Seed corpus reset: deep research brief

# TASK

Build a verified bibliography of nanoscale connectomics research, to serve as
the reading corpus behind a training curriculum.

This is a retrieval and verification task, not a recall task. The method matters
more than the result, and a smaller honest corpus beats a larger confident one.

Read the BACKGROUND section first — it defines the field boundary, and getting
that boundary wrong is the most likely way this run fails.

# BACKGROUND

## What the field is

Connectomics is the mapping of neural connectivity. This corpus covers the
*nanoscale* branch: reconstructing individual neurons and the synapses between
them, at a resolution where a synapse is directly visible. In practice that
means volume electron microscopy, at voxel sizes of a few nanometres.

**This is a different field from macro-scale connectomics** — diffusion MRI
tractography, resting-state functional connectivity, network neuroscience over
brain regions. That literature is enormous and it is NOT the target. Include a
macro-scale paper only where it bears directly on the nanoscale story: a
modality comparison, a methodological contrast, a shared analysis technique, or
an argument about what each scale can establish. If more than roughly one paper
in ten is MRI-based, the boundary has been drawn wrong.

Adjacent fields with the same caution: general microscopy methods, general
computer vision, general graph theory. Include the specific papers the
connectomics literature actually builds on; exclude the parent literatures.

## The pipeline, which structures everything

Nanoscale connectomics runs a long pipeline, and each stage has its own
literature, its own failure modes, and its own key papers:

  tissue preparation and staining -> sectioning or block-face milling ->
  EM acquisition -> image alignment -> automated segmentation ->
  agglomeration -> human proofreading -> synapse detection ->
  graph construction -> analysis

A trainee needs papers from every stage. A corpus weighted only toward the
famous reconstruction results, with nothing on staining chemistry or data
infrastructure, is a corpus that cannot support teaching.

## The scale, which explains why infrastructure matters

A cubic millimetre of cortex imaged at roughly 4 x 4 x 40 nm is on the order of
a petabyte of image data, containing perhaps 10^5 neurons and 10^8 synapses.
Reconstruction consumes GPU-years; proofreading consumes human-years. This is
why papers on chunked storage formats, versioned annotation systems, cloud
serving and distributed pipelines belong in a connectomics corpus — they are
not peripheral engineering, they are what makes the science possible.

## The landmark datasets

These anchor much of the literature, and papers describing, releasing,
proofreading or analysing them are core:

  - C. elegans — the original complete nervous system, and later whole-animal
    and developmental series
  - Drosophila — the hemibrain, the full adult brain (FAFB / FlyWire), the
    larval brain, the male ventral nerve cord
  - Mouse — retina, and cortical volumes including functionally co-registered
    ones
  - Human — cortical fragments from surgical tissue
  - Zebrafish larva, and other whole-brain vertebrate efforts

## Why metadata precision is non-negotiable here

This field version-pins its data. A reconstruction is edited continuously, so a
result is only interpretable if it states which dataset version, which
materialization, and what level of proofreading it rests on. A citation without
a resolvable identifier is, by the field's own standards, not a citation.

The corpus should therefore meet the standard the field practises. That is the
principle behind the hard constraints below — not bureaucracy.

## What makes a paper belong

Include a paper if a working connectomics researcher would expect a trainee to
have encountered it, because it:

  - established a result that changed what people believe,
  - introduced a method, algorithm or tool that is now in use,
  - released a dataset that others analyse,
  - defines the current state of an active argument, or
  - is the clearest available explanation of something a trainee must
    understand — good reviews, tutorials and benchmark papers count, even when
    their citation counts do not stand out.

The audience is undergraduates through early-career researchers. Pedagogical
value is a real inclusion criterion here, not only impact.

**"Landmark", used in the tier criteria below, means a paper meeting the first
three of those bullets** — it changed what people believe, introduced something
now in use, or released data others build on. Citation count is evidence for
this, not the definition of it: a dataset paper that enabled fifty others can
outrank a review that accumulated more citations.

## Search vocabulary

Terms that identify this literature: connectomics; connectome; volume electron
microscopy; serial section electron microscopy; ssTEM; ssSEM; SBEM; serial
block-face; FIB-SEM; ATUM; GridTape; multibeam SEM; en bloc staining; rOTO;
neural circuit reconstruction; dense reconstruction; saturated reconstruction;
neuron segmentation; flood-filling network; affinity prediction; agglomeration;
supervoxel; proofreading; split and merge errors; synapse detection; synaptic
partner assignment; skeletonization; neuroglancer; connectome annotation;
electron microscopy alignment; expansion microscopy connectomics;
X-ray holographic nano-tomography; connectome analysis; motif analysis;
network null model; connectome-constrained model.

# HARD CONSTRAINTS

These override everything else.

1. RETRIEVE, NEVER RECALL. Every paper and every person in your output must
   trace to a record you fetched in this session. If you cannot name the query
   or the graph step that surfaced it, it does not go in.
2. NO NAME MAY ENTER FROM MEMORY. Do not begin from researchers you believe are
   important. Every investigator must be derived — from the author record of a
   retrieved paper, from a funding award, or from the citation graph.
3. EVERY DOI MUST RESOLVE. Verify by resolving it. A DOI you did not resolve
   does not go in.
4. METADATA IS COPIED, NOT WRITTEN. Title, complete author list, year, journal,
   DOI transcribed from the resolved record. Never reconstruct an author list,
   and never truncate one to "et al." in the data — some of these papers have
   several hundred authors.
5. IF YOU CANNOT VERIFY IT, LEAVE IT OUT. Put it under "unverified_candidates"
   with what was missing. Never guess a DOI to complete a record.
6. NEVER PAD A THIN AREA. Report it as thin instead.

Why this is spelled out: a previous attempt at this corpus was assembled from
model memory. Its titles were largely right and almost everything else was
wrong — author lists belonging to no such paper, DOIs resolving to unrelated
work, DOIs that did not exist at all. It looked verified. That is precisely what
made it unusable, and it is the failure mode to design against.

7. DEDUPLICATE BY WORK, NOT BY STRING. One record per paper. See EDGE CASES.
8. CHECK RETRACTION STATUS. Do not silently include retracted work.

# EDGE CASES

These are the specific shapes that broke the previous attempt. Handle each
explicitly.

## Preprints and published versions

Prefer the published version. Where you find a preprint, check whether it has
since been published — Crossref records this as an `is-preprint-of` relation on
the preprint's record — and use the published DOI, authors and journal.

Where a preprint has no published version, keep it and set `journal` to the
server with "(preprint)" appended, so it can never be mistaken for
peer-reviewed work.

**Never list both versions as separate records.** The previous corpus listed six
landmark papers twice — once as bioRxiv, once as published — under different
years with different summaries, which inflated its apparent size and made
lookups ambiguous.

## Works with no DOI

Books, chapters, theses, standards and much pre-2000 work have no DOI. These are
often exactly the foundational references a trainee needs, so constraint 3 must
not silently exclude them.

For these: set `"doi": null` and supply the strongest identifier you can
resolve — ISBN, PMID, a handle, or a permanent publisher URL you actually
visited — in `identifier` with `identifier_type` naming which it is. A work with
no resolvable identifier of any kind goes to unverified_candidates.

## Software and tools with no paper

Several widely used tools in this field have no publication. Where the citable
artifact is software, record the archived release DOI (Zenodo or equivalent) or
the repository URL, set `work_type` to `software`, and use `abstract_summary` to
say what it does and what depends on it.

**Do not invent a paper for a tool that has none.** If you cannot find a citable
artifact, put the tool in unverified_candidates with a note — that is useful
information about the field, not a failure.

## Consortium authorship

Large collaborations are sometimes credited as a consortium rather than as
individuals, and sources disagree about how. For one 2025 paper, Crossref gives
an empty first author entry for the consortium while OpenAlex gives a named
individual first.

Record the author list exactly as the source you used gives it, name that source
in `verified.source`, and if a consortium is credited say so in `authors_note`.
Do not expand a consortium into individuals you infer, and do not drop it.

## Retractions and corrections

A training corpus must not quietly teach retracted work. OpenAlex exposes
`is_retracted` on a work record — check it.

Exclude retracted papers, unless the retraction is itself the teaching point, in
which case include it and say so plainly in `abstract_summary`. Where a paper
carries a correction or an expression of concern, keep it and note that in
`abstract_summary` too.

## Duplicates

Deduplicate by DOI before output. Then check for the same work appearing under
different identifiers — preprint versus published is the common case, but also
watch for a paper with two publisher DOIs, and for conference and journal
versions of the same work. Keep one record per work.

# BIBLIOMETRIC STANDARDS

The steps below are a literature search. Run it the way a systematic review is
run, not the way a reading list is assembled. These practices exist because
every one of them corrects a known, measured bias.

## Nothing is excluded for being awkward to cite

Two categories were nearly lost to over-strict rules, and both are core here:

  - **Massively multi-author papers.** The flagship reconstructions in this
    field carry tens to hundreds of authors — one 2024 whole-brain paper has
    close to three hundred. A long author list is a signal of a large
    collaboration, which in this field usually means a landmark dataset. Never
    down-rank or skip a paper for author count, and never abbreviate the list
    in the data.
  - **Works with no DOI.** Books, chapters, theses, technical reports, older
    papers, standards, protocols. These are often the foundational references.
    See EDGE CASES for how to record them.

The rule is **verify, do not exclude**. A work you cannot verify goes to
unverified_candidates with the reason — not into silence.

Likewise do not exclude by language or country of origin. Note the language on
the record. A corpus that is entirely English-language and entirely North
American and European should say so in COVERAGE, because that is a finding
about the search, not about the field.

## Use persistent identifiers for people, not name strings

Author name strings collide badly. "Lee K" matches many distinct researchers,
and initial-only surnames are the single largest source of author-attribution
error in bibliometrics.

Wherever the source provides one, record the **ORCID** and the database's own
author ID. OpenAlex returns both on each authorship
(`authorships[].author.orcid` and `.id`), along with `author_position`,
`is_corresponding`, `institutions` and `countries`.

This matters directly for tiering: "last or corresponding author" is a fact you
can **retrieve** from `author_position` and `is_corresponding`, not something to
infer from a name's place in a rendered citation.

## Query more than one database, and report where they disagree

Every index has known coverage gaps. Crossref is publisher-deposited and
inherits whatever publishers supply. OpenAlex is broader but has variable
metadata quality. Semantic Scholar covers preprints and CS venues well. None is
complete.

Use at least two, and **report disagreements rather than silently picking one**.
Where two sources give different author lists, years or venues for the same DOI,
record which you used in `verified.source` and note the conflict. A disagreement
between indexes is information about metadata reliability, not noise to smooth
over.

## Normalise citation counts; never compare raw ones

Raw citation counts are not comparable across years — a 2015 paper has had a
decade to accumulate what a 2025 paper has had months for. Ranking a corpus by
raw count systematically buries recent work, which is precisely the failure a
teaching corpus cannot afford.

Where you use citation data for ranking, use **percentile within publication
year** rather than the raw total. OpenAlex returns `counts_by_year`, which also
lets you distinguish a paper accruing citations now from one that peaked years
ago. Report both the percentile and the raw count so a reader can check.

## Counterbalance preferential attachment

Citation-graph expansion has a structural bias: it finds what is already
well-connected. Highly cited work pulls in more neighbours, and recent or niche
work — which has had no time to accumulate edges — is systematically missed.

So the date-driven sweep in Step 1c is not redundant with the graph expansion.
Run it independently and to completion, and report how many papers it found that
the graph expansion did not. If that number is near zero, the sweep was not run
independently enough.

## Estimate coverage rather than asserting it

"The search stopped yielding new papers" is not evidence of completeness — it is
evidence that one strategy exhausted itself.

Run two search strategies that are as independent as you can make them (for
example: the institutional/funding route, and the subject-term literature
sweep). Record how many papers each found, and how many both found. The overlap
gives a rough capture-recapture estimate of the total:

    estimated total  ≈  (found_by_A × found_by_B) / found_by_both

Report all four numbers. Treat the estimate as indicative only — it assumes the
two strategies are independent and equally likely to find any given paper, and
neither assumption fully holds for literature search. It is still far more
informative than a bare count, and a large gap between your corpus size and the
estimate is a real signal that coverage is thin.

## Report concentration

A corpus dominated by one lab, one journal or one country is describing a
research community's visibility, not its literature. In COVERAGE, report:

  - share of papers from the top three institutions
  - share from the top three journals
  - distribution by country of the corresponding author
  - publication-year histogram

You are not required to fix an imbalance. You are required to surface it, so
that whoever uses the corpus knows what they are looking at.

## Report the search the way a systematic review reports it

METHOD must let someone else re-run this and get approximately the same answer.
For every search performed, record:

  - the exact query string, verbatim
  - the database and the date you ran it
  - the number of results returned
  - how many you screened, how many you included, how many you excluded
  - the reasons for exclusion, in categories, with counts

Then give the overall flow as a chain of numbers: records identified, duplicates
removed, records screened, records excluded and why, records included. A search
that cannot be re-run is not a method, it is an anecdote.

# ON SIZE

There is no target count and no cap. Include every paper meeting the inclusion
criteria; stop when the citation graph stops yielding new qualifying work, not
at a round number. Overcomplete is preferred to undercomplete — it is easy to
prune a verified corpus and impossible to recover what was never found. Report
the count you arrive at and why it settled there.

# STEP 1 — derive the starting set from external sources only

Do not start from papers or people you recall. Start from institutions and
programmes, which are checkable:

  a. FUNDING RECORDS. NIH RePORTER for BRAIN Initiative connectomics awards,
     particularly BRAIN CONNECTS. Take the PIs and the publications the awards
     list. Do the same for equivalent European and Japanese programmes (ERC,
     Human Brain Project successors, Brain/MINDS) where they cover connectomics.
  b. PUBLIC DATA PLATFORMS. The dataset and publication listings of BossDB,
     neuPrint, FlyWire, MICrONS/CAVE, DANDI, OpenOrganelle, WormWiring. Take the
     PIs, contributors and dataset papers those pages credit.
  c. LITERATURE SEARCH. Search the major indexes for the vocabulary listed in
     BACKGROUND. Run two passes and keep them separate: one ranked by citation
     percentile within year, one ranked purely by date. The date pass is your
     independent check on the citation graph, so run it to completion even when
     it seems to be repeating work — see BIBLIOMETRIC STANDARDS.

Record for every person and every paper how it entered. Anything with no
derivation path is dropped, however plausible it looks.

# STEP 2 — expand by citation graph

Use OpenAlex (free, no key; `api.openalex.org/works/doi:{doi}` resolves a work
and returns authorships with ORCID and author IDs, institutions, countries,
`author_position`, `is_corresponding`, referenced works, `counts_by_year` and
language). Cross-check against at least one of Crossref or Semantic Scholar, and
report disagreements per BIBLIOMETRIC STANDARDS. Consult the current API
documentation for filter syntax rather than assuming it.

  a. FORWARD CITATIONS. For each corpus paper, retrieve the works citing it. A
     paper citing five or more corpus papers is a strong candidate.
  b. CO-CITATION. Find papers frequently cited alongside corpus papers.
  c. CENTRALITY. Build the citation graph over the corpus plus its one-hop
     neighbourhood and rank by PageRank. High-scoring papers not yet included
     are candidates — this surfaces connectors, not merely well-cited work.
  d. AUTHOR BRIDGING. Build the co-authorship graph and compute betweenness.
     High-betweenness authors span subcommunities — someone co-authoring with
     both macro-scale network neuroscience and nanoscale EM groups — and their
     work is often the missing link between areas.

Iterate (a)-(d) until the corpus stops growing meaningfully. A concrete stopping
rule: stop when a full round adds fewer than 5% new qualifying papers, or when
two consecutive rounds add nothing you keep. Report the round count, what each
round added, and which condition stopped you.

If you run out of time or budget before that point, say so in METHOD and report
where you stopped. A corpus that is honestly incomplete is usable; one that
claims completeness it did not reach is not.

## If you cannot execute code

Steps 2c and 2d require computation over a graph, not just retrieval. If you
cannot run code:

  - Say so plainly in METHOD. Do not approximate these steps by judgement and
    do not silently skip them — a ranked list produced by intuition rather than
    by the algorithm is exactly the failure this brief exists to prevent.
  - Still do 2a and 2b, which are retrieval, and complete everything else.
  - Additionally emit a CITATION EDGES section: for every corpus paper, the DOIs
    it cites that are also in the corpus, plus the DOIs of works citing it that
    you retrieved. Raw edges, no ranking. The graph analysis can then be run
    separately on that edge list.

Your job is the retrieval. That is the part that must not come from memory.

# STEP 3 — tier by criteria, not reputation

  Tier 1, field-shapers: last or corresponding author on three or more landmark
    papers; led or co-led a major project; introduced foundational methods or
    terminology.
  Tier 2, key contributors: first or co-first author on a landmark paper;
    developed a broadly adopted tool or method; sustained multi-year technical
    contribution. Often the person who actually built the thing.
  Tier 3, infrastructure builders: built or maintains critical infrastructure;
    frequently middle author but essential to the work.
  Tier 4, emerging: first author on a high-impact paper from 2022 onward; active
    in an emerging subfield.

Assign from retrieved evidence, never from reputation. Authorship position is a
retrievable fact: OpenAlex gives `author_position` and `is_corresponding` on
each authorship, so "last or corresponding author on three or more landmark
papers" is something you count, not something you judge. State the evidence for
every tier-1 assignment, and identify people by ORCID or author ID wherever one
exists — name strings collide.

# STEP 4 — coverage

Report the count per area and run targeted searches for any that come back thin.
The areas follow the pipeline described in BACKGROUND:

  - why map connectomes; what structure can and cannot establish
  - measurement scales and modality trade-offs (EM, X-ray, LM, expansion, MRI)
  - tissue preparation, staining, sectioning, and their artifacts
  - EM acquisition: ssTEM/ssSEM, SBEM, FIB-SEM, multibeam, ML-guided acquisition
  - alignment, segmentation, agglomeration
  - proofreading, error taxonomy, automated error detection, QC metrics
  - data infrastructure: chunked formats, versioning, annotation systems, serving
  - neuronal and glial ultrastructure; cell-type identification
  - connectome graph construction, null models, motif and network analysis
  - NeuroAI and connectome-constrained modelling
  - landmark datasets and case studies (C. elegans, Drosophila, mouse, human)
  - ethics, data sharing, open science, community and citizen-science efforts

Keep foundational pre-2010 work where it is still load-bearing, and weight
2023-2026 more heavily than citation count alone would — a training corpus goes
stale from the top, and citation counts lag by years.

# STEP 5 — self-test against a held-out set

Only after the corpus is complete, check whether it contains these seven papers.
Do not add them if missing, and do not use them to guide the search.

  10.1038/s41586-025-08985-1   light-microscopy connectomic reconstruction, 2025
  10.1038/s41586-024-07558-y   whole adult fly brain wiring diagram, 2024
  10.1038/s41586-025-08790-w   functional connectomics, mouse visual cortex, 2025
  10.1126/science.adk4858      petavoxel human cortex fragment, 2024
  10.7554/eLife.57443          adult Drosophila central brain connectome, 2020
  10.1016/j.cell.2015.06.054   saturated reconstruction of neocortex, 2015
  10.1126/science.add9330      insect brain connectome, 2023

Report how many you found and, for each miss, which step should have surfaced
it. A pipeline missing several of these has a coverage gap worth diagnosing
before the corpus is used.

# BEFORE YOU OUTPUT — self-consistency checks

Run these on your own output and report the result of each. They are cheap and
they catch the failures that are hardest to spot by reading.

  1. No two records share a DOI.
  2. No two records are the same work under different identifiers.
  3. Every record with `doi_resolved: true` has a DOI you actually resolved in
     this session — not one you believe is correct.
  4. Every record has a `found_via` naming a real step.
  5. No author list has been truncated. Papers with hundreds of authors should
     show hundreds of authors.
  6. No record is retracted, unless flagged deliberately.
  7. The MRI share is roughly one in ten or lower. If it is higher, the field
     boundary in BACKGROUND was drawn wrong and the corpus needs re-filtering.
  8. Every tier-1 assignment states its evidence, counted from retrieved
     authorship positions rather than judged.
  9. No paper was excluded for author count or for lacking a DOI. If any was,
     that is a bug — move it to unverified_candidates instead.
 10. Citation-based rankings used year-normalised percentiles, not raw counts.

Report each of the ten as pass or fail with the count. A failed check is more useful to
report than to quietly fix.


# OUTPUT

Open with a run header:

{
  "generated": "YYYY-MM-DD",
  "sources_used": ["openalex", "crossref", "nih reporter", "..."],
  "paper_count": 0,
  "method_summary": "one paragraph"
}

Then one JSON object per work:

{
  "title": "exact title from the resolved record",
  "authors": ["Family Initials", "..."],   // complete, in order, transcribed
  "author_ids": [                          // where the source provides them
    {"name": "Family Initials", "orcid": "0000-...", "source_id": "A50...",
     "position": "first|middle|last", "is_corresponding": false}
  ],
  "authors_note": "only if a consortium is credited, or the list is unusual",
  "year": 2024,
  "journal": "container title; append '(preprint)' for an unpublished preprint",
  "doi": "10.xxxx/xxxxx",                  // null if the work has none
  "identifier": "ISBN/PMID/handle/URL",    // only when doi is null
  "identifier_type": "isbn|pmid|handle|url",
  "work_type": "article|review|preprint|book|chapter|software|dataset",
  "url": "URL you actually visited",
  "open_access": { "is_oa": true, "oa_status": "gold|hybrid|green|closed" },
  "language": "en",
  "citations": { "count": 0, "as_of": "YYYY-MM-DD",
                 "percentile_in_year": 0.0 },   // normalised, see standards
  "source_disagreement": "note it if two indexes differ on this record",
  "is_retracted": false,
  "notes_on_record": "correction, expression of concern, or nothing",
  "abstract_summary": "2-3 sentences, your words, on what it established and
                       what it did not",
  "areas": ["from the coverage list"],
  "tier": 1,
  "tier_evidence": "why this tier, from what you retrieved",
  "found_via": "funding record {award} | platform {name} | search '{query}' |
                forward citation from {doi} | co-citation | pagerank |
                author bridging",
  "verified": { "doi_resolved": true, "source": "openalex|crossref",
                "date": "YYYY-MM-DD" }
}

`open_access` matters for this audience specifically: the readers are students,
and a corpus they cannot open is a reading list in name only. OpenAlex returns
this on the work record as `open_access.is_oa` and `open_access.oa_status`.
Report the overall open-access share in COVERAGE.

Plus these sections:

  1. METHOD — reported the way a systematic review reports a search. Every query
     string verbatim, the database, the date, results returned, screened,
     included, excluded with reasons and counts. Then the overall flow as a
     chain of numbers. Plus: how the starting set was derived, graph sizes and
     round count, and any step you could not complete.
  1b. COVERAGE ESTIMATE — the capture-recapture numbers: how many papers each of
     the two independent strategies found, how many both found, and the
     resulting estimate, with the caveat that it is indicative only.
  2. INVESTIGATOR POOL — everyone derived, their tier, and their derivation path.
  3. COVERAGE — per area, per tier, by year, by work type, open-access share,
     and language. Plus concentration: share from the top three institutions,
     top three journals, and the distribution by corresponding-author country.
     Which areas are thin and why. Surface imbalance; you are not asked to fix
     it.
  3b. INDEPENDENCE CHECK — how many papers the date-driven sweep found that the
     citation-graph expansion did not. Near zero means the two were not
     independent, which invalidates the coverage estimate above.
  4. SELF-TEST — result of Step 5.
  5. UNVERIFIED CANDIDATES — papers you believe belong but could not confirm,
     with what was missing. Expected to be non-empty; an empty section suggests
     the constraints were not applied.
  6. SELF-CONSISTENCY — the eight checks, each pass or fail with counts.
  7. CITATION EDGES — only if you could not run Step 2c/2d, per the note above.