Examples¶

Finished runs: one review in each mode, pages from a review article, and a gallery of lineage figures. Every figure here is unedited toolkit output. Click any figure to enlarge it.

How to read a lineage figure. Each horizontal lane is one theoretical family, labeled at left with its claim. Each dot is one verified paper, placed by publication year. Dot area is proportional to citation count, and a hollow dot has no count. Labeled dots are landmarks, chosen automatically by citation count and by how often the corpus itself cites them. A ring and a ★ mark a home-lab paper: a row from the lab's own record in lab mode, or any paper by an author named in LITREVIEW_LAB_AUTHOR. §7.2 of the manual gives the full key.

Topic mode: visual–cerebellar anatomy¶


Prompt	"anatomical connections between visual system and cerebellum, primate or human, any tractography method, back to the 1970s"
Deliverable	`visual_cerebellum_bibliography.xlsx`, 72 papers
From the search	42 papers, 1980–2025 ( cream rows)
From the cross-citation pass	30 papers, 1944–2010 ( green rows)
Caught by verification	3 fabrications: a real 2025 paper by Schmahmann et al. returned as "Olson et al.", a DOI off by one digit, and an invented PMCID
Built	May 2026, with an early version of the toolkit: before the antecedents pass and the reference gates
Time	about 7 minutes, without PDFs

This run is small and early. A current build runs more steps and takes hours; see the manual's reference card for measured times.

The cross-citation pass found the field's older anatomy (1944–2010) that the topic search missed. The output directory keeps that run's audit trail: agent_out.json (raw search results), verify_report.json (what verification caught), xref_visual_cerebellum.json (the cross-citation table), xref_picks.json (the 30 papers taken from it), and rows.json (the table the spreadsheet is built from).

Lab mode: the Gallant lab in context¶


Prompt	"review the Gallant lab's human-imaging work in the context of the broader field"
Front end	`lab_corpus.py` ingest → prune false positives → derive themes
Deliverables	`gallant_lab_in_context_bibliography.xlsx`, two lineage figures, an AI-authored review `.docx`

The Gallant lab's 61 human-imaging papers in six themes, 1990 to 2025 — **Lab mode first maps the lab's own work.** The 61 human-imaging papers from the Gallant lab, grouped into six research themes derived in Phase L3. Every paper is a lab paper, so every dot is ringed and starred. The lab's work moves from visual encoding and stimulus reconstruction (2000–2015) to semantic maps of visual cortex (from 2012) and then to language (from 2016). A methods-and-software lane runs throughout. This map defines the themes that the next step places in the wider field.

The same six themes with 287 field papers added around 72 lab papers — **The same themes, placed in the field.** Phase L4c searched outward from each theme, adding 287 papers from other groups to 72 lab papers (359 in total). Ringed, starred dots are lab papers; plain dots are the field. In visual encoding the lab's papers sit inside a dense literature that began decades earlier. In language, most of the field's papers postdate the lab's 2016 landmark. The figure shows which themes the lab entered as established literatures and which grew after its work.

A finished review article¶

Pages from the complexity_representation review .docx (Phase 7).

Review title page with author block, disclosure and abstract — **The title page states who wrote the review and how it was checked.** Below the title, the author block names the AI model. The author's disclosure explains that every citation was machine-verified and every reference rebuilt from its DOI, and that the author read abstracts, not full texts. The abstract and introduction follow.

First page of the reference list — **The reference list is generated from the verified corpus, not typed.** Each entry has its full author list, a sentence-case title and a DOI link. Entries follow APA-7 order: one author's works by year, and a sole author before that author's co-authored papers (Attneave, 1959, then Attneave & Arnoult, 1956). The one entry without a link is a DOI-less book, which keeps a hand-written reference.

Lineage figure gallery¶

Complexity representation: six families, 1948 to 2026 — **How the brain represents complexity (190 papers).** Six families, each a different answer to what complexity is: information, redundancy to compress, hierarchy, representational geometry, bounded capacity, or integrated information. The information, redundancy and capacity lanes rest on 1950s landmarks (Shannon, Attneave, Miller). Hierarchy follows in the 1960s, integration in the 1990s, and geometry in the mid-2000s. The six answers accumulated over seven decades; none replaced the others.

Cognitive map: six families, 1948 to 2026 — **The cognitive map (110 papers).** Six families, each a different account of what the map is for: a spatial metric, replayed sequences, abstract structure, prediction and planning, relational memory, or structure learned from experience. The spatial and memory lanes start with Tolman (1948), Scoville (1957) and O'Keefe (1971). The abstraction lane starts in 2008. Extending the map beyond physical space is a recent idea.

World models: six families, 1943 to 2026 — **World models (348 papers).** Families are named for what the model does: compress, infer, control, map, simulate, or emerge in trained networks. The compression lane is almost entirely pre-2005 (Attneave, Barlow, Olshausen). The simulation and emergence lanes are densest in 2023–2025. The field's recent growth is concentrated in its two newest families.

Distributed conceptual network: six families, 1950 to 2026 — **The distributed conceptual network (450 papers).** Six accounts of how concepts are stored: in sensory and motor systems, in an amodal hub, as a map tiling cortex, handed from vision to language, built from memory, or bent by goals. The sensory and hub accounts have run in parallel since the 1970s. Most home-lab papers (★) fall in the "meaning tiles the cortex" lane, from 2008 on.

Data-structure representation: six families, 1948 to 2026 — **How the brain represents data structures (117 papers).** Six strategies for holding a structure: compression, statistical learning, ordinal codes, cognitive maps, compositional symbols, or predictive maps. Each strategy has its own roots between 1948 and 1990 (Tolman, Miller, Sternberg, Smolensky). The cognitive-map lane grows fastest after 2005. Most of today's accounts of structure descend from ideas older than neuroimaging.

Comparative language network: five families, 1940 to 2026, with two editorial arrows — **The comparative language network (122 papers).** Five evolutionary arguments for how human language arose from a primate brain. The axis is linear (no time warp). The two arrows are editorial (`--spec`). Each links an older claim that a feature is unique to humans or apes to a recent paper that found it in monkeys. Every "revising the gap" paper dates from 2008 or later, so the case for gradual evolution is recent.

Conversation and brain recording: six families, 1989 to 2025 — **Recording brain activity during conversation (131 papers).** Six views of what the neuroscience of dialogue is about: brain-to-brain coupling, shared production and comprehension codes, turn timing, social context, decoding the message, and the two-brain method itself. Nearly all of it postdates 2004. The decoding lane holds only four papers, three from 2024–2025, so computational decoding of conversation has barely begun.

The HTML version is interactive

Each run also produces an interactive HTML figure. Hover a dot for its reference, click it for its summary, counts and DOI, and step through papers with Next/Prev or the arrow keys. For publication, each run also writes an SVG copy, plus PNG and PDF when rsvg-convert or Inkscape is installed. See Reading the lineage figure.