Multi-faceted vs ingest:dual — comparison memo

ASN-689 acceptance artifact. Updated 2026-04-29 with empirical probe data from the path-B kickoff. Full-book live runs still open.

Why we run both

ingest:dual and the multi-faceted pipeline serve different consumers. They are not substitutes; this memo is the running record of what each produces on the same source.

	ingest:dual	multi-faceted
Consumer	Mosaic editorial · retrieval · audits	/sources/[code] profile · factory corpus surfaces
Output shape	Curator passages (~15) + research-bulk (~50–150)	Models / Constructs / Instruments / Factor model / Chapter summaries / Case studies (+opt-in Propositions / Supports)
Granularity	Passage chunks (~500 words)	Typed entities, per chapter
Per-book cost (sonnet 4.6, default tier, with chapter caching)	~$0.30 (embed + chunk LLMs are cheap)	$4.30–$7.54 (12-chapter book)
Per-book cost (deep tier, +propositions)	n/a	$6.46–$12.34
Fail mode if absent	Retrieval against the source returns unrelated high-similarity hits	/sources/[code] panel is empty for the source

Probe data — Maslow, A Theory of Human Motivation (60KB chunk, 2026-04-29)

Path-B validation runs on a single representative chunk.

Probe 1: entities bundle vs three singles control (untuned)

	Bundle	Singles
Models	1	1
Constructs	10	8
Instruments	0	0
Cost	$0.115	$0.252
Stop	clean	clean

54.1% savings; comparable counts; bundle's first-model fields slightly shallower than singles (e.g. key_meta_analytic_correlates: [] vs 5 entries; aggregation_rules shorter).

Probe 2: entities bundle (tuned with FIELD COMPLETENESS) vs singles

	Bundle	Singles
Models	1	1
Constructs	9	9
Instruments	0	0
Cost	$0.121	$0.256
Stop	clean	clean (1 parse error on instruments singleton)

52.9% savings; counts match; bundle's aggregation_rules is now richer than singles (captured the specific 85/70/50/40/10 satisfaction percentages). Crucial finding: singles hallucinated first_publication_year: 1943; bundle correctly returned null because the chunk contains zero mentions of 1943. The FIELD COMPLETENESS rule made the bundle more source-faithful, not just cheaper. Memory: feedback_bundled_prompts_source_fidelity.

Probe 3: argument bundle (factor_model + propositions + supports)

	Result
stop_reason	max_tokens
Output truncated	yes, mid-string at position 46274
Cost	$0.257

The bundle exceeded the 12K output-token budget. Propositions extraction is intrinsically high-output (20–50+ per chapter, each ~50 tokens). The bundle pattern doesn't fit propositions. Decision: split argument back — factor_model becomes a singleton (small output), propositions becomes a separate opt-in single.

Probe 4: narrative bundle (chapter_summary + case_studies)

	Result
stop_reason	clean (end_turn)
chapter_summary	1 (rich — logline + pitch + story_script + 12 takeaways + 7 key_questions + 10 misconceptions)
case_studies	7 (each typed and outcome-coded)
Output tokens	15,442 of 16,000 budget
Cost	$0.308

Works at the limit of the 16K budget. High-value output but expensive.

Per-bundle decision (path-B post-probe)

Bundle	Status	Cost/chapter	In default tier?
entities	✓ ship	~$0.12	yes
factor_model (singleton)	✓ ship	~$0.05	yes
narrative	✓ ship	~$0.31	yes
propositions (singleton)	✓ ship	~$0.20–$0.40	opt-in via --include-propositions

Tier model

Mike, 2026-04-29: "books we need or want to go deep on, or that we intend to do a broader guide based on (in synthesis with other hand selected books) — then we get the more expensive propositions. That until we are flush with money and can spend the extra money to just have it on hand."

• Default tier (entities + factor_model + narrative): $4.30–$7.54 per 12-chapter book. Always run on vela-tagged sources we want public source profiles for.
• Deep tier (+propositions): $6.46–$12.34 per book. Reserved for books that anchor a guide / cross-source synthesis or that we want full argument-graph treatment on. Toggle: --include-propositions.

What multi-faceted captures that ingest:dual does not

Concrete examples from the Maslow probe outputs:

1. Named theoretical models with explicit factor structure. The entities bundle returned Maslow's Hierarchy as a typed entity with primary_factors: [Physiological, Safety, Love, Esteem, Self-Actualization, Self-Transcendence, Cognitive Needs], aggregation_rules capturing the partial-satisfaction percentages (85/70/50/40/10), and a cluster_category: Motivation. Passage chunks from ingest:dual could not surface this in queryable form.
2. Constructs as cards. 9 typed constructs (Physiological Needs, Safety Needs, Self-Actualization, Self-Transcendence, Prepotency, …) each with definition + level + linked models. A query for "self-actualization" hits the construct card directly rather than passage excerpts.
3. Chapter summary as a story-shape. logline, pitch, character / problem / guide / plan / call-to-action, transformative_journey, key questions, misconceptions, takeaways. Useful for editorial reading guides built from a source.
4. Case studies typed and outcome-coded. "The Chronically and Extremely Hungry Man" / "The Psychopathic Personality" / "Martyrs and Idealists" — each with type, outcome_type, conflict, illustrates, themes, quotes.

What ingest:dual captures that multi-faceted does not

1. Long-form passage retrieval for any claim. Multi-faceted indexes typed entities; it does not replace passage-level evidence retrieval.
2. Magazine-grade pull quotes with charge tier and arc-stage tagging (curator pass).
3. Embedding-backed similarity at passage level. Multi-faceted entities are not embedded in this iteration (follow-up scope).

Cost / time per book

Filled per book after the first live full-book run. The chunk-level probe data already published above.

Book	book_key	dual-grade passages	multi-faceted entities	tier	run log	actual cost
TBD	TBD	TBD	TBD	TBD	TBD	TBD
TBD	TBD	TBD	TBD	TBD	TBD	TBD
TBD	TBD	TBD	TBD	TBD	TBD	TBD

Decision matrix

When a vela-tagged book lands, which pipelines should we run, in what order?

1. Always run ingest:dual first. Cheap; fixes retrieval; everything else depends on a queryable corpus.
2. Run default-tier multi-faceted for vela-audience books we want to feature on the source profile. Reserve for sources reached from magazine articles, museum tours, or the reading list.
3. Add --include-propositions for books that anchor a guide — broader cross-source synthesis where the argument-graph queries pay off. Mike's "until we're flush" tier.
4. Skip multi-faceted entirely for factory-internal sources unless we later want a per-source factory studio surface. The current ASN-685 panel is public-facing; factory-internal extraction adds spend without UI payoff.

Open questions

• Should multi-faceted entities also be embedded for retrieval (parallel to mosaic_passages.embedding)? Decide once ASN-690's suggestion engine for the magazine editor defines the retrieval shape.
• Aggregator + referee passes are scaffolded in the design doc but not yet wired in the orchestrator's --execute path. Add when the per-chapter loop is validated end-to-end on a real book.
• Long books (>1MB source, chapters >200KB each) will land toward the high end of the cost envelope because input-cost grows. May want a sub-chapter split strategy honoring sub-headers (per book-ingestion rule 2). Defer until a real long-book run shows the actual hit.

How to extend this memo

After every multi-faceted run:

1. Append a row to the Cost / time per book table with concrete numbers.
2. If the run surfaced something the prompt set didn't anticipate, note it in What multi-faceted captures.
3. If a bundle's actuals differ materially from the envelope, update the per-bundle envelope in lib/factory/multi-faceted-pipeline.ts and re-publish here.
4. Revise the decision matrix if a pattern emerges.

The run log lives at data/factory-outputs/research/run-logs/<book_key>-<ts>.json and is the source of truth for these numbers.