerations.md](derivability-of-missing-considerations.md)): the\nmachine reliably reaches the well-articulated space — famous analogies, textbook cross-domain\nkinship — and cannot reach the *idiosyncratic* one, which lives in one person's particular library\nread in one particular order. That is the corpus-absent rung. A daemon that only self-seeds\ntherefore mints the analogies everyone already knows; the channel is how the rare, positioned\njuxtaposition enters at all. It is also the cheapest possible human act — one sentence, no ontology\nfluency — which places it on the lowest rung of the engagement gradient, exactly where the\ncorpus says valuable acts should live.\n\n**The contract.** A seed is one sentence plus the two claim or concept ids it joins. The daemon\nreturns, as propose-only structure with the seeder attributed: the mother claim (\"X and Y share\nmechanism M\"), instance subclaims, the strongest counter-analogy attached (the minting contract of\n§ 5), and the four-check verdict (Raised / Neutral / Lowered / Conditional). Nothing propagates\nstrength until it survives attack. A seed the daemon cannot develop is returned as *could not\ndevelop*, never silently dropped — the confession channel.\n\n**The instrument it creates.** Rama's claim becomes testable rather than debatable: compare\n**human-seeded** against **daemon-seeded** analogy claims on survival under attack and on the\nfour-check verdict distribution. If human seeds survive at a higher rate, the idiosyncratic channel\nis doing real work and the daemon's self-seeding should be throttled toward development; if they do\nnot, the machine reaches further than the objection allows. Either outcome teaches something —\nHerrán's own experimental-contrast method applied to creativity — and it is the clean falsifier\nshape, since a machine seed has no affordance to blame.\n\n**Preconditions**: unchanged from § 4 (propose-only, queue never ping, rejection-rate pause, the\nstrength layer settled first). One addition: seeds are attributed and the seeder gets the consent\nchannel over the daemon's rendering of *their* juxtaposition — a person may say \"that is not what\nI meant by the resemblance\", which is act three, and it is the one place the daemon's output can\nbe corrected by exactly one person.\n\n"}