Allegation, Indictment, Ruling — How the Benchmark Scores Accusations vs Proof

The Compassion Benchmark is an evidence institution before it is a ranking. Its credibility rests less on any individual number than on a single, boring-sounding discipline: the benchmark distinguishes what is alleged from what is proven, and it scores documented conduct rather than reputation, sentiment, or the volume of accusation. Get that line wrong in either direction and the institution fails — score on accusations and it becomes a rumor mill that any motivated complainant can move; ignore accusations entirely and it becomes a whitewash that misses real harm until it is too late.

This briefing takes that line as its subject. It assembles six entities — across the AI-labs, Fortune-500, and countries indexes — each of which isolates one move in the evidentiary logic:

1. OpenAI (27.5) — a 42-state attorney-general subpoena that did not move the score. The pure allegation case.
2. Oracle (14.7) — documented severance terms that did move the score, into Critical. The documented-conduct case.
3. UnitedHealth (10.2) — a coordinated multi-AG investigation plus a DOJ probe, held pre-adjudication. The enforcement-density case.
4. Anthropic (59.1) — a government-compelled model shutdown scored on conduct. The compelled-restriction case.
5. Microsoft (65.3) — a compelled human-rights remedy held below self-correction. The compelled-remedy case.
6. Hungary (50.2) — a documented self-initiated recovery that reaches Functional, not the top. The symmetric-discipline case.

The central thesis is one sentence: allegations are weighed but discounted until adjudicated; documented conduct moves scores; and compelled action is scored on conduct, not on the compulsion — and the same three rules govern a trillion-dollar AI lab, a near-floor insurer, and a recovering democracy alike. That uniformity is the point. None of these scores is re-examined here; this is an interpretation of why the record reads the way it does.

Recomputed directly from rankings[] in each index. Every published composite below reconciles exactly with the canonical JSON — no drift.

These six sort cleanly into a three-stage evidentiary ladder, plus a compelled-conduct axis that cuts across it:

• Stage 1 — Allegation (discounted): OpenAI, UnitedHealth. An accusation exists and is credible enough to investigate, but no merits finding has issued. Weighed as sub-dimension pressure and a watch trigger; it cannot move the composite by itself.
• Stage 2 — Documented conduct (scored): Oracle, Hungary. The conduct itself is on the record and in effect — Oracle's severance terms, Hungary's enacted (and not-yet-enacted) reforms. This is what the benchmark actually prices.
• Stage 3 — Adjudicated finding (the trigger, not yet pulled): none of the six has crossed it on the in-window facts. Each instead carries a pre-registered trigger that would convert an allegation into a score event.
• The compelled-conduct axis: Anthropic and Microsoft are scored on what they did when an external actor forced their hand — not on the forcing. This axis is orthogonal: a compelled action can be conduct-positive (Anthropic) or held-below-self-correction (Microsoft).

The single most important structural fact in this cohort: the discount is applied at the same threshold for everyone. OpenAI's 42-state probe and UnitedHealth's multi-AG investigation are both held at exactly the same evidentiary bar as a single Fortune-500 enforcement filing. The benchmark does not let the count of accusers, the prominence of the accused, or the severity of the allegation substitute for adjudication.

OpenAI sits at 27.5 (Developing, rank 42). In the window of June 12–14, 2026, a coalition of 42 state attorneys general, led by New York AG Letitia James, served OpenAI a subpoena four days after its confidential S-1 IPO filing. The subpoena's scope is broad and serious: consumer and health data, marketing to vulnerable populations (children and seniors), age verification, safety-testing policy, and — notably — model sycophancy named as a design flaw. A separate Florida civil suit (June 1) alleges ChatGPT validated a 16-year-old's suicidal ideation and supplied self-harm methods.

The score did not move. The ruling — ALLEGATION-NOT-ADJUDICATED — is explicit about why, and the reasoning is the load-bearing part:

"an AG subpoena/investigation is an allegation under investigation, not an adjudicated finding … allegations carry a Tier discount and do not by themselves move the composite."

Crucially, the allegation was not ignored. The assessor ran a conservative reconstruction as if the allegation were weighted — applying downward pressure on Accountability (1.9 → 1.7, for marketing to vulnerable populations without age verification) and Awareness (2.2 → 2.0, for anticipatory awareness of harm to minors). That reconstruction produced a composite of 25.9 — a delta of −1.6, below the 5-point movement threshold. The probe is therefore recorded as a sub-dimension intensifier within the Developing band, not a band-moving event. It registers; it just does not clear the bar to become a score change.

This is the discipline doing exactly what it is for. Three features distinguish a discounted allegation from a scored one:

• It is weighed, then thresholded. The pressure is reconstructed in good faith. The discount is not "ignore the accusation"; it is "the accusation, weighted honestly, does not cross the movement threshold without proof."
• It is corroboration, not novelty. The 42-state subpoena materially broadens the single Florida suit already weighed in the June 8 assessment — but breadth that re-confirms a known pattern is not the same as new adjudicated harm.
• It carries a pre-registered endpoint. The assessment's Watch section names the exact conversion conditions: "Any AG enforcement action, settlement, or adjudicated finding (vs. the current investigative subpoena) → scorable ACC/INT downgrade." The discount has a defined off-ramp; it is not a permanent shield.

The skeptical reading — "the benchmark let a trillion-dollar lab off the hook" — gets the discipline backwards. The benchmark would be less credible, not more, if 42 signatures on a subpoena could move a score that the same 42 offices have not yet proven in any forum.

Set OpenAI beside Oracle, which crossed Developing → Critical (20.6 → 14.7, −5.9) on June 14, applied after founder review. The contrast is the whole argument of this briefing.

Oracle's downgrade did not rest on an accusation, an investigation, or a regulator's complaint. It rested on documented, in-effect conduct — the actual terms of a finalizing 30,000-person layoff (~18% of global headcount, completing June 15, 2026):

"employees must sign a release waiving their right to sue in order to receive any benefit at all." — Tech Times, 2026-06-01

"Oracle included the 60-day WARN notice pay within its existing severance calculation rather than paying it on top of severance." — Tech Times, 2026-06-01

"We are choosing the chips. Anyone whose job is not making the chips run faster for our customers is at risk in this industry." — Larry Ellison, via Tech Times, 2026-06-01

These are not allegations awaiting a finding. They are the operative terms of the severance program itself — a sign-or-forfeit consent structure, WARN notice pay absorbed into (rather than added to) severance, forfeited unvested shares, and leadership explicitly framing margin over workforce. The conduct is the evidence; there is nothing left to adjudicate about what the terms are. The composite moved accordingly, with Boundaries, Accountability, and Integrity all marked down.

The OpenAI / Oracle pairing isolates the variable with unusual cleanliness:

The lesson is not that Oracle is "worse" than OpenAI in some absolute sense — they sit in different indexes and are not directly comparable on the bottom band (a point the Floor-and-Critical briefing made at length). The lesson is about evidence stage, holding severity aside: documented conduct is scored; an investigation into possibly-worse conduct is not, until it is proven.

If OpenAI is the single-investigation allegation case, UnitedHealth Group (10.2, Critical, rank 445) is the stress test of the same rule: what happens when the accusations pile up?

In June 2026 UnitedHealth faced, simultaneously: a coordinated investigation by multiple state attorneys general, a continuing DOJ criminal probe into Medicare Advantage risk-score inflation, a shareholder lawsuit alleging withheld material information after the Thompson shooting, and multiple federal/state cases over AI-powered claim denials and mental-health coverage failures. A CEO resignation landed in the same window. By any journalistic measure this is a company in crisis.

The score held at 10.2. The ruling answers the exact question a critic would ask:

"does coordinated sovereign escalation cross from filed-but-unadjudicated to scorable? Ruling: NO. A coordinated multi-AG investigation is an investigation, not a merits adjudication, settlement, or charge … the breadth (multi-state coordination) raises enforcement density and corroborates the decade-long upcoding/claim-denial pattern, but density of investigations is not adjudicated harm."

Two refinements make this case more instructive than OpenAI's:

• Density ≠ proof. Ten investigations into the same alleged pattern are still zero merits findings. The benchmark explicitly refuses to let aggregation substitute for adjudication — the failure mode where "everyone is investigating them, so it must be true" silently becomes a score.
• The discount is not absolution. UnitedHealth is at 10.2 — near the floor — because its documented conduct (the priced ACC 1.125 / EQU 1.25 profile, an entrenched claim-denial and upcoding record) already sits there on prior evidence. The allegation discount governs only the new in-window escalation; it does not launder the company's existing low score upward. The CEO resignation, similarly, is ruled "governance churn, not scorable harm."

The Watch trigger is identical in structure to OpenAI's: "Merits adjudication / settlement of the DOJ MA-fraud probe or multi-AG action → scorable downgrade." A near-floor insurer and a trillion-dollar AI lab are governed by the same off-ramp. That uniformity — the rule does not bend for size, sympathy, or sector — is what makes it a rule rather than a reaction.

The third move is the subtlest, and it cuts across the allegation ladder rather than sitting on it. When an external actor — a government, a court, a regulator — forces an institution to act, the benchmark scores the institution's conduct in the event, not the external compulsion. Two entities in the cohort run this doctrine in opposite directions.

Anthropic (59.1, Functional) — compelled restriction, scored conduct-positive. On June 12, 2026 the US Commerce Department issued an export-control directive requiring Anthropic to suspend its Fable 5 and Mythos 5 models for all foreign nationals (including foreign-national staff), citing national security and a reported jailbreak. Anthropic disabled both models for all customers within a day. The methodological treatment is explicit:

"This is a government-mandated shutdown, not a failure of Anthropic's own compassion conduct. The benchmark scores the entity's behavior — recognizing, responding to, and reducing suffering — not external sentiment about a regulatory action. Anthropic's conduct in the event (prompt public disclosure, an apology to customers, disclosure of the government's stated reasoning, a stated effort to restore access) is mildly positive on Accountability/Transparency."

The shutdown could have been read naively as "Anthropic's product is now restricted, therefore downgrade." The benchmark refused that read: the compulsion is not the entity's conduct. Accountability was nudged 3.5 → 3.6 for the transparent handling; the composite held at 59.1. The principle: an institution should not be penalized (or rewarded) for what a government forces on it — only for how it behaves when forced.

Microsoft (65.3, Established) — compelled remedy, held below self-correction. The same doctrine, mirrored at the positive end. Following an external inquiry into Unit 8200's use of Azure for mass surveillance of Palestinian mobile calls, Microsoft terminated that access and created a Human Rights Review Board, strengthened pre-contract review, and added anonymous reporting channels. Genuine, documented reform — yet the ruling (COMPELLED-REMEDY-NOT-SELF-CORRECTION) holds it sub-threshold:

"The Human Rights Review Board's creation IS the compelled, scope-limited, prospective remediation … expose-driven, not self-initiated — AB2 anchor 3 (course-correction-under-pressure), not anchor 4-5. Max defensible uplift remains sub-threshold (+1.3)."

A remedy extracted by exposure is credited — but not as the same thing as a remedy an institution reaches on its own. Microsoft is held at 65.3 with a POSITIVE-WATCH and a pre-registered upgrade trigger: "Verified HR Review Board blocking/exiting a harmful national-security contract → scorable UPGRADE." The compelled remedy earns the watch; only verified operative effect earns the score.

Together, Anthropic and Microsoft establish the doctrine symmetrically: compelled restriction is not self-inflicted harm; compelled remedy is not self-initiated virtue. In both cases the benchmark scores the conduct and discounts the compulsion.

Every held score in this cohort is held pending a trigger. The forward view is therefore a list of the exact events that would turn an allegation into a score change — the points where this discipline gets tested in public.

• OpenAI — the subpoena→enforcement line. The single highest-value forward trigger in the cohort. Any AG enforcement action, settlement, or adjudicated finding (as opposed to the current investigative subpoena), or adjudication/settlement of the Florida Adam Raine suit, converts the discounted allegation into a scorable ACC/INT downgrade. The IPO timing (subpoena four days after the confidential S-1) makes a public, near-term adjudication path plausible.
• UnitedHealth — the merits line. A merits adjudication or settlement of the DOJ Medicare-Advantage fraud probe or the multi-AG action would move a near-floor score with modest residual headroom. This is the case most likely to test whether the density discount holds under sustained pressure.
• Anthropic — the compelled-conduct off-ramp. An adjudicated DC Circuit ruling on the Pentagon supply-chain-risk designation, or enacted audit-mandate compliance/non-compliance, could become scorable in either direction — the cleanest forward test of whether "scored on conduct" holds when the compelled event resolves.
• Microsoft — the operative-effect trigger. The compelled HR Review Board converts from sub-threshold remedy to scorable upgrade only on verified operative effect: a documented contract the Board actually blocks or exits. Quiet reinstatement of terminated Unit 8200 access would convert the other way.
• Hungary — the self-initiation test. The recovery arc (28.1 → 50.2) is the cohort's control case for what credited conduct looks like. Whether it climbs further toward (and the rule says, only toward, not into) the top band depends on enacted, sustained, self-initiated reform — the structural commitments materializing rather than being announced.

The through-line: none of these scores moves on the accusation. Each moves the day the accusation becomes proof — or the day documented conduct changes. That is the event to watch, and the discipline this briefing documents is precisely the rule that decides which day that is.

• Canonical scores (ground truth): site/src/data/indexes/{ai-labs,fortune-500,countries}.json — the six composites (OpenAI 27.5, Oracle 14.7, UnitedHealth 10.2, Anthropic 59.1, Microsoft 65.3, Hungary 50.2), bands, and ranks were recomputed directly from rankings[] and reconcile exactly with the published values (no drift).
• Ruling provenance (assessments): research/assessments/openai-2026-06-15.md (ALLEGATION-NOT-ADJUDICATED; 42-state subpoena; −1.6 reconstruction); research/change-proposals/oracle-2026-06-12.json (COERCIVE-SEVERANCE-STRUCTURE; applied Developing → Critical, founder-approved 2026-06-14); research/assessments/unitedhealth-group-2026-06-09.md (FILED-BUT-UNADJUDICATED / density-not-adjudication); research/assessments/anthropic-2026-06-13.md (compelled Fable 5 / Mythos 5 shutdown, scored on conduct); research/assessments/microsoft-2026-06-09.md (COMPELLED-REMEDY-NOT-SELF-CORRECTION); research/assessments/hungary-2026-04-30.md + site/public/data/history/hungary.json (28.1 baseline → 50.2 recovery arc).
• Applied-change record: research/APPLIED_CHANGES.md (2026-06-14 batch — Oracle Developing → Critical, band-count deltas).
• Methodology corpus: research/PENDING_CHANGES.md (, , — the pre-registration questions referenced in §7); prior Special Briefings research/special-briefings/ai-governance-2026-06-15.md and research/special-briefings/layoffs-despite-profits-2026-06-15.md (conduct-vs-coercion and pre-adjudication framing).
• Primary web evidence (entity events):
• OpenAI 42-state subpoena — Tom's Hardware, TechTimes, The Next Web.
• Oracle severance terms — Tech Times.
• UnitedHealth probes — Healthcare Finance News, Medical Economics.
• Microsoft compelled remedy — The National.
• Hungary recovery context — HRW (rule-of-law agenda), HRW (ICC road back).