RICO Citation Verification Report
This is a substantive revision, not a reissue. Report versions v1.0 and v1.1 verified RICO v3.8 and v5.7 respectively. Those versions were previously described as permanently locked to the manuscript version they verified. That lock has been lifted by PCP decision of 2026-08-08, on the ground that the lock governs which manuscript version a report verifies, not whether the report's findings are correct, and v1.1 contained an error that could not be corrected without re-issuing.
The error: v1.1 recorded C2 (Olsson et al., 2022) as verified with no fabrication detected, and dispositioned it KEEP. The RICO reference at that time listed an author, "Rai, A.", who does not appear anywhere in the source, and gave DasSarma the wrong initial. The v1.1 record shows the verifier encountered the discrepancy and produced an unparseable note about it rather than assigning the AMBIGUOUS rating that CVP Section 5 Part C requires when support cannot be confidently determined. The certification statement issued on that basis was therefore inaccurate as to C2.
This version re-runs the full protocol against RICO v6.3, in which the reference has been corrected. The failure is documented rather than quietly repaired, because a demonstration artifact for a verification protocol is more useful to a practitioner when it shows the protocol's real failure mode than when it shows only a clean pass.
1. Live Browsing Access Confirmation
Per SF0037 v1.4 Section 4.1. Live access confirmed prior to verification. Full text was retrieved for every citation; no citation was assessed from abstract or metadata alone.
| URL Opened | Content Confirmed | Status |
|---|---|---|
| arxiv.org/abs/2005.14165 | Brown et al. (2020) landing page; 31-author byline and v4 submission history retrieved. | LIVE |
| arxiv.org/pdf/2005.14165 | 75-page full text extracted; few-shot definition located at p.6. | LIVE |
| transformer-circuits.pub/2022/in-context-learning-and-induction-heads/index.html | Full HTML retrieved (124,410 characters of body text); byline and induction-head definition confirmed. | LIVE |
| aclanthology.org/2024.tacl-1.9.pdf | 17-page TACL full text; lost-in-the-middle finding located at p.1. | LIVE |
| arxiv.org/pdf/2404.06654 | 27-page RULER full text; context-length degradation results located at p.1. | LIVE |
| arxiv.org/pdf/2309.17453 | 21-page ICLR 2024 full text; attention-sink mechanism located at p.1. | LIVE |
| doi.org/10.5281/zenodo.18289391 | Concept DOI resolved to record 19155138 (CRD v1.6); PDF openly downloadable. | LIVE |
2. Citation Inventory
Per SF0037 v1.4 Section 4.2. Total citations in RICO v6.3: N = 6.
| ID | Reference | Cited In | Access Class |
|---|---|---|---|
| C1 | Brown et al. (2020). Language models are few-shot learners. arXiv:2005.14165. | 5.3 | FREE-FULLTEXT |
| C2 | Olsson et al. (2022). In-context learning and induction heads. Transformer Circuits Thread. | 5.3, 7.2 | FREE-FULLTEXT |
| C3 | Liu et al. (2024). Lost in the middle. TACL, 12, 157-173. | 1.2 | FREE-FULLTEXT |
| C4 | Hsieh et al. (2024). RULER. arXiv:2404.06654. | 1.2 | FREE-FULLTEXT |
| C5 | Xiao et al. (2023). Attention sinks. arXiv:2309.17453. | 7.3 | FREE-FULLTEXT |
| C6 | Gantz, T. W. (2026). Context representation drift. SF0039. Zenodo. | References only | FREE-FULLTEXT |
3. Claim Map
Per SF0037 v1.4 Section 4.4. Claim text is verbatim from RICO v6.3.
| Cite | Claim ID | Claim Text (verbatim) |
|---|---|---|
| C1, C2 | CLAIM-01 | in-context learning, where models adapt to patterns from examples provided within a prompt (Brown et al., 2020; Olsson et al., 2022) |
| C2 | CLAIM-02 | Olsson et al. (2022) describe induction heads -- circuit components that allow transformer models to continue patterns observed earlier in the context window |
| C3 | CLAIM-03 | Liu et al. (2024) demonstrated that models frequently underutilize information positioned in the middle of long contexts, retrieving content near the boundaries of a context more reliably than content buried in the middle |
| C4 | CLAIM-04 | the limits of effective context utilization across different architectures (Hsieh et al., 2024) |
| C5 | CLAIM-05 | Research on attention sink phenomena demonstrates that attention does not distribute evenly across context tokens (Xiao et al., 2023). Certain tokens accumulate disproportionate attention weight regardless of their semantic relevance. |
| C6 | NONE | No inline claim in RICO v6.3 depends on this citation. It appears in the reference list only. See Section 4, C6, and Section 7 item 1. |
4. Individual Citation Verification Records
5. Verification Summary
| Metric | Result |
|---|---|
| Total citations inventoried (N) | 6 |
| Citations fully processed | 6 / 6 |
| Citations passing PVC (FREE-FULLTEXT) | 6 (C1-C6) |
| Citations mapped to a manuscript claim | 5 (C1-C5); C6 unmapped |
| Citations rated FULL SUPPORT | 5 (C1-C5), across six claim mappings |
| Citations failing Section 4.4 claim-map requirement | 1 (C6) |
| Citations removed or replaced | 0 (one pending PCP decision) |
| Fabricated author names detected and corrected | 1 (C2, corrected in manuscript v6.3) |
| Prior-report findings superseded | 2 (C2 disposition; C6 PVC flag cleared) |
| Bibliographic style notes recorded | 1 (C1 truncation without ellipsis) |
| Verification tier achieved | Tier 1 with PAR records; see note below |
Tier note, stated plainly. SF0037 Section 7 sets Tier 2 (dual-platform verification) as the default for public release. This run achieved Tier 1: each citation was verified against one authoritative first-party source with full text opened and evidence anchors captured, and PAR persistence records (canonical identifiers plus retrieval URLs, and a file hash for C6) were recorded for all six. Independent second-repository confirmation was not performed for every citation. The honest tier is therefore Tier 1 with PAR, not Tier 2. Report v1.1 claimed Tier 2 on the basis of dual confirmation for two of six citations, which overstated the tier; that claim is superseded. Achieving Tier 2 requires a second independent repository check per citation and is recommended before this report is deposited.
6. Certification Statement
I certify that I executed SF0037 CVP v1.4 using live retrieval against public sources. Full text was opened for every citation; no support rating was assigned from an abstract, from metadata, or from model priors.
Total citations inventoried: N = 6
Citations fully processed: 6 / 6
Citations retained after PVC and support verification: 5 unconditional (C1-C5); 1 pending PCP disposition (C6)
Citations removed or replaced: 0
Verification tier achieved: Tier 1 (Single-Platform Verified) with PAR persistence records
Verification completed: August 8, 2026
Verifier: Claude Opus 4.6 (claude-opus-4-6), Anthropic | Prepared by: Thomas W. Gantz, Synthience Institute
7. Recommendations and Notes for Publication
1. C6 (Gantz 2026, CRD, SF0039) — action required. The citation satisfies PVC but fails the Section 4.4 claim-map requirement, because no body-text claim in RICO v6.3 depends on it. Two dispositions are available and the choice is a PCP decision. Remove it from the reference list, which costs nothing evidentially since no claim rests on it. Or reattach it by adding an inline citation at a point where RICO's argument genuinely draws on context representation drift; Section 1.2, which surveys long-context literature, is the natural site, and CRD is thematically adjacent to the Liu and Hsieh material already cited there. Reattachment is the stronger option if and only if the manuscript actually uses the source; adding an inline citation solely to legitimise a reference-list entry would invert the requirement.
2. C1 (Brown et al. 2020) — optional style correction. The seven-name list is a truncation of a 31-author byline presented without an ellipsis. All named authors are correct. If house style follows APA, the form is the first nineteen authors, an ellipsis, then the final author; the common short form is Brown, T. B., et al. (2020). No integrity issue and no urgency.
3. C2 (Olsson et al. 2022) — closed. The fabricated author and the incorrect initial are corrected in RICO v6.3 and verified against the source byline. No further manuscript action.
4. C3 (Liu et al. 2024) — closed. The final TACL citation replaced the preprint numbering per the v1.1 recommendation.
5. Tier upgrade before deposit. Section 5's tier note records that this run is Tier 1 with PAR. A second independent repository confirmation per citation would achieve the Tier 2 default that SF0037 Section 7 sets for public release.
6. Publication location. Report v1.1 was deposited as a standalone Zenodo record. Canon Published records that in hindsight it should have been included in the SF0037 record as a Supporting Artifact under a single DOI, and that future Supporting Artifacts will follow that pattern. This version supersedes v1.1 within the existing SR001-VR1 record, so no new DOI is minted and the structural anomaly is not repeated.
7. On the value of this report as a demonstration artifact. Report v1.1 was published to show what CVP output looks like. It now also shows what a verification failure looks like from the inside: a verifier that retrieved the correct source, encountered a name it could not place, wrote a fluent sentence that resolved nothing, and passed. Section 5 Part C already prescribes the correct behaviour, which is to rate AMBIGUOUS and treat it as failure. The gap was not in the protocol but in its execution, and the class of error is one an automated author-list comparison cannot commit, since a name either appears in the retrieved byline or it does not. Practitioners running CVP should treat any author-name discrepancy as AMBIGUOUS until resolved by direct string comparison against the retrieved source.