The State of Institutional Compassion — 2026

This is the inaugural edition of the Compassion Benchmark's flagship annual report — the institution's signature artifact, the read-it-first synthesis of everything the benchmark measures. Where the daily pipeline scans all entities for what changed in the last fortnight, and the year's Special Briefings each took one pattern and went deep, this report does the opposite of both: it pulls back to the whole field and asks the simplest, hardest question the benchmark exists to answer.

As of mid-2026, how do the world's institutions recognize, respond to, and reduce suffering — and what does the complete record actually show?

The benchmark applies one 0–100 framework, built on eight dimensions of conduct, to every institution it assesses: a sovereign state, a Fortune 500 corporation, an AI lab, a robotics lab, a US state, a US city, a global city. As of this writing it has scored 1,156 entities across seven indexes. That single shared scale is the institution's defining bet: that "how an institution treats those it affects" is a coherent thing to measure across radically different kinds of institution — and that putting them on one axis reveals patterns no single-sector ranking can.

This report is an honest first snapshot. The benchmark's structured data is 2026; this is not a multi-year trend report and makes no claim to be one. What it can do, and does, is three things at once: describe the state of the field (the full distribution across the five bands), name the one cross-cutting pattern that defines it (the Equity gap), and synthesize the year's evidence — the nine deep-dives and the roughly twenty scoring cycles applied since April — into a single, citable read.

The central thesis is plain and uncomfortable: the state of institutional compassion in 2026 is mediocrity at scale. The modal institution is Developing; the field median is 35.9; the tails — the genuinely failed and the genuinely exemplary — are thin. And underneath the distribution sits one near-universal pattern: institutions of every kind, at every level of overall performance, are weakest at equity — fairness toward those with the greatest need and the least power. That single finding unifies the top and the bottom of the scale and is, by some distance, the most diagnostic thing the record contains.

All counts in this section are recomputed directly from rankings[] in each index JSON, using the canonical band cutoffs (Critical ≤ 20, Developing 20–40, Functional 40–60, Established 60–80, Exemplary ≥ 80). They reconcile exactly with the published per-index data and with the field figures reported in this year's Middle of the Scale briefing.

The shape of the field is the headline. The modal institution is Developing — 535 entities, nearly half the record, sit in the 20–40 band. Add the Functional band and 783 of 1,156 entities (67.7%) live in the middle. The field median composite is 35.9 — a high-Developing score — and the mean is 38.9. The dramatic cases the public hears about — the 23 institutions at the absolute floor, the 63 in the Exemplary band — together account for fewer than one entity in eleven.

This is the state-of-the-field finding, and it is sobering precisely because it is undramatic: the world's institutions, measured on how they treat those they affect, are overwhelmingly mediocre. Not collapsed, not exemplary — partial. A Developing or Functional score means, in the benchmark's own terms, that core practices may exist and meet a basic bar, with significant gaps remaining: systems on paper or in pilot, without the disaggregated data, independent audit, or sustained-under-pressure evidence that the higher bands require. It is the band of we started, but we cannot prove it worked. The tails are exceptional; the middle is the world.

Each index has its own center of gravity, and the differences are themselves a finding.

Three structural facts stand out:

• The Fortune 500 is the largest and the most middling index. 448 corporations cluster heavily in Developing (215) and Functional (117), with only 7 Exemplary (1.6%) and zero at the floor. The corporate field is a dense mediocre middle with very few standouts in either direction.
• Robotics labs are the highest-scoring index (mean 56.7; 26% Exemplary) and countries and global cities are the lowest-centered (means 35.2–36.5; the deepest Critical concentrations). This is not because robotics labs are "better institutions" than states — it is because a single pro-social product line clears the band on a far narrower evidentiary surface than a whole-of-population sovereign test does (see §4).
• The bands are an artifact of evidentiary surface as much as of conduct. US states show a barbell (8 Exemplary, 4 Critical, nothing in Functional/Established) reflecting a partial index; US cities have zero Critical entities even though the United States itself is now Critical, because cities are scored on domestic-service delivery insulated from the federal-conduct downgrade. The distribution must be read per-type, not as one league table.

• The bottom: 177 entities are Critical, and 23 sit at the absolute 0.0 floor — every one carrying an identical profile, all eight dimensions collapsed to the minimum anchor. The floor is not a low score; it is a discrete state, reserved by ruling for conduct that leaves no remediation surface to credit: active in-territory state perpetration (Russia, Myanmar, North Korea, Eritrea, Turkmenistan, Belarus, Syria), conflict-driven famine (Sudan, South Sudan, Yemen), formally recognized structural atrocity (Afghanistan, Israel), and product-is-the-harm labs (xAI/Grok, Character AI, Palantir AI, Ghost Robotics).
• The top: 63 entities are Exemplary (composite ≥ 80). They are led by Nordic and Alpine cities (Copenhagen, Helsinki, Oslo, Stockholm, Vienna all at 100), Switzerland (97.5) among countries, assistive-robotics labs (Open Bionics 97.5, Ottobock 95.9), and a handful of corporations (Target 92.8). Exemplary is an integration achievement, not a peak score — it is reached by having no weak dimension, not by excelling at one.

The single most important structural fact about the tails is their symmetry: the floor and the ceiling are produced by the same formula mechanic run in opposite directions (total collapse vs. uniform high conduct), and both are reached on different evidentiary surfaces across types. That makes the cross-type read — "this lab is as good as that country," "this insurer is as bad as that state" — the one thing the framework cannot support, at either end.

If §2 is the shape of the field, this is the single thread that runs through all of it. It is the most diagnostic finding the benchmark has produced, and it is the throughline of this report.

The benchmark scores eight dimensions: Awareness, Empathy, Action, Equity, Boundaries, Accountability, Systemic Thinking, and Integrity. Equity (EQU) measures whether care is distributed fairly, with explicit priority for those who have the greatest need and the least power — its subdimensions are Universality, Priority for Vulnerable, Bias Awareness, Access Design, and Historical Harm Acknowledgment. It asks not "does this institution help?" but "does it help the people who are hardest to reach, and can it prove it?"

Equity is the weakest or tied-weakest dimension for 1,046 of 1,156 entities — 90.5% of the field — and the strictly lowest single dimension for 593 (51.3%). No other dimension is the minimum for anything approaching a majority. Recomputed directly from scores{} across all seven indexes:

The global Equity mean is 2.21 on the 1–5 scale — the lowest of all eight dimensions (next-lowest is Integrity at 2.45; Action peaks at 2.69). It is the lowest-mean dimension in six of the seven indexes; the lone exception is global cities, where Boundaries edges it by a sliver and Equity is a close second.

Two facts make this the report's signature cross-cutting finding:

• It is universal across the quality spectrum. The gap between an entity's other-seven-dimension average and its Equity score holds at every level of overall performance. Switzerland (composite 97.5) carries it; Sudan (composite 0.0) carries it. It is not a marker of failed institutions — it is a near-constant of institutional conduct as the benchmark measures it.
• It unifies the top and the bottom. At the floor, collapsed equity sits alongside collapsed accountability and integrity. At the ceiling, equity is the lone thing still missing: of the 63 Exemplary entities, 61 have Equity as their weakest dimension, and 36 carry Equity as their single below-threshold score. The benchmark's top band is, in practice, "excellent at everything except fairness." From either end of the scale, Equity is the dimension institutions reach last.

The bar Equity sets is genuinely demanding — its top anchors require disaggregated outcome data, independent audits, and co-design with affected communities — and those are uncommon in the real world and, by at least one external measure, becoming rarer. The unavoidable open question is whether the near-universal equity gap is a finding about how institutions actually behave or an artifact of a bar almost no one meets. The report's view is that it is mostly the former (the gap holds across the entire quality range, including institutions that clear high bars everywhere else), but the question is real and is filed for methodology review.

The benchmark's defining feature is that it scores governments, corporations, AI labs, robotics labs, and cities on one framework. This is a genuine and unusual asset. No other institution scores robotics labs on harm accountability at all — the purpose-to-score gradient in that index (§6) is a finding only this record can produce. The shared eight-dimension framework lets a reader ask, for the first time, whether a trillion-dollar AI lab and a sovereign state recognize and reduce suffering on comparable terms.

But the year's deep-dives establish, repeatedly, that comparability is within-type, not cross-type — and saying so explicitly is part of what makes the benchmark citable rather than glib.

• The three dominant types fail differently and are judged on different evidence. States fail by active perpetration or codified impunity and can reach the 0.0 floor. Corporations fail by accumulated, unadjudicated enforcement — and are held not scorable until a merits adjudication, which produces a dense corporate cluster jammed just above the Critical line and zero corporations at the floor. AI and robotics labs fail by governance posture and contracting, and reach the floor only when the product itself is the harm.
• A shared number is not a shared meaning. UnitedHealth (10.2, a Fortune 500 health insurer) and Turkey (10.3, a sovereign state) sit within a tenth of a point. Turkey's failure concentrates in political boundaries, accountability, and integrity — governance failure distributed across a state serving ~85 million people. UnitedHealth's concentrates in accountability, empathy, and equity in claims handling — stakeholder harm bounded within a commercial relationship. The number is comparable; the kind of suffering is not.
• The same logic holds at the top. An exemplary robotics lab is judged on whether one assistive product is safe and accessible; an exemplary country on whether a whole state treats millions fairly. Both can score 81–98, but "Exemplary" rests on a far narrower evidentiary surface for a single-product lab than for a sovereign. Robotics is 26% Exemplary — far above any other index — partly because a narrow pro-social mandate satisfies the band easily.

The institution's unique asset is therefore double-edged: the one framework is what makes the comparison possible, and the within-type calibration is what keeps it honest. The report's recommendation, consistent with the year's briefings, is that cross-type "league table" framing be avoided at the extremes specifically, and that band placement be published with a per-type interpretive frame. Filed for methodology review.

Since the first applied cycle on 2026-04-15, roughly twenty scoring cycles have been logged in research/APPLIED_CHANGES.md. As a mid-2026 snapshot this is the benchmark's first scoring season, not a multi-year trend; but within 2026 the direction is unmistakable. The dominant motion is downward, and the few upgrades are hard-won and tightly disciplined. Two patterns account for most of the movement.

The deepest and most concentrated movement is democratic backsliding among formerly-Developing states, scored on governance evidence — courts, opposition, franchise, civil society — rather than on conflict or famine:

On the corporate side, the year's clearest new doctrine is the layoffs-despite-profits line, which set two anchors in a single window:

• Procter & Gamble: 86.1 → 79.0 (Exemplary → Established). 7,000 cuts announced while profitable and projecting profits, with a tariff-cost-protection rationale. The downgrade lands on exactly two dimensions — Boundaries and Integrity — and turns on the distinction between distress-driven and profit-protection layoffs, not on headcount.
• Oracle: 20.6 → 14.7 (Developing → Critical). 30,000 cuts wrapped in a coercive-severance architecture — sign a legal release waiving the right to sue or forfeit all severance. The decisive element was structure, not scale: the first application of a coercive-severance trigger, hitting seven of eight dimensions at once.

Other notable 2026 downgrades, each grounded in a documented event, include State Street (92.5 → 60.2), Abbott (87.4 → 57.8), Microsoft (87.8 → 66.4), CVS Health (50.0 → 25.6), Johnson & Johnson (48.4 → 27.5), and the AI-lab repricings of OpenAI, Anthropic, and Mistral away from inflated first baselines.

Against that tide, the upgrades are few and instructive. The benchmark credits genuine reversal — but only when it is proven, self-initiated, and sustained, and it never confers a top score on momentum:

• Bangladesh: 34.4 → 39.8 (+5.4, rank +35). The clearest 2026 upgrade. Bangladesh exited the ITUC's ten-worst countries for workers' rights for the first time since 2017, after November 2025 labour-law reforms (union formation eased to 20 workers, 120-day maternity leave, anti-discrimination protections) and ratification of ILO Conventions C155, C187, and C190. Tempered: it retains ITUC Rating 5 ("no guarantee of rights"), and SEZ workers remain excluded — so the gain is real but lands just below Functional.
• Hungary: 28.1 → 50.2 across multiple cycles. A documented, self-initiated institutional recovery (16-minute cabinet, audit ordered, ICC reversal sustained) — the year's control case for what reversal looks like. It reaches only the middle, two full bands below "good," precisely because the benchmark grades sustained conduct, not trajectory.
• Smaller documented gains: Venezuela (4.4 → 18.0), Botswana (60.9 → 67.5), Saudi Arabia (4.4 → 9.4), New York City (48.4 → 56.3).

The asymmetry between §5.1 and §5.2 — many disciplined downgrades, few capped upgrades — is the scoring story of 2026.

What unifies every entry above is a pre-adjudication discipline applied symmetrically. The benchmark prices proof, not accusation. A 42-state attorney-general subpoena did not move OpenAI's score (allegation, not adjudicated); Oracle's documented severance clause did (in-effect conduct). UnitedHealth faces a coordinated multi-state AG investigation, a DOJ criminal probe, and a shareholder suit at once — and is held at 10.2, because density of investigation is not adjudicated harm. The same rule runs in the good-news direction: Microsoft's compelled human-rights remedy is held below a self-initiated one, and Hungary's genuine recovery is credited as conduct but stops at the middle. This discipline — and not any single score — is the load-bearing reason the record is defensible and citable.

This report sits atop nine Special Briefings published this year. Each took one pattern and went deep; together they triangulate the field. Read as a set, they are the evidentiary spine of this synthesis.

1. The Floor and the Critical Band — 177 Critical, 23 at the 0.0 floor; the floor is a discrete "total-collapse" state reached by four entity types on different evidentiary bars, and corporations are exempted from it by a formula mechanic rather than a stated rule.
2. What Good Looks Like (Exemplars) — 63 Exemplary; "good" is an integration achievement (no weak dimension), the cleanest exemplars are assistive-robotics labs (the mirror of the floor), and even the best are weakest at equity.
3. The Equity Tax — the foundational dimension-level finding this report elevates to its cross-cutting pattern: Equity is weakest for 90.5% of the field and the lone sub-band score for most exemplars.
4. AI Governance Under Pressure — a government-mandated shutdown, a 42-state subpoena, and a union vote in one fortnight, none of which moved a composite; external governance is testing AI labs faster than their internal practice is maturing.
5. Layoffs Despite Profits — the P&G and Oracle anchors; the benchmark now distinguishes distress-driven from profit-protection layoffs and prices coercive-severance structure as scorable harm.
6. State of Exception — the codified-impunity pattern: governments legislating their way to the bottom; Bolivia's 35.9 → 6.3 arc as the first "predicted-trigger-realized" sequence.
7. What the Product Is For — sorting the 100 technology labs by end-use yields a single gradient: defense/surveillance/weapons at the floor, healthcare/accessibility at the ceiling; Boston Dynamics scores 65.6 (research) and 20.3 (weaponized demo) — a 45-point gap inside one institution.
8. Allegation, Indictment, Ruling — the pre-adjudication discipline made explicit across six clean test cases; the benchmark scores proof, discounts allegation, and disciplines good news too.
9. The Middle of the Scale — the on-ramp: 783 entities (67.7%) in the Developing/Functional middle, where the integration premium is switched off and the composite is the plain dimension average; the median institution is mediocre, not lopsided.

The throughlines across all nine are the three this report foregrounds: the mediocre middle, the universal equity gap, and the within-type discipline that keeps cross-type comparison honest.

This is a snapshot, and snapshots have edges. The standing patterns most likely to define the next edition:

• Democratic backsliding is the live entry pattern into the bottom of the scale, and it is accelerating on legal grounds, not battlefield ones. The "Dismantler" cohort just above the Critical line (Croatia, Slovakia, Bulgaria, Italy) is the next-likely set of entrants; Hungary's upgrade arc is the control case for what reversal looks like. Watch the gap between codified impunity (which floors a score) and contested-but-reversible conduct (which holds it in Critical) — a functioning court is, repeatedly, the difference between Critical and the floor.
• AI governance is testing labs faster than their practice is maturing. Subpoenas, compelled shutdowns, and union votes are arriving as forward indicators ahead of any scored event. The open question is what converts external pressure — especially worker voice — into a scorable accountability signal, and the first adjudication of a major AI enforcement action (the OpenAI multi-AG probe is the leading candidate) will set the precedent.
• The corporate boundary cluster is a coiled spring. A dense band of Fortune 500 entities sits just above the Critical line under unadjudicated enforcement; a single merits adjudication could reprice several into Critical at once — and would directly test whether a corporation that crosses into Critical ever faces a floor question, which the framework currently does not answer.
• The Equity gap is the standing structural finding — the one pattern that will almost certainly survive into every future edition, and the one most worth watching for movement. An institution that closes its equity gap while holding the other seven dimensions is the rarest and most significant possible motion in the record; as of mid-2026, only Open Bionics and Switzerland have cleared it at the top of the scale.

The honest summary of the inaugural state of the field is this: the world's institutions are, on the whole, partial — they have started, and cannot yet prove it worked — and they are most partial, almost universally, at fairness to those with the least power. That is the baseline. Future editions will measure the field against it.

• Canonical scores (ground truth): site/src/data/indexes/{countries,fortune-500,global-cities,ai-labs,robotics-labs,us-states,us-cities}.json — the five-band distribution (177/535/248/133/63), the per-index breakdowns, the field median (35.9) and mean (38.9), the Equity diagnostics (1,046/1,156 = 90.5% weakest/tied; 593 strictly lowest; global EQU mean 2.21; 61/63 and 36/63 among Exemplary), and the 23-entity floor roster were all recomputed directly from rankings[] and scores{} and reconcile exactly with the published data (no drift).
• The year's scoring record: research/APPLIED_CHANGES.md (all ~20 cycles since 2026-04-15: Bolivia ×4, United States, UAE, Israel, India, P&G, Oracle, Bangladesh, Hungary, and the corporate and AI-lab repricings, each with its applied delta, dimensions, and ruling rationale); research/change-proposals/*.json (per-entity proposals).
• The nine Special Briefings synthesized here: research/special-briefings/{floor-and-critical-2026-06-11, exemplars-2026-06-11, equity-tax-2026-06-16, ai-governance-2026-06-15, layoffs-despite-profits-2026-06-15, state-of-exception-2026-06-16, what-the-product-is-for-2026-06-16, allegation-indictment-ruling-2026-06-16, middle-of-the-scale-2026-06-16}.md — their publicSummaries supply the cited per-pattern findings (the floor taxonomy, the integration-premium mechanic, the purpose-to-score gradient, the pre-adjudication discipline, the middle-of-the-scale field figures).
• Formula / methodology: site/scripts/lib/scoring.mjs::computeCompositeFromDimensions (canonical composite), site/src/data/dimensions.ts (the 8 dimensions, 40 subdimensions, 5 bands), and the floor/integration-premium logic in the assessor agent definitions.
• Fresh web evidence (external grounding for cited facts):
• Bangladesh exiting the ITUC ten-worst list: New Age — "Bangladesh exits world's worst 10 for workers' rights"; ITUC Global Rights Index.
• UN "hallmarks of genocide" finding in Darfur and UAE arms supply to the RSF (basis for the UAE downgrade): Human Rights Watch — "UN Body Finds 'Hallmarks of Genocide' in Darfur"; allAfrica — "UN Accuses UAE of Funnelling British-Made Arms to Militias in Sudan".
• Independent corroboration of the top-of-scale geography (Nordic/Alpine clustering): Gallup / World Happiness Report rankings — convergent (Finland, Iceland, Denmark, Sweden, Norway lead; Switzerland top-tier), not circular evidence for the benchmark's distinct conduct-based methodology.