The Rope and the Record

The trend, in the points the data state plainly

The 26% undetermined share is the report’s “misclassification” signal, read through the peer-reviewed literature as Black suicides hidden as undetermined — not homicides hidden as suicides. Peer-reviewed reading And on autopsies: almost every homicide- or undetermined-coded hanging is autopsied, yet nearly half of Black hanging suicides were ruled without one. High confidence

Click a column header to sort. supp. = CDC suppresses counts of 1–9 deaths; the report never estimates them. Undetermined share = share of all hanging deaths (all races) left undetermined, Appendix A.2; “—” where the appendix does not state one. By system type, undetermined shares run: State ME 0.68% · County ME 1.19% · Coroner 1.11% · Mixed 1.0%. Alaska (3.5%), Arizona (2.6%) and Mississippi (2.5%) leave the largest shares undetermined — and high-undetermined states include every system type.

Extremes worth knowing

Context from the same chapter: in custody, hanging is how ~90% of jail suicides happen — yet the Black jail suicide rate (16 per 100k inmates) is the lowest of any group, against a White rate of 86; 71% of jail suicides were White inmates (BJS, Mortality in Local Jails 2000–19). State-level custody hanging data by race is not published — a data gap the report flags.

Both can be true at once: a system that under-investigates Black deaths can hide suicides and miss homicides.

What the studies found

All seven are peer-reviewed. Peer-reviewed

Coroner vs medical examiner

Centralized / state ME systems leave fewer hanging deaths undetermined (0.68%) than coroner or county systems (1.0–1.2%): County ME 1.19%, Coroner 1.11%, Mixed 1.0%. Coroner-only states under-report suicides by 17.4% and perform 20.4% fewer autopsies (Fernández & Jayawardhana 2024). The report’s own caveat: “Correlation, not proof.”

Mississippi is one of 13 county-coroner states. Its coroner qualifications: age 21+, a high-school diploma or equivalent, and a state death-investigation course (generally 40 hours) before taking office; coroners are elected (Miss. Code §19-21-103, -105). The State ME performs autopsies on referral.

The logic test — what a hidden-homicide pattern would look like

If homicides were being systematically certified as suicides, the national data should show a recognizable shape. It shows the opposite shape.

Bottom line: national data do not show a large hidden-homicide pattern — and they cannot rule out a small number of individual cases, because a homicide certified as suicide is invisible in the statistics.

Also in Mississippi: Rankin County’s “Goon Squad” — deputies who tortured two Black men in January 2023; six were sentenced in 2024 after their group chat was exposed on May 29, 2024. High confidence

The flyers — and the timeline reading

KKK flyer drops are documented in 2025–26 in Tennessee (Jan 2025, >2,000 flyers), Baltimore (Jul 17, 2026), Indiana (Jul 24, 2026), Michigan (Aug 2026), Columbus, Ohio (Sept 28–29, 2026) and Louisiana (Oct 11, 2026, Unverified). No 2025–2026 KKK flyer reports in Mississippi were found. Flyers are documented as intimidation; none has been linked to any hanging.

The report’s timeline verdict Medium confidence: what lines up is a climate — a run of reported hangings (June–Sept 2026), the Senatobia texts, flyers in four other states, a 60-member congressional letter (Sept 8–10). What doesn’t line up is place: the hanging cases spread over NC, FL, MI, MS and MD; the flyers over MD, IN, MI, OH and LA — no overlap; and the only Mississippi hanging in the cluster (Fortune) was charged as a drug-related staged homicide. The timeline shows a climate — documented police racism and organized intimidation — that explains why families distrust rulings; it does not show a causal link between any flyer, any officer and any death.

The Sea Tow call — what a single phone call can and cannot establish

A teen on the friends’ boat called a private towing company; MDMR released the audio ~July 25, 2026. Primary audio Three things about it remain disputed:

• The words: asked “Is everyone in good health?”, the answer is transcribed as “Yeah, everyone is on board” (WLOX/Fox) vs “Yeah, everybody’s [inaudible]” (independent channel analysis) — disputed.
• The time: 3:48 pm (file label) / ~4:00 pm (WLOX) vs 4:36 pm (family examiner) / 4:48 pm (friends’ counsel exhibit) — unresolved; four times.
• The location: the caller gives coordinates — “not aground, ~31 ft” — matching an anchor point, yet GPS shows the boat moving from 4:31 pm — conflict, unexplained. Nolan is not identifiably heard on the released audio.

The report’s reading Medium confidence: the call establishes that a call happened and what was asked; it does not establish whether Nolan was aboard, and it does not show a crime.

What the AI process concluded — and what it did not

The AI jury voted 15–1 to reopen rather than let the record stand; the tiers ran Reopen → Reopen (Tier 2 affirmed 3–0) → Reopen. At Tier 3, 4 of 5 models found “no evidence” of a hate group (1 said “insufficient”); 3 of 5 marked “possible violation — refer” on civil rights. The report’s gloss: the AI consensus is a process verdict — the record warrants an independent federal review and release of the ranger report, photos and full timeline. It is not a finding that anyone killed Nolan Wells. AI simulation

No official body, court or credible outlet has found a cover-up or a racial motive. The family’s attorney: “we don’t know if it was racist or not.” Three friends have sued six online creators and Terrell Owens for defamation; at least four people were charged over threats. Jackson County context: pop. 145,249 (18.9% Black, 66.1% White); EJI documents 8 lynchings, 1877–1950; death investigation by elected county coroner, State ME autopsy.

The pattern underneath: cases from elected-coroner jurisdictions ended “not stands” 71% of the time, against 41% in medical-examiner jurisdictions. The recurring problems were self-review (the same local system investigating and presenting its own case), unreleased records, fast closures, and no full autopsy. The report’s caveat: small numbers, non-random selection (disputed cases by definition), AI verdicts — this shows where scrutiny is warranted, not a rate.

The cross-case pattern, grouped

Coroner and ME rows cover 39 of the 42 cases, as grouped in the report. On the report’s third question: at Tier 3, the hate-group finding was “no evidence” for every case reviewed, except C31 and C33, where it was “insufficient evidence.” No case produced evidence of hate-group involvement at any tier.

Eight near-misses were excluded with reasons, including Titus Tucker (no news or official record found), Kindra Chapman (family publicly accepted suicide), Cory Zukatis (no documented dispute of the ruling) and Robert Elliott (a homicide investigation, not a disputed ruling). Three roster cases — Torrance Medley, Dominique Alexander and Uche Okafor — surfaced from the social-media pulse and were verified in major outlets before inclusion.

What the juries believed, in probabilities

Jurors gave probabilities for three propositions; these are the medians, reported per docket — not any single model’s verdict.

Jurors also judged the defense the more faithful advocate — 10 of 16 on docket 1 and 10 of 13 on docket 2.

Tier 1: the judge’s verdict

On Q1 (Gemini 3.1 Pro): no data support a “meaningful rate” of Black hanging homicides misclassified nationally; the literature finds the inverse error — but local systems have structural deficiencies and conflicts of interest that make several individual rulings unreliable. On Q2: unanimously unproven.

Tier-1 highlights: unanimous “ruling stands” for 8 docket-1 cases; a 15–1 “homicide” vote for C16 (Fortune, already charged); C1 and C4 drew the only homicide minorities among disputed hangings. On docket 2: unanimous “stands” for Sandra Bland, Raynard Johnson and Feraris Golden; 12–1 “unlawful killing” for Elijah McClain and 11 votes for Vanessa Guillén — both matching official or court findings; Birchmore split 7 homicide / 6 reopen.

The appeals that moved cases

Tier 3: the final findings

• Hate groups: no evidence in any case on either docket. A minority said “insufficient evidence” (not “no evidence”) for C1, C4, C31 and C33. Median chance any case involved a hate group: 5% (docket 1), 3% (docket 2).
• Patterns: a misclassification pattern was judged unlikely — median 18% / 15%. A deficient-investigation pattern was judged likely — median 72% / 72%. “Pattern-or-practice” was marked found by 3 of 5 models on docket 1 and 4 of 5 on docket 2 (Claude Opus 5.5 and Grok 4.7 dissented: no single jurisdiction in the record shows multiple deficient death investigations).
• States named: Mississippi was named a state of concern by every Tier-3 model; then NC, AL, TX, GA (Texas, Georgia, Montana, Pennsylvania and Colorado appear on docket 2).
• Civil rights — “possible violation, refer” majorities: C20 Birchmore, C21 Nolan Wells, C32 Kendrick Johnson, C35 McClain, C37 Guillén, C39 Christian Hall. Most other flagged cases were marked investigative deficiency.
• Final aggregate: of 42 cases, the last tier to review each left 19 stands · 20 reopen · 3 homicide/unlawful killing (Fortune, McClain, Guillén — all matching charges or court findings). 19 + 20 + 3 = 42; no cases remain.

How the votes moved

Probabilities, Rounds 1 → 2: the median chance of a meaningful hidden-homicide share stayed at 10%, but the range narrowed from 2–20% to 5–12%. The median chance of hate-group involvement fell from 3% to 2%. In Round 4, 17 of 27 models said the evidence rule changed their view (mostly on C21 and trust in local rulings). The key self-correction: the CDC figures are produced by the same certifiers whose accuracy is in question, so national data cannot fully referee the dispute.

Round-3 consensus

• Death certification for hanging deaths differs by race: Black deaths ~5× as likely coded homicide, ~3× undetermined; the literature shows Black suicides under-counted. (~100% agree on direction.) High confidence
• The documented bias mostly runs the other way: suicides coded undetermined or accidental. No study estimates hidden hanging homicides by race. High confidence
• Hidden homicides: “Weak” — possible in isolated cases (C8, C10, C12, C18 carry red flags; C16 and Birchmore show staging happens), not shown as a pattern. Black hangings are autopsied more; Black rates are lowest in the Deep South. (28/28.) Medium confidence
• Certificate data cannot detect a homicide recorded as suicide, so neither “no evidence” nor “disproven” is available for Claim B. Medium confidence
• No evidence links organized hate groups to any listed case; the Senatobia texts and 2025–26 flyers overlap with the hangings only in time, not place or content. (28/28.) High confidence

Named dissents: 13 models (including Claude Sonnet 5.5, GLM-5.3, DeepSeek V4 Pro, Grok 4.7 and Gemini) would keep Claim A at Moderate, not Strong. Nemotron’s Baltimore-flyer point was rejected by the panel; Gemini’s p(B) = 5% argument was rejected by the others. Dissents are recorded — not averaged away.

A retrieval control: Perplexity’s Sonar models ran through OpenRouter in academic mode (search_mode: "academic", scholarly sources only), with Sonar Pro in normal web mode as control. Academic mode returned 21 citations, 16 (76%) from journals/PubMed; web mode 15 citations, 10 (67%). Impact on conclusions: small — the gain was sourcing quality, not direction.

“US death records do treat Black hanging deaths differently: they are more often labeled homicide or ‘undetermined’ than White ones. The best research suggests the main error is Black suicides being undercounted, not murders being disguised as suicides. A few cases have real unanswered questions, and staged hangings do happen, but no case has been officially reclassified and no study shows a hidden pattern. Because a murder recorded as suicide would be invisible in national statistics, the panel cannot rule it out and puts the chance of a meaningful hidden share at roughly 10%. There is no evidence that organized hate groups were behind any of the listed deaths.”

Who sat on the panel: Anthropic (Claude Sonnet 5.5 thinking, Claude Opus 5.5) · OpenAI (GPT-5.5, o3, gpt-oss-120b) · Google (Gemini 3.1 Pro, Gemma 4 31B thinking) · xAI (Grok 4.7) · Perplexity (Sonar Deep Research academic, Sonar Reasoning Pro academic, Sonar Pro web) · DeepSeek (V4.1 Flash 1017 thinking, V4 Pro thinking) · Alibaba Qwen (3.7 Max thinking, 3.5 397B thinking) · Moonshot Kimi K3 · Zhipu GLM-5.3 (thinking) · MiniMax M3 (thinking) · Mistral (Large, Medium 3.1, Small 4) · Meta Llama 4 Maverick · NVIDIA Nemotron 3 Ultra (thinking) · Nous Research Hermes 4 405B (thinking) · Xiaomi MiMo v2.6 Pro · Baidu ERNIE 5.1 (thinking) · Cohere Command A Reasoning · Amazon Nova Pro · StepFun Step 3.7 Flash (thinking) · Meituan LongCat 2.0 (thinking).

Per-case attention on TikTok Medium confidence: Elijah McClain 28.3M views of relevant posts, Kendrick Johnson 11.1M, Vanessa Guillén 10.3M lead the roster — while 11 names, including Otis Byrd, Dennoriss Richardson and Roy Veal, returned no relevant posts at all. The loudest accounts spanned both stances: alarm-raisers (@msrachelforlittles, 12.0M followers; @nowthisimpact, 6.67M; Nancy Grace) and skeptics (ex-CSI analyst @amy_santoro; commentator Officer Tatum, 1.01M views on YouTube).

Claims vs evidence

The claims that travelled furthest, checked against the record — with the report’s own verification status on each.

The pulse also worked as a reporting tool: three roster cases (Medley, Alexander, Okafor) surfaced from social media and were verified in major outlets before inclusion, while the loudest misattribution — the “DC police station lynching” — was traced to a real death, wrongly described, and excluded.

1. DEEP PAST

   Violence between groups needs no color line

   Chimpanzees kill across community lines — Wilson et al. 2014 (Nature) compiled 152 killings in 18 communities over five decades [V]; Goodall documented Gombe’s “Four-Year War,” 1974–78 [C]. Chimps target group membership — territory, familiarity — not appearance; the Gombe victims were former companions [C]. Humans favor their in-group even on trivial labels (Tajfel’s minimal-group studies) [C]; infants’ own-race face preference reflects exposure, not innate hostility [C].

2. ≥13,400 YEARS AGO

   The earliest human intergroup violence

   Jebel Sahaba, Sudan — repeated episodes of intergroup violence (Crevecoeur et al. 2021, Scientific Reports) [V]. Nataruk, Kenya, ~10,000 years ago: 27 individuals (Lahr et al. 2016, Nature) [V; the “warfare” reading is debated, S].

3. ANCIENT WORLD

   Difference noticed — hierarchy not built on it

   Greeks and Romans noticed skin color but did not build a color hierarchy of human worth (Snowden, Blacks in Antiquity 1970; Before Color Prejudice 1983) [C, with debate]. Isaac (2004) finds Greco-Roman “proto-racism” — hereditary stereotypes about peoples, Aristotle’s “natural slaves” — but not color-coded [C]. “Barbarian” meant non-Greek speaker; China’s Hua–Yi line was cultural; ancient slavery came from war, debt and piracy, not race [C].

4. 1449 – 1705

   Race is invented — in law

   1449: Toledo’s Sentencia-Estatuto barred descendants of converted Jews from office — limpieza de sangre made lineage a heritable stigma [C]. The Atlantic slave trade embarked ~12.5 million Africans; ~10.7 million survived the crossing (SlaveVoyages) [V]. Virginia law followed: 1662 (a child’s status follows the mother), 1667 (baptism doesn’t free), 1691 (interracial marriage banned), 1705 (slave codes) [V] — a hardening historians link to elite responses after Bacon’s Rebellion, 1676 (Morgan 1975) [C, interpretive]. Linnaeus (1758) split humans into four varieties with temperament stereotypes; Blumenbach (1775/1795) proposed five and coined “Caucasian” [V].

5. 1883 – 1950

   Science and law harden it

   Galton coins “eugenics” (1883) [V]; Plessy v. Ferguson (1896) [V]; the 1924 immigration quotas and Virginia’s Racial Integrity Act [V]; Buck v. Bell upholds forced sterilization (1927) [V]; the Nazi Nuremberg Laws (1935) drew partly on US race law (Whitman 2017) [C, interpretive]; apartheid’s Population Registration Act (1950) [V]. Tuskegee recorded 4,743 lynchings 1882–1968, 3,446 of Black victims [V]; EJI documents 4,400+ racial-terror lynchings of Black people, 1877–1950 [V]. Recorded lynchings peaked in 1892 — 230, of whom 161 Black — and fell to near zero by the 1950s; Tuskegee counts are a floor.

6. 1954 – 2022

   Dismantling — and what genetics says

   Brown (1954), the Civil Rights Act (1964), Voting Rights Act (1965), Loving (1967), Fair Housing Act (1968); apartheid ends in 1994 [V]. Lewontin (1972): ~85% of human genetic variation lies within populations [V]; Rosenberg et al. (2002, Science): 93–95% [V]. The AABA Statement on Race and Racism (2019): race does not accurately describe human biological variation; it is a social construct with real effects [V]. The 2013 Shelby County ruling ended VRA preclearance; the 2022 Emmett Till Antilynching Act made lynching a federal hate crime. High confidence

The panel’s median cause weights for the hierarchy’s origins (ranges in the report): economics of slavery & labor 30 · colonialism 20 · law 15 · political elites 10 · scientific racism 9 · evolutionary in-group psychology 5 · religion 5. Caveat: the answer template’s example period may have anchored some models’ wording; the cause weights were unprompted. AI simulation

Progress, measured: approval of interracial (Black–White) marriage rose 4% (1958) → 94% (2021); the Black–White segregation index fell 72.7 (1980) → 55.2 (2020); intermarriage among newlyweds rose 3% (1967) → 17% (2015). High confidence Hate crimes moved the other way recently: anti-Black incidents jumped to 2,971 in 2020 and peaked at 3,421 in 2022; the BJS victim survey finds roughly a quarter-million hate victimizations a year, most never reported to police. Medium confidence Of the panel, 21 of 27 scored models said the post-2020 rise is “partly” real and 6 said “yes”; none called it purely a reporting artifact.

What works — the fixes, graded by evidence

The grade is the report’s, based on the evidence — not the panel’s vote, which is shown alongside.

“To an observer, sorting the avatars by color looks both tragic and arbitrary. But the panel’s answer is that the sorting was never really about color: it was built to divide labor, land and power, and it is maintained the same way. So the fix is unglamorous — investigate every disputed death in the open, by people with no stake in the answer; close the economic and housing gaps that keep the groups apart; teach the history straight; and build AI that audits power instead of automating it.”

What would settle it — six steps