Candidate dossier: corrections and familiarity backfire

Source content

Candidate dossier: corrections and familiarity backfire

Status: **draft research candidate; corrected after three failed exact-head reviews; not accepted

catalog knowledge**

Task: EM-0019

Candidate identity:

em:dossier:sha256:4339d6d6d52b9b534d2e63c95f52ff3cf90be5264f567762f47bab9af4d945a7

Human question

Does repeating misinformation inside an evidence-based correction generally make people

believe or rely on that misinformation more than they otherwise would?

This is narrower than “can corrections ever fail?” It separates:

• familiarity backfire: belief or reliance rises above a no-correction or pre-correction

baseline after a correction repeats a claim;

• continued influence: a correction helps but does not fully erase an earlier effect;
• worldview backfire: a proposed identity- or prior-attitude mechanism; and
• context-specific backfire: an adverse result in one message design, sample, delay, or

measure that does not establish a general familiarity mechanism.

Collapsing those outcomes would manufacture a stronger claim than the experiments test.

Current answer, with uncertainty preserved

Interpretation, not admitted global fact: target-comparable backfire has been reported in

some tested formats and contexts, including unlicensed negated corrections and two

vaccine-message studies. Two other often-cited supportive experiments are memory or rehearsal

boundaries without the target baseline. Twelve modeled counterevidence data roots include

flu-message, retraction, correction-versus-control, standalone-negation, replication, and

brand-misinformation tests. This bounded record does not support a blanket rule that

communicators should avoid repeating misinformation inside corrections. It also does not show

that backfire is impossible.

The causal story remains unresolved. A result compatible with familiarity can also involve

pragmatic licensing, worldview, measurement, correction skepticism, or another mechanism.

Derivable counts and their unit

The builder declares one counting unit: **one participant-data lineage per publication-defined

data series**. Multiple experiments within one publication are collapsed. Separate publications

with separately reported samples remain roots unless data reuse is evidenced. Author, method,

material, and research-program overlap remain separate disclosed dimensions.

Assertion sets are also derived. They are the unique assertion endpoints of modeled support

or rebuttal relations; review/report assertions are support endpoints whose own lineage has no

data dimension. The builder passes those derived keys to independence_summary; no displayed

number is an editorially typed evidence total.

The deterministic calculation currently reports:

| Derived quantity | Count |

| --- | ---: |

| Modeled apparent-support assertions | 10 |

| Known support-side participant-data roots | 6 |

| Unresolved support-side participant-data roots | 1 |

| Target-comparable supportive participant-data roots | 4 |

| Target-comparable unresolved participant-data roots | 1 |

| Modeled review/report assertions receiving no new-data credit | 4 |

| Modeled counterevidence assertions | 12 |

| Modeled counterevidence participant-data roots | 12 |

The six known support-side roots are Skurnik et al. 2005, Peter and Koch 2016, Autry and

Duarte 2021, Pluviano et al. 2017, Pluviano et al. 2019, and Thomas and Autry 2024. Skurnik

2005 is excluded from the target-comparable subcount because Ecker et al. 2020 report that it

lacked the needed baseline. Peter and Koch is also excluded because the full method shows that

all participants read the same correction article and only judgment instructions varied; the

reported truth-memory error is not correction versus no correction or pre-correction belief.

The unresolved root is the unpublished Skurnik, Yoon, and Schwarz 2007 flu-flyer manuscript.

The four review/report assertions are the 2011 handbook page, Schwarz et al. 2016’s general

claim and separate flu report, and Ecker et al. 2020’s historical report. They add mentions and

lineage information, not participant-data roots.

The twelve counterevidence roots are Cameron et al. 2013; Ecker, Hogan, and Lewandowsky 2017;

Rich and Zaragoza 2016; Swire, Ecker, and Lewandowsky 2017; Gordon, Ecker, and Lewandowsky

2019; Ecker et al.’s short-format study; Ecker, Lewandowsky, and Chadwick 2020; Carnahan and

Garrett 2020; Swire-Thompson et al. 2022; Prike et al. 2023; Ecker, Sharkey, and

Swire-Thompson 2023; and Nibat et al. 2026.

These are participant-data counts, not “independent research program” counts.

The corrected 2005/2007 lineage

The first independent review found a material error in the initial candidate: it treated a

flu-flyer result as if it came from the published Skurnik et al. 2005 paper. The corrected model

keeps two roots distinct:

1. Skurnik et al. 2005 is a published two-experiment consumer-claim study, with a notable

older-adult result and no target-comparable baseline.

2. Skurnik, Yoon, and Schwarz 2007 is an unpublished flu-flyer manuscript described by later

sources. Its manuscript, sample record, and participant data remain unavailable, so it is an

unknown lineage and receives zero evidential credit.

The public 2011 handbook page describes the flu flyer and a 30-minute result but gives the 2005

paper as its sole numbered reference. Schwarz et al. 2016 instead maps its older-adult passage to

reference 42 (the 2005 paper) and its flu-flyer passage to reference 6 (a 2007 review chapter).

Ecker et al. 2020 identifies that review chapter as discussing the unpublished 2007 manuscript.

The citation mismatch is retained as an undercutting relation. It is not silently repaired by

making the flu-flyer proposition depend on the nonmatching 2005 data.

Source search and cutoff

Search cutoff: 2026-08-22T17:32:37Z.

The search began from the 2011 handbook and Skurnik 2005, then followed backward references,

forward discussion, author/title searches, and replication/counterevidence terms for

“familiarity backfire,” correction, repetition, and misinformation. Exact editions and metadata

were checked through publisher or institutional pages, Europe PMC, Crossref, and linked OSF

projects. Inclusion required one of:

• participant data testing the target or a material boundary condition;
• an exact source span needed to resolve a data, method, or citation lineage; or
• material replication, rebuttal, qualification, or undercutting evidence.

This is a bounded lineage case study, not a systematic review or an exhaustive effect estimate.

Within that frame, a directly modeled result requires an exact primary or authoritative target

span that could be independently retrieved. Reviews, guidance pages, and historical reports

remain visible but receive no new-data credit. Restricted works retain only quote-minimal

attributed spans. Search results and inaccessible candidate full texts are not treated as source

evidence.

Source editions, custody, and reuse

Only exact excerpt packets are committed. Full provider artifacts and data files remain outside

the repository.

| Work | Exact edition and artifact receipt | Public treatment |

| --- | --- | --- |

| [Skurnik et al. 2005](https://doi.org/10.1086/426605) | University-hosted journal PDF; SHA-256 `c4893504537d256cff0c37a58b17aca0d6ceda11d9c5cea1aa3c066a544108f2`; 331,989 bytes | Restricted; two short abstract sentences; PDF not redistributed |

| [Cook and Lewandowsky 2011](https://skepticalscience.com/Debunking-Handbook-Part-2-Familiarity-Backfire-Effect.html) | Live HTML with later update notices; SHA-256 `bfd2e4ff7e2a5cc5a0c8f2e7aee70680cdc5eabbcf2333c4ef2a7ac3e3cfc3c4`; 89,663 bytes | Copyright notice present; short excerpts and reference only |

| [Schwarz, Newman, and Leach 2016](https://doi.org/10.1177/237946151600200110) | SAGE publisher PDF, journal pp. 85–95; automated capture returned HTTP 403; legacy DOI `10.1353/bsp.2016.0009` identifies the same work | No article-specific CC grant confirmed; independently read quote-minimal spans only |

| [Peter and Koch 2016](https://doi.org/10.1177/1075547015613523) | LMU institutional full-text PDF; two byte-identical retrievals; SHA-256 `ab18125517a34d469a3b644e0c7a3b6cfc0a59f82ac4c39e14845a65d73bc4ec`; 118,545 bytes | No open reuse grant confirmed; quote-minimal result and method spans; PDF not redistributed |

| [Wilson and Park 2008](https://doi.org/10.1016/j.pec.2008.02.008) | PubMed XML, PMID 18394856; two byte-identical retrievals; SHA-256 `230d805c865e5f4348b46e5adaeba5cb9a09081ab1d87c7a2363c4597ebc44e1`; 7,630 bytes | No open article reuse license confirmed; quote-minimal abstract method and result spans |

| [Berinsky 2017](https://doi.org/10.1017/S0007123415000186) | Author-hosted Cambridge journal PDF; two byte-identical retrievals; SHA-256 `d5b189db8963f8d3b412d4bbf6cddb9cfa52198f330f54929519224f99975c78`; 209,237 bytes | No open reuse grant confirmed; quote-minimal design, result, and qualifier spans; PDF not redistributed |

| [Cameron et al. 2013](https://doi.org/10.1016/j.pec.2013.06.017) | NCBI BioC XML, PMCID PMC3772650; two byte-identical retrievals; SHA-256 `40b29203eb82caf07c0a0c8d46438f01efb21795b39941aea03b3c898cd36ecb`; 78,951 bytes | Author manuscript permits text mining/fair-use treatment; short abstract results only |

| [Ecker, Lewandowsky, and Chadwick 2020](https://doi.org/10.1186/s41235-020-00241-6) | Europe PMC JATS XML, PMCID PMC7447737; SHA-256 `15620a5e73fe87bb19f3f37c84a4d1a0f7efa9b50be3a5d2749a07f588eb4047`; 208,195 bytes | CC BY 4.0; exact result, history, method, and implications spans |

| [Ecker et al. short-format study](https://doi.org/10.1111/bjop.12383) | Crossref deposited abstract record; SHA-256 `07408e83b8ff2702d8e7615829427bc2e93c5d0d586642ee5b67d05eb78c6ee5`; 10,954 bytes | CC BY-NC-ND 4.0 for accepted manuscript and VOR; short attributed abstract result only |

| [Ecker, Hogan, and Lewandowsky 2017](https://doi.org/10.1016/j.jarmac.2017.01.014) | Bristol accepted manuscript; raw captures are provider-generated but reproduce `pdftotext` SHA-256 `cbf940e71d4d1eb854ccac6150418ed0bbe31a0c697cee8408f4199097806e97`, 65,679 bytes | Bristol metadata labels the manuscript CC BY-NC-ND; quote-minimal result and method extents; PDF not redistributed |

| [Wahlheim, Alexander, and Peske 2020](https://doi.org/10.1177/0956797620952797) | Coauthor-lab publisher PDF; two byte-identical retrievals; SHA-256 `588addfe9659b726d04b1cc51e004d76a76a8cf73f6053d7bb10c00b1e378781`; 1,095,068 bytes | No open reuse grant confirmed; quote-minimal abstract design, endpoint, and result spans; PDF not redistributed |

| [Kemp, Alexander, and Wahlheim 2022](https://doi.org/10.1186/s41235-022-00434-1) | Europe PMC JATS XML, PMCID PMC9481799; two byte-identical retrievals; SHA-256 `6a5fc0db04a875ece25cb374392b63c6a2d43d46f1b4f0ae295bce017d8606ef`; 169,173 bytes | CC BY 4.0; exact design, control-boundary, and mixed-result spans |

| [Kemp, Loaiza, and Wahlheim 2022](https://doi.org/10.1038/s41598-022-25649-6) | Europe PMC JATS XML, PMCID PMC9758464; two byte-identical retrievals; SHA-256 `5540d1b2e31d199bf3623bd0e1f8e1788929a617aca80bd713ac09315e0bda01`; 113,222 bytes | CC BY 4.0; exact design, result, and isolation-limitation spans |

| [Carnahan and Garrett 2020](https://doi.org/10.1093/ijpor/edz037) | Oxford Academic publisher full-text read-back; automated capture returned HTTP 403 | No open reuse license confirmed; four independently read quote-minimal result fragments |

| [Rich and Zaragoza 2016](https://doi.org/10.1037/xlm0000155) | APA publisher-PDF read-back; automated capture returned an anti-bot HTML challenge | No open reuse license confirmed; independently read quote-minimal result and design fragments |

| [Swire, Ecker, and Lewandowsky 2017](https://doi.org/10.1037/xlm0000422) | Coauthor-lab accepted-manuscript PDF; two byte-identical retrievals; SHA-256 `e0f690593e78099431b4d555ea1714d448482b2dd352f1acfa451c0da433c11b`; 654,850 bytes | No open reuse license confirmed; three quote-minimal result sentences; PDF not redistributed |

| [Gordon, Ecker, and Lewandowsky 2019](https://doi.org/10.1016/j.jml.2019.104028) | Coauthor-lab preprint PDF; two byte-identical retrievals; SHA-256 `23d389cd19eecb76c8e21ae52a6e5f029bcd44a9e0e399f013376f1ac670cf06`; 747,048 bytes | UWA metadata labels the accepted manuscript CC BY-NC-ND; two quote-minimal result extents; PDF not redistributed |

| [Prike et al. 2023](https://doi.org/10.1186/s41235-023-00492-z) | Europe PMC JATS XML, PMCID PMC10317933; SHA-256 `a546f640397adf7ac208f040634e480181588e5c29e1a2c76f5a301ef728e019`; 172,943 bytes | CC BY 4.0; exact outcome and method spans |

| [Ecker, Sharkey, and Swire-Thompson 2023](https://doi.org/10.1371/journal.pone.0281140) | Europe PMC JATS XML, PMCID PMC10096191; SHA-256 `44ebae212686059b211ae7a9e8757958d8851d73b51513c6f9a2f9c9a518f88e`; 167,791 bytes | CC BY 4.0; exact outcome, sample, material, and ethics spans |

| [Autry and Duarte 2021](https://doi.org/10.1002/acp.3823) | Crossref record; SHA-256 `39293b19eb31c866b74e5017f7af0377f52a227162b034ad1f4c84471e82906d`; 9,943 bytes. OSF experiment-data receipts: `bc069a90152b887116cd4972df4376844570ba25cefbfa1055668c78eeaca80a` (148,954 bytes) and `de9f27b026ceadea0a6309ef695be06bc06c1fa35ba9209260f246d89b108ea5` (216,736 bytes) | VOR identified as CC BY 4.0; OSF project `yweu6` declares CC0 1.0 Universal at project scope; publisher qualifier and timing spans require independent read-back |

| [Pluviano, Watt, and Della Sala 2017](https://doi.org/10.1371/journal.pone.0181640) | Europe PMC JATS XML, PMCID PMC5547702; SHA-256 `87e363cb1cae5896c42c7e7e0295bd2e6969637473d14ff76deea3c740024377`; 114,003 bytes | CC BY 4.0; exact result and limitation spans |

| [Pluviano et al. 2019](https://doi.org/10.1007/s10339-019-00919-w) | University of Edinburgh accepted manuscript; provider-generated raw captures `42968a…` and `97b713…`, each 762,139 bytes; both produce the same raw `pdftotext` output, SHA-256 `cd054a8f7e1bcb31a6d7f700b775eb936c9e0d3f43899cd8411b01f6252ff670`, 36,848 bytes | Restricted; raw PDF identity is capture-specific because the endpoint regenerates a timestamped cover; semantic text identity, short spans, and both raw capture receipts are preserved; PDF not redistributed |

| [Thomas and Autry 2024](https://doi.org/10.1002/acp.70004) | Crossref record; SHA-256 `f71d065da32f2179ff262008a863cae22aeec81b01e59997c994a1871e87bf23`; 11,291 bytes. OSF data receipt: `4a65116f6f10b02dc34b8353a228a16604a48ebcca576b2aae91ad5073f4d66c` (644,279 bytes) | Restricted VOR; short abstract spans only |

| [Swire-Thompson et al. 2022](https://doi.org/10.1037/xge0001131) | NCBI BioC XML, PMCID PMC9283209; two byte-identical retrievals; SHA-256 `0480d8d31af7c1525433838c91b8fc49f33997b95038ded5de40f255e54af5a2`; 93,558 bytes | Author manuscript permits text mining/fair-use treatment; short abstract design and result only |

| [Nibat et al. 2026](https://doi.org/10.1016/j.ijresmar.2026.03.007) | Exact Sabancı University PDF URL for the open-access Elsevier corrected proof; SHA-256 `99075c8a58b19d3be262486ccb78b3508747887e350802e5b41b74390b13d3bc`; 2,525,138 bytes; landing page retained separately | CC BY 4.0; exact abstract scope and result spans |

Static artifact hashes identify the exact editions examined. The Pluviano 2019 exception is

explicitly typed as dynamic: every reviewer records their capture-specific raw digest and must

also reproduce the declared pdftotext semantic digest. A changed semantic digest fails closed.

Material qualifiers and counterevidence

| Result | What must travel with it |

| --- | --- |

| Skurnik 2005 | Older-adult and consumer-claim scope; no baseline for target-comparable backfire |

| Unpublished Skurnik 2007 | Later reports describe it, but the manuscript and participant-data lineage are unavailable |

| Peter and Koch 2016 | All participants read the same correction article; the study varies judgment instructions and delay, so its truth-memory errors lack the target baseline |

| Cameron 2013 | Flu-message knowledge and recall outcomes; supports counterevidence for this target without proving every facts-and-myths format safe |

| Autry 2021 | Experiment 2’s relevant contrast was nonsignificant; Experiment 1 used a five-minute delay, Experiment 2 no delay, and longer persistence remains open |

| Pluviano 2017 | Target-comparable result, with a convenience sample and stated generalizability limits; the result alone does not establish familiarity as mechanism |

| Pluviano 2019 | New parent participants, but the same author program and a delay chosen for direct comparison with the 2017 study |

| Thomas 2024 | Target-comparable unlicensed-negation result with a proposed Gricean/pragmatic mechanism, not an established familiarity mechanism |

| Ecker 2020 | A weak first result did not recur in two follow-ups; its tested conditions do not prove impossibility elsewhere |

| Ecker short-format study | Simple retractions reduced belief without familiarity backfire in the tested designs and delays |

| Ecker, Hogan, and Lewandowsky 2017 | Every retraction reduced reliance versus no retraction; explicit repetition improved the retraction versus no reminder, but every retraction also supplied an alternative causal account |

| Rich and Zaragoza 2016 | Experiment 1’s control contrast was null and Experiment 2’s reduced reliance; the nominal control story still mentioned the suspect and may have weakly implicated him |

| Swire 2017 | Post-correction myth belief remained below the pre-correction baseline in both experiments and across the tested delays |

| Gordon 2019 | The only-negated comparison sits inside a continued-influence paradigm; Experiment 2 supplies a direct no-misinformation control, but the Bayes factor was weak and the negated claim was licensed by a stereotype |

| Carnahan and Garrett 2020 | Study 2’s repeating correction formats improved accuracy versus an unrelated-story control; processing style moderated format effects and Study 1 lacked a true control |

| Swire-Thompson 2022 | Two correction-versus-control experiments found no item backfire; reliability and item-level measurement remain central to its interpretation |

| Prike 2023 | No immediate or one-week result in two experiments; a skepticism result appeared in open responses but not the rating-scale measure |

| Ecker 2023 | No vaccine-myth backfire observed; a replication attempt using the 2017 materials |

| Nibat 2026 | Five samples totaling 4,337 participants form one paper-series data root, not five independent roots |

Independence limits

The candidate records new participant collections separately while refusing to call them fully

independent programs:

• Ecker 2020 and Prike 2023 used the same US MTurk/CloudResearch eligibility frame, including

5,000-HIT and 97% approval thresholds; cross-paper participant overlap is unreported.

• Prike 2023 and Ecker 2023 share authors and UWA ethics approval RA/4/20/6423.
• Ecker 2023 directly reused Pluviano 2017 stimuli.
• Pluviano 2017 and 2019 share three authors and a comparison-oriented method/material family.
• Autry 2021 and Thomas 2024 share Autry and the unlicensed-negation program. They remain separate

data roots, but no direct method-dependence edge is asserted without an exact Thomas methods

span.

• Ecker, Hogan, and Lewandowsky 2017 and Gordon, Ecker, and Lewandowsky 2019 share two authors

and the continued-influence correction program. Their participant collections remain separate,

while the method/social dependence is modeled explicitly.

• Seven modeled counterevidence papers include Ecker; four include Swire/Swire-Thompson; four

include Lewandowsky. Different paper titles do not establish social or theoretical independence.

Method, material, and social lineages are separate from participant-data lineages. This prevents

the current dimension-agnostic dependency traversal from collapsing a genuinely new sample merely

because its method was reused.

Candidate and negative-result log

EM-0019 considered the initial product-direction candidates before selecting this one:

• Selected and narrowed: “Why repetition can change belief without adding evidence” became

the mechanism-specific correction question. It has accessible primary experiments, a visible

lineage collapse, counterevidence, qualifiers, and a practical human question.

• Deferred: “Information is not meaning.” Shannon, semantic-information, and truth

distinctions matter, but the first slice would be mainly conceptual and would not expose an

equally clear empirical lineage-count interaction.

• Deferred: “When justified true belief still is not knowledge.” A Gettier dossier remains

promising, but counting argumentative lineages before specifying an argument-dependence model

would overstate what the alpha dossier represents.

• Deferred: worldview backfire. It risks making politics the first realm and combines a

different mechanism with familiarity.

• Deferred: “What can an AI responsibly claim to know?” It is fast-moving and too adjacent to

Epistemedia itself for the first outward-facing proof.

• Excluded from evidential counts: the 2012 Association for Psychological Science summary;

its broad “sometimes backfire” framing does not distinguish the relevant mechanisms.

• Included after review expanded the bounded frame: Cameron et al. 2013; Ecker, Hogan,

and Lewandowsky 2017; Rich and Zaragoza 2016; Swire et al. 2017; Gordon et al. 2019;

the Ecker et al. short-format study; Carnahan and Garrett 2020; and Swire-Thompson et al.

2022 as directly captured counterevidence roots. Peter and Koch 2016 remains a known

support-side memory-error root outside the target-comparable count.

• No new-data credit: handbook, review, and historical-report assertions that summarize an

upstream experiment.

• Zero credit, preserved as unknown: the unpublished 2007 flu-flyer participant-data root.
• Not target-comparable, preserved as known: Skurnik 2005 because it lacks the needed baseline.

Closed bounded-study disposition ledger

The following ledger disposes every participant-data work named in Nibat et al. 2026 Table 1

and every direct counter-study named in the captured Prike et al. 2023 search passages. “Not in

the target count” is not a claim that a study is unimportant; it means its captured comparison

does not ask whether a correction raises belief or reliance above a no-correction or

pre-correction baseline.

| Surfaced work | Disposition | Reason | Exact dossier basis |

| --- | --- | --- | --- |

| Skurnik et al. 2005 | Modeled support-side root; not target-comparable | Published data are captured, but the later primary report states that the needed baseline was absent | `span-skurnik-summary`; `span-ecker-2020-no-baseline` |

| Skurnik, Yoon, and Schwarz 2007 | Modeled unresolved support root | Later sources describe the flu-flyer result; manuscript and participant-data identity remain unavailable | `span-ecker-2020-unpublished-lineage`; `span-schwarz-flu-2007-memory`; `span-schwarz-flu-2007-intentions` |

| Wilson and Park 2008 | Excluded from target count | Positive/negative health-statement recognition task; no misinformation correction or target baseline | `span-wilson-park-2008-method`; `span-wilson-park-2008-result` |

| Cameron et al. 2013 | Modeled counter root | Facts-and-myths flu messages improved knowledge and were not counterproductive to recall in the captured outcomes | `span-cameron-2013-knowledge`; `span-cameron-2013-no-counterproductive` |

| Peter and Koch 2016 | Modeled support-side root; not target-comparable | Everyone read the same correction article; the randomized factor was judgment timing and the outcome was truth-memory error | `span-peter-koch-2016-shared-stimulus`; `span-peter-koch-2016-randomization`; `span-peter-koch-2016-measure`; `span-peter-koch-2016-error-definition` |

| Rich and Zaragoza 2016 | Modeled counter root | Corrected versus uncorrected nominal-control comparisons were null in Experiment 1 and beneficial in Experiment 2; the story’s mention of the suspect remains a qualifier | `span-rich-2016-experiment-1-result`; `span-rich-2016-experiment-2-result`; `span-rich-2016-control-qualifier` |

| Berinsky 2017 | Excluded from target count; political/source/rehearsal boundary | Study 2 had no control group; the reported rehearsal difference within rumor-plus-correction was not significant at the stated threshold | `span-berinsky-2017-no-control`; `span-berinsky-2017-rehearsal-result`; `span-berinsky-2017-rehearsal-qualifier` |

| Ecker, Hogan, and Lewandowsky 2017 | Modeled counter root | Every retraction reduced reliance versus no retraction, and explicit misinformation repetition improved the retraction versus no reminder; every retraction also supplied an alternative causal account | `span-ecker-hogan-2017-result`; `span-ecker-hogan-2017-no-retraction-result`; `span-ecker-hogan-2017-repeat-result`; `span-ecker-hogan-2017-alternative-account` |

| Pluviano et al. 2017 | Modeled support root | Target-comparable vaccine-message result and limitations are captured | `span-pluviano-2017-result`; `span-pluviano-2017-limitation` |

| Swire, Ecker, and Lewandowsky 2017 | Modeled counter root | In both experiments post-correction myth belief stayed below the pre-correction baseline | `span-swire-2017-experiment-1-result`; `span-swire-2017-experiment-2-result`; `span-swire-2017-conclusion` |

| Ecker, O'Reilly, Reid, and Chang 2019/2020 | Modeled counter root | Same short-format paper under online-first and print citations; simple retractions showed no familiarity backfire | `span-ecker-short-2020-result` |

| Pluviano et al. 2019 | Modeled support root | Target-comparable parent-sample result; comparison method remains dependent on the 2017 program | `span-pluviano-2019-result`; `span-pluviano-2019-dependence` |

| Carnahan and Garrett 2020 | Modeled counter root | Study 2’s repeating correction formats all improved accuracy versus an unrelated-story control; Study 1 receives no separate target credit | `span-carnahan-2020-control-result`; `span-carnahan-2020-pairwise-result`; `span-carnahan-2020-processing-boundary` |

| Ecker, Sharkey, and Swire-Thompson 2023 | Modeled counter root | Direct vaccine-material replication found no target backfire | `span-ecker-2023-result`; `span-ecker-2023-materials` |

| Thomas and Autry 2024 | Modeled support root | Target-comparable unlicensed-negation result is captured | `span-thomas-2024-result`; `span-thomas-2024-mechanism` |

| Nibat et al. 2026 | Modeled counter root | Five-study publication directly compares repetition and correction and reports no target effect | `span-nibat-2026-scope`; `span-nibat-2026-result` |

| Gordon, Ecker, and Lewandowsky 2019 | Modeled counter root | Experiment 2’s standalone negation produced fewer inferences than a no-misinformation control, with very weak Bayesian evidence for the difference | `span-gordon-2019-result`; `span-gordon-2019-qualifier` |

| Kemp, Alexander, and Wahlheim 2022 | Excluded from target count; memory/retrieval boundary | The control is a different real-news item introduced in Phase 2; the mixed result is conditioned on later recollection and concerns real-news memory | `span-kemp-2022a-design`; `span-kemp-2022a-control-boundary`; `span-kemp-2022a-mixed-result` |

| Kemp, Loaiza, and Wahlheim 2022 | Excluded from target count; reminder/label boundary | Compares correction formats and reports memory/belief accuracy; the paper states that the design does not fully isolate the reminder effect | `span-kemp-2022b-design`; `span-kemp-2022b-result`; `span-kemp-2022b-limitation` |

| Wahlheim, Alexander, and Peske 2020 | Excluded from target count; memory/belief boundary | Compares corrections with and without reminders, but belief ratings attach to generated recalls rather than a common false-claim baseline | `span-wahlheim-2020-design`; `span-wahlheim-2020-endpoint`; `span-wahlheim-2020-result` |

| Ecker, Lewandowsky, and Chadwick 2020 | Modeled counter root | Two higher-powered follow-ups yielded evidence against target backfire and the paper rejects a blanket no-repetition recommendation | `span-ecker-2020-result`; `span-ecker-2020-guidance-conclusion` |

| Prike et al. 2023 | Modeled counter root | No immediate or one-week target result; the skepticism finding was measure-specific | `span-prike-2023-result` |

| Swire-Thompson et al. 2022 | Modeled counter root | Two correction-versus-control experiments found no item backfire | `span-swire-thompson-2022-result` |

| Autry and Duarte 2021 | Modeled support root | One target-comparable unlicensed-negation contrast was significant and one was not; timing and persistence limits remain | `span-autry-2021-result`; `span-autry-2021-qualifier`; `span-autry-2021-timing-limit` |

| Ecker et al. 2011 | Excluded from target count | Historical search span reports stronger continued-influence reduction from repeated retractions, not the target baseline | `span-ecker-2020-ecker-2011-search` |

| Ecker et al. 2022 and Swire-Thompson et al. 2020 | No new-data credit | Review/synthesis sources used to discover and classify upstream studies | `span-prike-2023-counter-search` |

This closes the declared ledger through the search cutoff. It does not claim that no other paper

exists; newly discovered directly comparable participant data must reopen the ledger rather than

being silently absorbed into the displayed counts.

The first independent review of candidate head

25ef8c1434034cec92b62c0c02c4b5b56e8d1926 failed materially. No pass receipt was created. The

shared branch was not rewritten. A second review of successor head

fa99719b0c1e0f9815e6e97f5aef673b8696bb3d also failed and produced no receipt. It found the

dynamic Pluviano capture, noncontiguous Schwarz spans, Nibat URL, Autry OSF license,

CloudResearch provenance, search-frame, and count-derivation defects repaired in this successor.

The third review of ad8bec769c1dac1d30be3b6984456850db0b3054 independently passed all

artifact, exact-span, derivation, and repository checks but failed the search-ledger gate because

Swire 2017 and Gordon 2019 were publicly recoverable and unmodeled. No receipt was created.

The corrected candidate must receive a fresh review against its exact final head.

Reproduce the candidate and counts

From the repository root:

.venv/bin/python research/how-we-know/corrections-backfire/build_candidate.py --check

Regenerate only when reviewed material intentionally changes:

.venv/bin/python research/how-we-know/corrections-backfire/build_candidate.py --write

Place the twenty-two captured primary artifacts and three public data workbooks under one temporary

directory using the filenames declared in verify_retrievals.py, then verify their bytes and all

artifact-backed excerpts:

.venv/bin/python research/how-we-know/corrections-backfire/verify_retrievals.py \
  --artifact-dir /tmp/em0019-artifacts

Before an independently authored pass receipt exists, the following must fail closed:

.venv/bin/python research/how-we-know/corrections-backfire/verify_retrievals.py \
  --artifact-dir /tmp/em0019-artifacts \
  --require-complete

After review, validate the tracked receipt explicitly:

.venv/bin/python research/how-we-know/corrections-backfire/verify_retrievals.py \
  --artifact-dir /tmp/em0019-artifacts \
  --review-receipt \
  research/how-we-know/corrections-backfire/review-receipts/<receipt>.json \
  --require-complete

Finally run:

make check

Independent-review receipt gate

The authoring agent cannot satisfy the gate by checking its own work again. A fresh, independently

rooted reviewer must retrieve the sources without using the authoring agent’s temporary artifacts,

review the exact committed head, and author an append-only JSON receipt under:

research/how-we-know/corrections-backfire/review-receipts/

The local format is epistemedia-research-review-receipt-v0.1; it does not claim to be a global

protocol schema. The verifier requires:

• reviewer codex-independent-reviewer, fresh-clone and independent-retrieval declarations, and

an explicit declaration that authoring-agent artifacts were not used;

• repository URL, exact reviewed head, current base, tree, binary-diff digest, dossier identity,

packet digest, and digests and lengths of all four packet files;

• UTC start and completion times;
• exact argv, working directory, exit code, and stdout/stderr digests for the candidate build,

retrieval verifier, and full repository check;

• one captured, captured-dynamic, or unavailable receipt for every primary edition, plus recorded

data and read-back sources; dynamic captures must include both raw and semantic identities;

• every candidate span’s locator, expected and observed extent digest, verification mode, and

match result;

• findings, limitations, clean pre/post Git state, and decision pass.

The packet digest sorts the four repository-relative paths, then hashes for each path: UTF-8 path,

a NUL byte, ASCII byte length, a NUL byte, and the exact file bytes. The binary-diff digest hashes

the raw output of:

git diff --binary --full-index --no-ext-diff \
  <reviewed-base> <reviewed-head> -- \
  research/how-we-know/corrections-backfire

The receipt is stale if the base, reviewed head, tree, diff, dossier, README, builder, generated

candidate, verifier, source record, or span changes. After the reviewed head, only receipt files

may change before validation. An ordinary task run receipt cannot substitute for this source-span

review receipt.

Until the exact corrected head has a valid independently authored pass receipt, this remains a

reviewable research proposal, not accepted knowledge and not homepage copy.