Evidence, witnessing, and platformed amnesia in a harassment archive.
Records do not become evidence simply because they are true, numerous, or painfully preserved. Evidence does not become public truth simply because it exists. Between a saved trace and a shared reality sit claims, readers, witnesses, circulation paths, platform systems, and the social power to decide whose records matter.
This page is written from two specific projects on this site: a closed case study of coordinated harassment across Tumblr, AO3, and adjacent communities (2024–2026), and a forensic analyzer that runs syntax-first counts across five persona corpora. The terms below are named after what these materials needed and what they failed to receive. Theoretical traditions named at the end of the page locate this work in nearby intellectual lineages, but the terms come from the case first.
First-time readers may want to start with the case page and then return here; this page names the method after the archive is already in view.
The harassment archive began as a survival record: an attempt to preserve attacks, responses, platform decisions, and changing narratives before they could be erased or reversed.
That record assumed a painful but common premise: if enough reaction, repetition, and contradiction were preserved, the truth would become self-evident. It did not. By the time the case closed, around 5,700 archived posts existed across five attacker corpora; roughly 100 outlets and NGOs had been contacted; fewer than five had replied, all rejections, none from a journalist or academic.
The failure was not only evidentiary. It was also social: without witnesses willing to receive, repeat, cite, or defend the claim, evidence can remain trapped inside the archive that proves it.
I preserved what happened.
This material supports a specific judgment.
Others can receive, repeat, cite, or act on the claim.
Evidence Spine
The problem above is not only that records were ignored. It is that truth changes form as it moves from raw trace to readable public record. Each step below names the conversion and binds it to a concrete artifact in this archive.
Raw record → Event → Claim → Evidence → Pattern → Public record
In this archive, that movement runs through 217 raw files, a 51-event timeline, a narrow claim about AYWS as post-exposure infrastructure, counted excerpts from the LadyYomi document, cross-corpus harassment patterns, and the public case page, full report, and analyzer.
The operative claim here is deliberately narrow: not simply "a campaign existed", but that these accounts coordinated, across these platforms, through these specific patterns, with these outcomes.
Not stages in a sequence. Each is a fate a record can take or be denied, often in parallel within the same case.
Forms truth takes
The record exists. It protects memory against deletion, denial, and self-doubt, but may still remain unread outside the person who preserved it.
The 50-chapter AO3 report was complete the day before AO3 ruled against the appeal. Completeness was never the missing piece.
The record has been organized around a claim and shaped into a form that can travel: a timeline, a stable frame, a short version, a concept that lets others repeat it without rereading the whole archive.
The forensic brief, the timeline, the network graph, and the analyzer each take the same materials and make a different claim travelable.
Forms truth fails
A platform, institution, moderator, or rule system recognizes some version of the claim. This version may be narrower, cheaper, or less truthful than the lived record.
AO3 received 60 pages of cross-linguistic harassment evidence and returned a single 15-day ban — addressing a few posts in isolation, not the coordinated pattern across five accounts and 18 months.
Deletion, account loss, social withdrawal, platform churn, or hostile reframing breaks the circulation chain, making a documented event appear as if it never mattered.
Bluesky: permanently lost. Medium: permanently lost. AO3: account self-deleted after the ruling ('申诉是一种耻辱' — appeal as humiliation). Three of the four platforms that hosted this case no longer host the survivor's voice at all.
Five terms named after what this case needed and what it could not receive. The first three name absences. The last two name active hostilities present in the wrong direction. Each is followed by the specific material it points at.
Forms of absence
The people, links, summaries, citations, archives, feeds, search paths, and trusted nodes that let a truth survive beyond its original witness.
Demonstrated by absence: 100+ outlets contacted, fewer than five replies, zero from journalism or academia. The infrastructure that would have carried the claim outward did not exist for this case.
A record that has not yet found readers. It is not useless, but its public force is delayed, suspended, or limited to future retrieval.
The full 33k-word AO3 report, the 60-page brief, and 217 raw evidence files are preserved here precisely because they were nearly erased — a record waiting for a reader the original platforms did not deliver.
The way deletions, bans, migrations, private drama spaces, bookmarks, and disappearing posts can make harm look socially unconfirmed.
Three of four platforms involved no longer host the survivor's account; from the inside of any one platform, the case can be made to look like nothing happened.
Forms of active hostility
The power to save, quote, frame, and withdraw other people's words while treating the target's own preservation practices as obsessive or abusive.
The same screenshot mechanics used to circulate one annotated character-voice line as a 'death threat' in 14 different framings were, in the opposite direction, used to label the target's contemporaneous documentation as 'stalking' and 'obsession'.
Not the absence of witness infrastructure but its active inversion: the interpretive default by which a target's self-protective behavior is read as attack, her documentation as offense, her fiction or essay as evidence of pre-positioning rather than evidence of stance. The platforms and audiences that did exist read against her by default.
AO3's Section II.H states `Harassment in response to other harassment is not allowed` — a clause that pre-loads the platform's interpretive default. Banning the target's documentation required no platform-side credence of any specific accusation circulated by the harasser network; only the rule had to apply.
A coordinated harassment campaign inside a fan community produced a pattern that no parameter inside the archive could correct. The community surrounding the canon was built on affect, identification, and shared object-attachment — not on civic accountability, evidentiary scrutiny, or the patience to read 60 pages of forensic documentation. These are different kinds of community, held together by different bonds.
This is a structural mismatch, not a moral failure. A fan space is not a civil society in miniature; asking affect-based bonds to perform civic-grade witness work — receiving evidence, weighing claims, registering coordinated harassment as a public event — asks them to carry weight they were never load-bearing for. The roughly 100 outlets contacted and fewer than five replies are the same observation made one zoom level out: civic-grade witness infrastructure exists somewhere, but it is not concentrated around the cultural products that gathered the original audience.
The method consequence is direct. If the witness infrastructure a case needs is structurally absent from the community in which the case occurs, the archive cannot wait for that infrastructure to arrive. It has to build durable forms — citations, timelines, concepts, public pages — that other readers, in other contexts, can pick up and carry. This page, the case study, and the analyzer are that work. The aim is not to shame the community that did not witness; it is to acknowledge what kinds of community can carry which kinds of truth, and to build the carrying device.
Documentation as offense
The harassment archive's most direct administrative outcome was not vindication. It was a 15-day AO3 ban issued against the target, followed by a ruling against the appeal. The records the target had preserved — screenshots, timeline entries, contemporaneous notes — were read by the platform not as evidence of the campaign against her but as her own offense against the campaigners. Preservation crossed a threshold: not stigmatized, but prosecuted. This is asymmetric evidentiary privilege at its limit. The right to save and frame other people's words is granted; the parallel right to save and frame the harassers' words is reclassified as harassment of them.
AO3's own Terms of Service, Section II.H, states: Harassment in response to other harassment is not allowed. This sentence pre-loads the platform's interpretive default — a target's documentation of attacks against her is, on its face, the same category of conduct as the attacks themselves. The ruling against this case did not require the platform to credit any of the attacker network's specific accusations — death threats against readers, anti-LGBTQ branding, plagiarism, sockpuppet alts — none of which appear in the moderator's actual citation. It only required the rule to apply. The structurally pre-loaded ToS clause does the silencing work that fabricated accusations could not.
"Harassment in response to other harassment is not allowed."
Bookmark substrate
The platform's hostile environment finding under II.H had a specific material substrate: AO3's bookmark feature — public commentary on a work-author's pages that the author cannot delete. Starting 2025-09-12, an Autumn-side account (noonesshadow, later renamed Tsaverna) used bookmarks on the target's works to attempt covert outreach, establishing the affordance that bookmarks-on-her-works read as routine fan behavior. From October onwards, LadyYomi, AYWS, vice, and other network accounts bookmarked the target's works in coordinated chorus — performing the register of concerned fan while accumulating commentary the target could not remove. The wall, sustained over six months, became by ruling time exactly the hostile environment the rule names. The target wrote a chapter of formal notice against the weaponization mid-case; that chapter was added to the harassment ban.
The same enforcement window holds the asymmetry's other half. The target filed nine separate abuse reports against the harasser network across summer 2025 — none received a case number, an investigation acknowledgment, or any communication during the case window. The first report against her own documentation, by contrast, produced a 15-day suspension within days. Sanctions against the harasser network did eventually follow — a parody work targeting the target was banned, two attacker accounts were banned, defamatory content within other harasser works was removed, and one prominent harasser's comments on the platform were deleted — but on a markedly delayed timeline that did not reach the target during the period the rulings against her were forming. The asymmetry is not that one side was prosecuted and the other was not; it is that the architecture is dramatically slow and unresponsive in one direction and fast and decisive in the other.
Nine abuse reports filed by the target against the harasser network.
One report filed against the target's own documentation.
Accusation cloud
10 circulated claims
0
overlap
Ruling basis
3 cited ToS sections
Circulated against the target
Cited in AO3's actual ruling
Zero overlap. The platform did not require credence of any specific accusation circulated by the harasser network; only the ToS clause had to apply on its face.
Founding lie
The accusation chain started from an event that did not happen — a claim that the target had appeared in the original poster's comment section to insult her and her character. The fabricated origin was never required to survive verification, because by the time anyone would have asked, it had already become ambient. Each subsequent grievance cited the founding event as established fact. In an adverse reading regime, the initial factual basis of an accusation does not need to exist; it only needs to be repeated until it can be cited.
Reading as inversion
Psychoanalytic Reflections on My Relationship with Kazama Chikage — an 18,000-word long-form character study and trauma autoethnography posted on AO3 well before the campaign reached full force — contained the writer's own naming of the attack mechanism then forming around her. The work satirizes the character-fundamentalist fans who had already attacked her on a Chinese platform as today's Red Guards in cosplay — anti-intellectual, yet convinced that you alone hold the keys to truth, and dissects the discursive structure by which any subject who is unclassified or complex in attitude gets shamed back into a recognizable framework. The campaign network demonstrably read this work: its critical voice was stripped of context and recycled as the foundation of the anti-LGBTQ and homophobic accusations against the writer — despite her own slash and queer fanwork output elsewhere on the same archive. Reading occurred; what occurred was reading-as-inversion. The text was not functioning as authorial speech but as raw material the hostile reading regime could re-shape into evidence against its author. The text cannot speak for itself; it requires a reading infrastructure willing to read it as speech, and in this case that infrastructure was actively hostile rather than absent.
"today's Red Guards in cosplay — anti-intellectual, yet convinced that you alone hold the keys to truth"
Distributed complicity
Three forms of participation enabled the campaign. Co-attackers contributed directly; their corpora are the materials the analyzer counts. Platforms chose which testimony to credit — and the choice to reread the target's documentation as harassment was an active editorial decision, not a procedural error. Bystanders, exposed only to the ambient narrative and never to the source materials, ratified the result by not reading. None of these positions is innocent. In a hostile reading regime there is no such thing as a passive bystander, because reading is active labor and the choice not to read is the choice to outsource judgment to whoever is producing the ambient noise.
Adverse reading regime
These four observations together name a condition the previous section did not fully address. Witness infrastructure was not merely absent — the reading regime around the case was actively adverse. Self-protective behavior was read as attack. Documentation was read as offense. The work that named the witch-hunt was read as evidence of pre-positioning rather than evidence of stance. Absence (no civic witness) and active presence (hostile default reading) are paired conditions, not the same one twice. A method built only against absence does not survive the second condition. It has to assume that any record produced will, in some quarters, be read against its author — and that durable forms must be designed to remain interpretable even when the surrounding interpretation is hostile.
Record held
The AO3 case report remains readable here after takedown: every chapter, note, and structural claim kept in public form.
Platform exception
The suspended account was restored five days after deletion, the only platform restoration in the case record.
Tool survived
The syntax-first monitor remains inspectable, rerunnable, and able to carry claims beyond any single post or platform.
Reading produced
Later forensic passes surfaced the 9-vs-1 asymmetry, the rhetoric-vs-ruling gap, and the ToS clause's adverse default.
The harasser network did not emerge from the case unchanged. The ladyyomiart Tumblr account self-deactivated on 2025-08-18; the replacement attack blog's original production collapsed after December 2025 and now emits content only reactively, within hours of the target posting elsewhere. The campaign's ability to set the ambient narrative independently of the target's own visibility — once its central capacity — is gone.
The page being read is itself an instance of the method working. Each paragraph here points at a verifiable source file in the archive; each named concept can be checked against a specific corpus or document. Forensic correlation passes by reading agents — including the ones that wrote and then corrected this page — have surfaced findings the original case window did not have time to derive: the 9-vs-1 platform-enforcement asymmetry, the gap between attacker rhetoric and AO3's cited ToS sections, the specific clause that pre-loads adverse reading. These findings are now durable, citable, retellable. Reading occurred. Reading produced yield.
None of this is vindication. The platforms that mattered most to the case (AO3, Bluesky, Medium) remain hostile or absent. The witness infrastructure that would have made this analysis matter at the time it was needed did not exist. But the forms held, the record survived, and the analysis can be cited. Yield is not victory; yield is what gets carried forward when victory is not on offer.
The aim is not only to prove that something happened. It is to build forms through which the truth can be read, verified, received, repeated, searched, cited, and remembered after the immediate social scene has moved on or pretended nothing happened.
This page is a frame for rereading the materials linked below, not a replacement for them. The terms here are accountable to those materials: any reader can check whether each claim survives contact with the source files.
Public knowledge is not a property a record has by itself. It is an achieved relation among record, claim, reader, witness, and circulation.
Public knowledge is an achieved relation.
These references locate the method's inheritance: each tradition supplies a specific concept the case extends or specifies. The connections are tight — what the case made visible, the tradition had a name for; what the tradition names, the case shows in operation. The page is not borrowing legitimacy from theory; it is reading the case and the theory through one another.
Shapin and Schaffer argued that what counts as a fact is settled by social arrangements, not by the truth of a position alone: Boyle's experimental knowledge became authoritative through virtual witnessing, gentlemanly replication, and the Royal Society as institution. The case here is a small-scale version. The 60-page forensic brief is Boyle-equivalent apparatus; the platform's II.H ruling is the equivalent of declaring `this method does not count as fact-making`. The dispute is not whether the harassment occurred — records exist, attackers admit much of it — but whether the target's documentation method is recognized as evidence-producing. Solutions to the problem of knowledge are solutions to the problem of social order; in this case, the social order delivered the platform's solution, not the target's.
Berger and Luckmann describe how repeated actions become typified, then sedimented into shared memory, then institutionalized as objective facts about the world — a cycle that runs regardless of any single act's truth. The case shows two competing sedimentation chains. The network's claim `ygpgsgl is autumn's alt` started as a single LadyYomi assertion, was typified through chorus repetition, sedimented into ambient cross-platform knowledge, and reached platform staff as a reasonable starting assumption. The target's counter-claim produced more documentation but lacked the network and institutional pathways for the same sedimentation cycle. Public knowledge does not weigh evidence; it weighs sedimentation throughput.
Fricker distinguishes testimonial injustice (a speaker is unfairly believed less because of identity prejudice) from hermeneutical injustice (a subject lacks the public vocabulary to make her social experience legible). Both are present. The target's 9 abuse reports went unprocessed while the harassers' single report received instant action — a credibility allocation that tracks identity-of-access and register, not evidence quality. More structurally, the case lacked ready public vocabulary for `coordinated cross-platform harassment from a fan circle that performs the register of concerned fan while accumulating an undeletable bookmark substrate that platforms can later read as II.H hostile environment`. Each clause of that sentence had to be coined or sharpened on this page. Fricker's framework predicts the case; the case in turn names a third mode Fricker's typology does not yet contain. Where testimonial injustice is credibility deficit and hermeneutical injustice is vocabulary absence, `adverse reading regime` (named earlier on this page) is interpretive inversion: speech is heard, vocabulary exists, credibility is not the issue — the speech is read in the opposite direction of its content. Self-protective behavior is read as attack; documentation as offense; the writer's own essay critiquing witch-hunts is read as evidence of witch-hunt complicity. Adverse reading is a distinct epistemic injustice mode — not deficit, not absence, but inversion — and the case is an instance available for theoretical pickup.
Star's structural insight: infrastructure is invisible when it works and visible upon breakdown. `Witness infrastructure` enters this page's vocabulary because the case made the absent infrastructure visible — the fan-circle social trust that should have carried the message outward, the platform abuse pathway that should have processed reports, the civic-media network that should have picked up the story. Each `should have` became namable retrospectively because each broke. Bowker and Star's second move: classifications embedded in infrastructure produce outcomes that hold regardless of evidence. AO3 II.H pre-classifies `harassment in response to harassment` into the same category as harassment itself; the target experiences classification torque — her conduct slotted into the wrong category, her own situation lacking a category that would protect it. Bowker and Star treat classification torque as bureaucratic byproduct: tragic, structural, but not designed against the misclassified. The case extends the framework. Hostile actors here read the ToS, located II.H, and `strategically built the bookmark substrate knowing II.H would later convert it into evidence of hostile environment`. Classification torque was not an accident; it was weaponized. Adversarial classification weaponization — pre-loaded categories deliberately exploited by adversaries to convert a target's defensive conduct into platform-actionable harm — is a B&S-tradition concept the case names but the original framework, in its bureaucratic-systems frame, did not yet articulate.
Kate Manne, Sara Ahmed, and Lauren Berlant describe overlapping mechanisms by which women's persistence against harm gets converted into evidence of their own disorder. Manne names the reframing of women-who-insist as `unhinged` / `shrill` / `unable to let go`. Ahmed names the feminist killjoy: the one who, by continuing to name a problem, becomes the problem disturbing collective happiness. Berlant names the affective discipline by which `move on` constructs persistent grievance as attachment to suffering rather than response to ongoing harm. The case condenses these three lineages into a single rhetorical move the harasser network deployed repeatedly: `you reap what you sow`. The proverb form is the payload — folk-wisdom authority makes the dismissal sound obvious without engaging the substance — and the case's specific contribution is naming what the proverb form does: it functions as projected-forward adverse reading. Once deployed, every future utterance from the target reads as `she's still not accepting her lesson`, regardless of whether the original alleged sow was fabricated (founding lie) or trivial (writing fanfic, using AI). Adverse reading regime has two faces: a spatial one (past speech read in reverse, developed earlier on this page) and a temporal one (future speech pre-discounted by proverb-form before utterance). Spatial inversion answers `how did documentation become the offense?`; temporal projection answers `why did persistence become evidence of disproportion?`. The case names both faces and demonstrates their joint operation.
This page does not offer a universal taxonomy. It names patterns this archive can verify, and leaves their portability to other cases open.
This page reads two specific projects in a single archive. The patterns named here are derived from one closed case across a particular configuration of platforms, languages, and community types. Whether adverse reading regime or founding lie or asymmetric evidentiary privilege describe other cases is an open question for those cases' own forensic work — not a claim made here.
The narrator of this page is also the target of the case the page reads. That dual role shapes what is visible from inside and what may not be. An external reviewer would have seen different things; some claims that read as forensic finding from inside may read as motivated interpretation from outside. The materials linked below let any reader make their own corrections.
The terms coined or borrowed here — adverse reading regime, witness infrastructure, platformed amnesia, asymmetric evidentiary privilege, founding lie, bookmark substrate, adversarial classification weaponization — are after-the-fact names for patterns the case made legible. They are not pre-existing taxonomies validated across corpora. They are useful inside this archive; whether they hold elsewhere is a question for elsewhere.
The forensic instruments referenced — regex categorization, stylometry comparison, time-zone fingerprinting, ask-graph mapping — are precision-recall trade-offs, not exact measurements. The analyzer over-counts and under-counts in known directions. The numbers cited (1,412 posts, 5,700 archived posts across five attacker corpora, the 14× decontextualization of the Okita-voice line, 9 vs 1 reports) come from corpora the reader can re-check; the interpretations attached to those numbers can be contested without contesting the numbers themselves.
A suggested path through the source material, from overview to raw record and back to method.
The closed case study (2024–2026) this method note rereads. 50-chapter report, 51-event timeline, network graph, evidence archive.
217 source files across six folders. This is the layer everything else is shaped from.
The 50-chapter case documentation in full, ~33,000 words. Originally posted on AO3 as work 67181098; preserved here after the takedown. The single most comprehensive primary source for everything described above.
Forensic analyzer that runs syntax-first counts across five persona corpora — a working example of evidence shaped into a public, inspectable form.