{"path":"research/kartpaus-auto-enrichment-design.md","content":"# Kartpaus Auto-Enrichment: Recursive Decomposition and Research-Feeding, Designed\n\n**Date**: 2026-08-25 · **Type**: design doc, graduated from the 2026-08-25 chat conversation on founder instruction, folding in the capture-and-scaffold ruling that ratified its direction. **Status**: direction RULED; build sequenced behind the live-test readiness suite; the founder's reflect-more-later flag applies to detection-architecture details.\n\n## The founder's ask, verbatim spine\n\nTwo statements, one arc. First: *\"In a live test for an election debate... I want Deliberus to automatically recursively decompose all claims in each mapping-break. It should find as much research of high quality as possible and feed everything into the graph.\"* Then the ruling that generalized it: *\"All unstated premises need to be clarified and scaffolded. Ideally whenever the weighing questions are applicable they should be auto opened AND tentatively responded to in the graph as faithfully as possible according to our reading of the source/sources... automatic recursive decomposition and deepening exploration and relevant research/googling should fire/engage, at the very least... directly after a live test/recording that's transcribed and diarized and ingested... and I'm leaning towards that this should also occur (to some degree/cap at least) in normal use.\"*\n\n## What this is, located in the corpus\n\nThis is the **graph-daemon vision's first concrete workload** and the **supply-side ingestion mode running live** — not a new species. Most constraints are already ruled; most machinery exists (decomposition endpoints + readiness scoring, evidence routing built-but-unadopted, methodology tiers, the injection guard, the Claude backend making call volume affordable, the shared post-store layer as the natural mounting point).\n\n## The one collision, and its resolution (the design's core)\n\n\"Decompose ALL claims recursively\" collides with what this doc called a ratified principle, *decompose along an axis only when someone contests along it* — **which was never ratified and was corrected 2026-08-27** (it misattributed a completeness criterion to the founder). The collision is milder than stated: contest-gating survives as a *budget* heuristic for where to spend machine effort, so eager decomposition into a draft tier does not violate a principle, it spends a budget. The genuine constraint is the evidence-alignment warning below (the decompose-then-verify literature: decomposition DEGRADES verification without per-sub-claim evidence alignment) and with the T-cell/pre-digestion warning. Resolution — **capture immediately, publish deliberately, applied to the machine's own work**:\n\n- **Do eagerly, show on contest.** The machine explores every claim's descent during the pause — recursive, research-fed, weighing questions auto-opened AND tentatively pre-answered from the faithful source reading — all landing at **draft lifecycle, machine provenance**, invisible on the default map. What participants see stays contest-gated: when someone attacks a claim, the descent is already there, instantly. The enzyme constraint holds: the machine changed the RATE of a move a human was about to make, never the equilibrium.\n- **Surfaces present questions, never finished analysis.** The cognitive-commons read supplies the second independent warrant (question-framed assistance is developmental; finished outputs attenuate — Danry, Everett; and the DeepMind result: question-asking facilitation clean, restatement steering). Machine pre-answers appear as \"our reading of the source suggests — confirm or correct?\", requiring the participant's own attempt (ratification, polarity answers, counterweight-naming).\n\n## Research-feeding: the adversarial-balance rule (pre-registered)\n\nAuto-fetched evidence carries a NEW flattening mechanism: selection bias — fetching research that resolves disagreement manufactures the convergence the wager tests. Counter-rule, pre-registered: **for every contested claim, fetch the strongest available evidence on BOTH sides**, methodology-tiered, always-mint-labeled (machine-fetched, asserted-by-nobody), with disagreement-preservation run over the fed material. Two shelved pieces become load-bearing the moment this ships: **evidence routing with fair-share discounting** (bulk evidence attached to decomposed trees without alignment is the literature's measured failure — systematic badge inflation in the flattering direction) and the **phantom-citation detector** (auto-fetched research is exactly where fake references arrive).\n\n## The stopping rule: worth-asking as the machine's own action price\n\nNo new constant. The machine spends its next action — one more decomposition level, one more fetch — where the expected badge movement across a display band is highest, and stops where nothing it could add would change what a reader sees. The active-inference threshold applied to the daemon's own exploration budget; this is also the founder's \"to some degree/cap at least\" for normal use, made mechanical.\n\n## Cadence and cost (live test)\n\nPause-pipelined: pause N displays while N+1's enrichment runs. A full recursive pass over a pause's claims ≈ order-100 backend calls ≈ minutes at 8-way concurrency, $6–16 API-equivalent per pause on subscription window limits — heavy but feasible for an evening. The readiness suite's per-pause ceiling asserts the latency envelope before any guest sits down.\n\n## Generalizability\n\nAlmost everything generalizes, because the live test is just the highest-cadence trigger of one loop: eager decompose + adversarial research into draft tier, surfaced by contest and hinge. Async use triggers on extraction completion; the needs-help feed consumes the same drafts; solo use runs it privately. **Post-session enrichment is RULED MANDATORY for the live test** (full premise structure of everything said/typed/implied, right after diarized ingest). **Normal-use enrichment is the founder's recorded lean, capped by the action price.** Two things do NOT generalize by default: the **personal-domain carve** (the proportionality principle — machine auto-research into claims about someone's own body/spirit waits on consent; the fallible-self-knowledge caveat governs what may be challenged: self-interpretations yes, experience reports never) and the live test's pre-registration instrumentation.\n\n## Build pieces (four, none large — sequenced AFTER the readiness suite)\n\n1. The orchestrator loop (recursive decompose + adversarial fetch, worth-asking stopping rule) — mounts on the shared post-store layer.\n2. Draft-tier provenance flag + surfacing rules for machine decompositions and pre-answers.\n3. The adversarial research fetcher (quality tiers, corpus dedup, always-mint labels, injection guard on fetched text).\n4. The kartpaus ratification surface (already the live-test plan).\n\n## Cross-references\n\n[graph-daemons-design-space.md](graph-daemons-design-space.md) (tiers, enzyme constraint, co-stimulation) · [incentives-analysis.md](incentives-analysis.md) §6b (ingest by cluster; the gates) · [cognitive-commons-and-deliberus.md](cognitive-commons-and-deliberus.md) (question-framed warrant; the ratification tether this workload must not erode) · [decomposition-axes.md](decomposition-axes.md) (the separability floor, and why contest-gating survives only as a budget heuristic; the evidence-alignment warning) · [weighing-scheme-and-terminus-classifier-design.md](weighing-scheme-and-terminus-classifier-design.md) (capture-and-scaffold addendum) · [active-inference-context-acquisition-and-deliberus.md](active-inference-context-acquisition-and-deliberus.md) (the action price) · [live-election-test-design.md](live-election-test-design.md) (the venue; the readiness gate this sequences behind)\n"}