The Floor and the Critical Band — How the Benchmark Judges the Worst

The Compassion Benchmark applies one 0–100 scale to every entity it assesses: a sovereign state, a Fortune 500 corporation, an AI lab, a robotics lab, a city. At the top of the scale this is unremarkable — "more compassionate institutional conduct" reads coherently across types. At the bottom it becomes a genuine tension. A state fails its population through the apparatus of governance; a corporation fails stakeholders within a commercial boundary; a lab fails through a product or a contract. These are not the same kind of failure, yet they are scored on the same axis and sorted into the same two bottom categories: the Critical band (composite ≤ 20) and, beneath it, the 0.0 floor.

This briefing takes the bottom of the scale as its subject. It asks three questions of the existing record:

1. What is the floor, mechanically and substantively — and is it coherent across the types that reach it?
2. Why does the largest index (Fortune 500) have 52 Critical entities but zero at the floor — is that a judgment or an artifact?
3. Can a number near the bottom be read across types — is UnitedHealth at 10.2 "as bad as" Turkey at 10.3?

The central thesis: the floor is a defensible, principled state — but the single 0–100 scale invites a cross-type severity read at the bottom that the methodology cannot actually support, and one entire entity class (corporations) is exempted from the floor by a mechanic rather than by a stated rule. Both are interpretive features worth making explicit; neither is a reason to re-score anything.

Recomputed directly from rankings[] in each index (Critical = composite ≤ 20; floor = composite ≤ 0.05). The pre-supplied snapshot is confirmed exactly — no drift.

The single most important structural fact: every one of the 23 floor entities carries an identical dimension vector — all eight dimensions at the minimum anchor ([1/1/1/1/1/1/1/1]). The floor is not "a very low score." It is the simultaneous collapse of the entire profile to the anchor floor, which the canonical composite formula renders as exactly 0.0 (see §4). The floor is therefore a discrete state, not a point on a continuum — and it is reserved, by ruling, for a specific class of conduct.

Countries (12): Afghanistan, Belarus, Eritrea, Israel, Myanmar, North Korea, Russia, South Sudan, Sudan, Syria, Turkmenistan, Yemen.

Global cities (7): Pyongyang, Naypyidaw, Kabul, Asmara, Ashgabat (capitals of floor states) · Bangui, Port-au-Prince (parent countries not at floor — see §6).

AI labs (3): xAI/Grok, Character AI, Palantir AI.

Robotics labs (1): Ghost Robotics.

These cluster into four defensible families, each defined by the same underlying test — no remediation surface to credit:

The principle is consistent across all four: the floor marks conduct for which there is no remediation surface to credit — not merely "very bad outcomes." A bad outcome with a contested remediation channel (litigation, regulatory oversight, reversible policy) is scored above the floor by design.

The Critical band looks different in each index because the three dominant entity types fail differently and are judged on different evidence:

• States fail by active perpetration or codified impunity. They are scored on attested conduct — lethal force, signed impunity laws, judicial findings — and can reach an absolute 0.0 floor. The scoring question is floor entry vs. priced continuation (does new evidence cross the threshold, or merely re-confirm what is already priced?).
• Corporations fail by accumulated, unadjudicated enforcement. They are explicitly held not scorable until a merits adjudication or settlement (the FILED-BUT-UNADJUDICATED rulings). The scoring question is enforcement density vs. adjudication. This produces the persistent corporate cluster jammed just above the 20.0 Critical line (Oracle 20.6, Cigna/3M 20.3) and the absence of any corporate floor.
• AI / robotics labs fail by governance posture and contracting. They are scored net-neutral on announcements, pending regulatory enactment or court rulings — except where the product itself is the harm, which is the only route by which a lab reaches the floor. The scoring question is posture vs. operative conduct.

The per-type Critical profile signatures bear this out:

The countries Critical band is conduct-clustered (repression, conflict). The Fortune-500 Critical band is sector-clustered — fossil fuel and coal (Exxon, Chevron, Halliburton, Marathon, ~25 names), weapons/defense (Sturm Ruger, Vista Outdoor, TransDigm, Boeing), private prisons (GEO Group, CoreCivic), platform harm (Meta), and healthcare-denial (UnitedHealth). The same band, in two indexes, is organized by two entirely different logics.

The Fortune 500 is the largest index (448 entities) and has the most Critical entities in absolute terms (52), yet zero at the floor. Three converging reasons, two structural and one substantive.

(a) The all-anchor-collapse threshold is almost never reachable by a corporation. The floor requires every dimension at the minimum anchor simultaneously. The lowest-scored corporation in the entire index — GEO Group (private prisons) at 6.6, with a uniform [1.5 × 8] profile — still retains some detectable structure on every dimension. A going concern still detects customers (AWR), still has bounded scope (BND), still runs a governance system (SYS). It does not collapse all eight at once.

*(b) The composite formula structurally penalizes total collapse, not partial. The canonical composite is baseComposite + integrationPremium, where baseComposite = ((mean_anchor − 1) / 4) × 100 and the integration premium drops to 0 the instant any single dimension hits the absolute minimum. A uniform all-anchor-1 profile maps to baseComposite ((1−1)/4)×100 = 0.0 — the floor. A uniform all-1.5 profile maps to ((1.5−1)/4)×100 = 12.5. This is exactly why the Fortune-500 Critical band piles up at the quantized values 12.5, 14.1, 15.6, 18.8* — uniform or near-uniform low-anchor profiles. Verified histogram of the 52 F500 Critical composites:

The effective mathematical floor for a corporation that has not been ruled an active perpetrator is roughly 6–12, not 0. Critically, this is a side effect of the formula, not a deliberate judgment. UnitedHealth at 10.2 demonstrates corporations can drive single dimensions toward the floor (its ACC is 1.125, below GEO Group's uniform 1.5) — they simply never collapse all eight at once.

(c) Substantively, the floor rulings are conduct categories corporations rarely fit. The four floor families are in-territory perpetration, conflict famine, recognized structural atrocity, and weaponized/harmful product. Of the four product-floor entities, three are AI labs and one is a robotics lab — entities defined by an unremediable-harm product. A diversified corporation almost never matches: even GEO Group's harm has a (contested) remediation surface — contract terms, litigation, regulatory oversight — that the rulings credit above the floor.

Is the exemption defensible? The substantive reason (c) is: corporations are judged on a remediation-surface logic, and most genuinely have one. The structural reason (a/b) is the weak point — an entire entity class is held off the floor by a formula mechanic rather than by an explicit "corporations are judged differently, and here is why" ruling. The benchmark should make that distinction a stated judgment rather than let the math do it silently. This is the most defensible criticism an external party could level, and it is

Note the corporate Critical cluster is not only deep but also tight against the line from above: Oracle (20.6), Cigna and 3M (20.3) sit fractions of a point above the boundary under multiplying-but-unadjudicated sovereign actions, held out of Critical only by the FILED-BUT-UNADJUDICATED discipline. The corporate bottom is a pressure zone defined by adjudication timing, not by a floor.

The two entities sit within a tenth of a point. Their dimension profiles (verified from the index JSONs) show why that proximity is an artifact of a shared scale, not a shared meaning:

• Turkey's failure concentrates in BND (1.2), ACC (1.15), INT (1.15) — collapse of political boundaries, accountability, and integrity (opposition party leadership judicially removed, party HQ physically seized). This is governance failure, distributed across the apparatus of a state serving ~85 million people.
• UnitedHealth's failure concentrates in ACC (1.125), EMP (1.25), EQU (1.25) — collapse of accountability, empathy, and equity in claims handling (algorithmic claim-denial, DOJ probe). This is stakeholder-harm failure, bounded within a commercial relationship with members and patients.

The number is comparable; the kind of suffering is not. The benchmark's own design encodes this incommensurability — BND (Boundaries) means something structurally different for a sovereign than for a firm, and the integration-premium term behaves differently across the two. Severity is calibrated within-type, not across-type. Within countries, Turkey at 10.3 sits clearly above Russia at 0.0 and below the developing band — coherent. Within corporations, UnitedHealth at 10.2 sits below GEO Group at 6.6 and above the rest of the Critical cluster — coherent. The cross-type read ("a health insurer is as bad as an authoritarian state") is the one the methodology cannot support.

Recommendation: publish Critical-band placement with an explicit per-type interpretive frame, and avoid cross-type "league table" framing for the bottom band specifically.

The live entry pattern. The fastest-moving route into the Critical band in the 2026 cycles is not conflict or famine — it is documented, structural democratic backsliding among formerly-Developing democracies, scored on governance evidence (courts, opposition, press, franchise):

The countries Critical count rose from roughly 39 to 45 across the window. The "Dismantler" cohort (Croatia, Slovakia, Bulgaria, Italy) hovers just above the line within Developing. The conspicuous counter-trend is Hungary — a sustained multi-cycle upgrade (37.5 → 50.2) — the natural control case for "what reversal looks like." This pattern is distinct from the floor families: these entities are Critical, not floored, because functioning courts still reverse some of the conduct, giving a remediation surface the rulings credit above 0.0 (India at 15.6, not 0.0, is the clearest illustration).

Outliers worth naming:

• The city-inheritance asymmetry. Five floor cities are capitals of floor states (coherent inheritance). But Bangui (parent Central African Republic, 4.7) and Port-au-Prince (parent Haiti, 4.7) are at the 0.0 floor while their parent countries are not. The placement is defensible on the facts (non-state-actor collapse concentrated in the capital), but it rests on case-by-case judgment, not a stated rule.
• us-cities has zero Critical even though the United States itself is now Critical (17.5). US cities are scored on a domestic-service basis insulated from the federal-conduct downgrade — a within-system inconsistency worth flagging, though not necessarily an error.
• The cap/floor boundary is thin. The UAE is held at 18.4 (capped, Critical-but-not-floored) for external sponsorship of a documented atrocity, explicitly because it is an external sponsor (not in-territory perpetrator) and the ICJ case is unadjudicated. Russia and Israel, as in-territory perpetrators, are at 0.0. The distinction — in-territory vs. external, adjudicated vs. filed — is load-bearing and the UAE sits just on the safe side of it. The adjudication trigger that would move it toward the floor should be pre-registered.
• The quantization clustering (§4: 16 entities at 12.5, 12 at 14.1, 13 at 18.8) likely reflects placeholder first-baseline profiles rather than measured ones — a candidate for first-baseline re-assessment, not an interpretive finding about conduct.

• The corporate boundary cluster. Oracle (20.6), Cigna (20.3), 3M (20.3) sit fractions above the Critical line under unadjudicated enforcement. A single merits adjudication (e.g., the multi-state AG health-insurer suit) could reprice several Fortune 500 entities into Critical at once — and would directly test Q1 (does a corporation that crosses into Critical ever face a floor question?).
• The democratic-backsliding arc. The Dismantler cohort (Croatia, Slovakia, Bulgaria, Italy) is the next-likely set of Critical entrants on governance grounds. Hungary's upgrade arc is the control case to watch for what reversal looks like.
• The cap/floor line. An ICJ ruling on the RSF/Sudan case is the single highest-value forward trigger; it tests whether the complicity-by-proxy cap holds or converts toward the floor (Q4).
• Floor exits. No floor entity has exited in this cycle. The record is explicit that floors exit only through verified perpetration cessation, never through evidence accumulation or the passage of time. A documented cessation signal at any of the 23 would be the most significant possible movement at the bottom of the scale.

• Canonical scores (ground truth): site/src/data/indexes/{countries,fortune-500,global-cities,ai-labs,robotics-labs,us-states,us-cities}.json — all cohort counts, the 23-entity floor roster, the UnitedHealth/Turkey dimension vectors, and the F500 quantization histogram were recomputed directly from rankings[] and reconcile exactly with the pre-supplied snapshot (no drift).
• Primary analysis sources: docs/PATTERNANALYSISFLOORCRITICAL2026-06-11.md (floor roster, per-type Critical characterization, ruling taxonomy, cross-type comparability); docs/PATTERNANALYSISTHEMES_2026-06-11.md (three failure modes, democratic recession, cross-index meta-patterns).
• Ruling corpus / provenance: research/APPLIEDCHANGES.md, research/PENDINGCHANGES.md (rulings, boundary watch, the Oracle/UAE/US/Bolivia entries), research/change-proposals/2026-0[56]-.json.
• Formula / methodology: .claude/agents/benchmark-research.md, .claude/agents/overnight-assessor.md (floor-formula logic), site/src/data/dimensions.ts (8 dimensions), canonical composite baseComposite + integrationPremium.
• Fresh web evidence: none required for this edition. This is a structural/longitudinal analysis of the existing record; the marquee entity statuses (Turkey, UnitedHealth, UAE, Oracle, US, Bolivia) are grounded in the in-window ruling corpus and reconcile with canonical scores. Web-search budget (≤10) was not drawn down.