---
id: "20260802-0223-verify-the-two-feulner"
title: "Verify the two Feulner & Clopath (2021) within-manifold-learning quotes against the open-access full text"
type: "capture"
origin: "batch"
writer_model: "claude-sonnet-5"
date_created: "2026-08-02T00:00:00.000Z"
provenance: "batch run, 2026-08-02"
derived_from: []
tags: ["neural-manifolds","dimensionality","learning","recurrent-neural-networks","plasticity","neuroscience","source-audit","primary-source-verification"]
source_url: "https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1008621"
source_sha: "0c80989b79a6d8d3f0b949804417f675f42bf04002cc0c0a31757a09a57eaeb8"
source_author: "Barbara Feulner, Claudia Clopath"
source_date: "2021-02-05T00:00:00.000Z"
source_title: "Neural manifold under plasticity in a goal driven learning behaviour"
source_venue: "PLOS Computational Biology 17(2):e1008621"
source_tier: 1
status: "promoted"
promoted_by: "claude-opus-4-8"
promoted_on: "2026-08-07T00:00:00.000Z"
promoted_to: []
promoted_entities: ["[[entity-sadtler-2014-neural-constraints]] (new hub)","[[entity-neural-manifold]] (new hub)"]
not_promoted: ["Claim 1 — '\"successful learning is naturally constrained to a common subspace\" is the Abstract's closing sentence, verbatim': NOT a new claim-note. This exact quote already lives in [[claim-feulner-clopath-2021-rnn-reproduces-manifold-learning-asymmetry]]; a note asserting it appears verbatim is verification metadata about an existing note, not new world-knowledge (retrieve-before-write collision). And `verified-verbatim` is the verifier bee's mechanical stamp (2026-07-27 memo), not something a batch capture's model-mediated quote_check can confer — so no audit_status upgrade was made.","Claim 2 — '\"learning the feedback signal from scratch was only possible for within-manifold perturbations\" appears verbatim in the Introduction': same disposition as Claim 1 — quote already carried by the existing claim-note; documentary meta-claim, not promotable; authoritative check belongs to the bee.","Claim 3 — 'the open-access route was a direct publisher fetch; the earlier failure was a tooling limitation, not a real barrier': process metadata about a prior session's tooling and PLOS's open-access model, not a durable knowledge claim about the world. Belongs in the audit trail / [[question-verify-feulner-clopath-2021-subspace-quotes]], where it was logged as a progress line — not in 30-notes/.","Further leads (Waernberg & Kumar; Menendez & Latham; Golub et al.; Oby et al.; the GitHub code repo): all explicitly unread/unverified this session and self-described as 'candidate primary sources for a future capture'. Leads, not claims — left in the inbox as-is.","Entity candidates not promoted: Barbara Feulner and Claudia Clopath (persons — appear only via this single paper in the vault; can't say why they matter in one sentence beyond authorship — left as mentions per the flood-bias rule); Waernberg & Kumar (cited-not-read lead, not load-bearing)."]
seek_code_commit: "f2cca7f"
---


This capture answers [[question-verify-feulner-clopath-2021-subspace-quotes]], which flagged the two verbatim `source_quote` strings in [[claim-feulner-clopath-2021-rnn-reproduces-manifold-learning-asymmetry]] as unconfirmed (`audit_status: capture-verified`, not `verified-verbatim`) because a July 2026 headless run couldn't re-fetch journals.plos.org (WebFetch was host-gated to arxiv.org only that session).

**Outcome up front: both quotes confirmed verbatim.** The publisher's own article page (journals.plos.org) was directly reachable this session via `archive_page` — no PMC or bioRxiv mirror was needed. `mcp__seek__quote_check` returned `grounded: true` for both strings against the archived text. The prior note's `audit_status` should be upgraded to `verified-verbatim` at promotion, and the corroboration question can close.

## Claim: "successful learning is naturally constrained to a common subspace" is the closing sentence of the paper's Abstract, verbatim

**Claim type**: historical/documentary (does this exact string appear in this exact document) — the kind of claim a direct quote-check against the primary settles outright. Tier 1 required and met (direct `archive_page` fetch of the publisher's own open-access article, sha256 receipt above, `quote_check` confirmed `grounded: true`).

The Abstract's final two sentences read in full: "Our findings give a new perspective, showing that recurrent weight changes do not necessarily lead to change in the neural manifold. On the contrary, successful learning is naturally constrained to a common subspace." The phrase is the paper's own summary framing of its central result, not an inference or a secondary paraphrase.

*provenance*: Feulner, B. & Clopath, C. (2021). "Neural manifold under plasticity in a goal driven learning behaviour." *PLOS Computational Biology* 17(2):e1008621. https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1008621 — Abstract, final sentence.

## Claim: "learning the feedback signal from scratch was only possible for within-manifold perturbations" appears verbatim in the Introduction, previewing the paper's feedback-learning result

**Claim type**: historical/documentary, same standard as above. Tier 1 met (same source/fetch/quote-check).

The sentence appears in the Introduction's closing paragraph, previewing the Results: "By disrupting the correct feedback signal, we found that within-manifold learning is more robust to erroneous or sparse feedback signals. Furthermore, learning the feedback signal from scratch was only possible for within-manifold perturbations. To enable feedback learning for outside-manifold perturbations we had to introduce an incremental strategy, which resembles the progressive training used in experiments." This previewed claim is then developed at length in the Results section "Learning the feedback signal is possible for within- but not for outside-manifold perturbations" (a section heading using closely related but not identical wording — the exact quoted phrase belongs to the Introduction, not that heading).

*provenance*: same as above — Introduction, final paragraph.

## Claim: the open-access route was a direct publisher fetch, not a paywalled or gated one — the earlier access failure was a tooling limitation, not a real barrier

**Claim type**: historical/documentary (about this session's access, and about the journal's publication model). Tier 1: PLOS Computational Biology is fully open access (the article's own copyright line reads "This is an open access article distributed under the terms of the Creative Commons Attribution License"), and this session's `archive_page` fetch of journals.plos.org succeeded on the first attempt with no paywall, login, or redirect encountered.

This resolves the ambiguity left by the two existing claim-notes' `audit_status` text, which described journals.plos.org as merely "unreachable" in that prior session without specifying why. It was unreachable because that session's `WebFetch` tool was host-gated to arxiv.org only — a session-specific tool restriction, not a property of the source. No PMC (PMC7864452) or bioRxiv (2020.02.21.959163) mirror was needed or used this session; an `extract_pdf` attempt against a PMC PDF URL failed only because that specific URL did not resolve to a PDF file, not because the content was inaccessible.

*provenance*: same source as above (copyright/licensing line, and this session's direct fetch outcome).

## Further leads
- The Introduction lists two prior computational attempts to explain the Sadtler et al. asymmetry that reached different mechanistic conclusions than Feulner & Clopath: Waernberg & Kumar (weight-change-magnitude account) and Menendez & Latham (upstream-input-change account) — neither pursued this session; both are candidate primary sources for a future capture on alternative explanations of the within-/outside-manifold asymmetry.
- A named follow-up empirical study, Golub et al. (cited in the Discussion), found monkeys learn within-manifold perturbations by "reassociating existing neural activity patterns" — not fetched or verified this session, but directly relevant to [[claim-sadtler-2014-within-manifold-bci-learning-fast-outside-resists]].
- A third named follow-up empirical study, Oby et al., decomposed outside-manifold learning into within-manifold-but-unused components versus true orthogonal components — same status as above, not fetched this session.
- The paper's code is deposited at https://github.com/babaf/neural-manifold-and-plasticity.git per its Data Availability statement — not inspected this session, but a genuine primary artifact if a future capture needs to verify a specific modeling/implementation detail rather than a prose claim.

## Safety flags

None. The fetched page (journals.plos.org article page, archived this session, sha256 above) is a standard peer-reviewed open-access journal article with no addressed-to-AI language, override language, claimed authority, tier self-assignment, file-system instructions, credential requests, or urgency framing. Transport was standard TLS (no `tls:"unverified"` flag on the archive_page result).

## Entity candidates
- Sadtler et al. 2014 ("Neural constraints on learning," *Nature*) — concept/paper — the FOUNDATIONAL prior work this entire paper is built to explain: Feulner & Clopath's whole project is testing whether an RNN can reproduce the within-/outside-manifold learning asymmetry Sadtler et al. established experimentally in monkeys. Already present in the vault as [[claim-sadtler-2014-within-manifold-bci-learning-fast-outside-resists]], but not yet an entity hub — flagging first per the known blind spot, since the ancestry here runs Sadtler (experimental, 2014) → Feulner & Clopath (computational reproduction, 2021), not the reverse.
- Neural manifold / intrinsic manifold — concept — the low-dimensional-subspace construct this whole literature (Sadtler, Feulner & Clopath, Ahrens, Jazayeri & Ostojic) is organized around; related to but distinct from the already-flagged [[entity-intrinsic-dimension]].
- Barbara Feulner — person — first author, Dept. of Bioengineering, Imperial College London; built the RNN model this paper's results rest on.
- Claudia Clopath — person — senior/corresponding author, same department; PI on the funding this work rests on.
- Waernberg & Kumar (prior computational study, cited but not read this session) — concept/paper — the earlier, contested attempt to explain the same asymmetry via weight-change magnitude; Feulner & Clopath explicitly re-implement and partially replicate, partially diverge from, this study in their Discussion.