Atlas of atlases: eighteen research surfaces organised by the four-rung ladder of a delegated act — existence, authority, oversight, and verifiability.

Draft — not peer-reviewed. Working edition for review only; every number carries a proof ceiling and may change.
atlas of atlases.

The same gap, re-measured one rung at a time

Eighteen research surfaces, reorganised not by data source but by the ladder of a delegated act: does the act even exist on a record, under whose authority, with what oversight, and can a third party verify it? The same legibility gap is strong at the bottom and collapses at the top — with a hard floor where the actor is a machine.

apparatus mean ≈ 2.76 ai-legibility ≈ 1.32 paired d ≈ 2.86 27 / 27 EU states 0 / 11 agentic-security cases reach a court

The four-rung ladder

Each rung asks one question of a delegated act, and a different set of observatories answers it. The apparatus is strongest at the bottom (mature trust infrastructure) and weakest at the top (does the act exist at all).

rung 1
existence

Does the act or agent exist on any record?

apparatus: weak rung 2
authority

Under whose authority does it act?

apparatus: weak rung 3
oversight

With what oversight is it run?

apparatus: mixed rung 4
verifiability

Can a third party verify it?

apparatus: strong

How to read every claim

One grammar is reused on every surface. A headline is one sentence. A warrant names the observatory cell, the indicator, the count, and the test statistic. A bedrock is the primary record itself, carrying its evidence tier and a proof-ceiling label. Every headline number clicks down to a primary record.

headline

one sentence

The finding, stated plainly and falsifiably.

warrant

cell + n + stat

Observatory cell, indicator, count and test statistic.

bedrock

the primary record

The trusted-list entry, case, or watch-event — with its tier.

rung 1

existence

does the act or agent exist on any record?

The weakest rung. Most agentic acts leave no register entry at all; what we can count is the harm catalogue and the thin set of public registries. ai-legibility sits near 1.32.

gaia

Traces of AI systems across public records.

gaia.eatf.eu
public-ai-registry

Where public-sector AI is (and is not) registered.

public-ai-registry.eatf.eu
obscure-ai

An atlas of 297 AI edge-case harms.

obscure-ai.eatf.eu
limen

The boundary cases that feed the catalogue.

limen.eatf.eu
figure 4
The publication mirage

Apparent US dominance of AI enforcement is a publication artefact: harm count tracks language family, not AI-law readiness (r = −0.17 overall; flips to +0.14 once English-primary jurisdictions are removed).

Scatter of evidence-grade harm count against PALLAS AI-law readiness, points coloured by language family; English-primary jurisdictions carry most harm and the correlation sign flips when they are removed.
source: obscure-ai harms × PALLAS AI-law readiness, 297 cases descriptive counts of public records, not incidence
figure 5
The visibility ceiling

Evidence reaches a regulator or court (tier T3) in 53.2% of cases overall, but all 11 security and agentic cases stall at the CVE disclosure tier — 0% adjudicated: disclosed, never adjudicated.

Stacked bars per harm category showing the highest evidence tier reached; the security and agentic category sits entirely at the CVE disclosure tier with zero reaching a court.
source: obscure-ai proof-ceiling distribution, 297 cases by category proof ceiling: T1–T4 per case; the floor cell is hard
rung 2

authority

under whose authority does it act?

Law and standards say who may act and on what basis; the crosswalk surfaces map AI-Act obligations onto the underlying primitives. This rung is still weak — duties are declared faster than the evidence that would discharge them.

atlas

Law and standards crosswalk for AI governance.

atlas.eatf.eu
pallas

AI-governance index, indicators I01 / I02 (AI-law, authority).

pallas.eatf.eu
legal

Case law on AI use, including CLIO and NOMAD lines.

legal.eatf.eu
patent

Who claims the primitives, by assignee and CPC class.

patent.eatf.eu
figure 6
The drift atlas

Of 2,500 re-read cells only 12.3% show a true score move (53.8% re-affirmed, 33.9% confidence-only); moves skew up, but 306 of 307 carry a human-review flag — proposed, not merged.

Diverging bars of net true-score move per indicator alongside a donut of 2,500 re-read cells split into re-affirmed, confidence-only, and true-move classes; almost all true moves remain proposed rather than applied.
source: PALLAS v0 freeze vs 17 dated re-read sprints proof ceiling: drift is proposed, not applied to the frozen matrix
rung 3

oversight

with what oversight is it run?

Procurement signals and supervisory indicators (PALLAS I05 / I08) show oversight is mixed: declared in policy, uneven in practice, and rarely tied back to an attributable act.

rhr

Procurement and rights-holder oversight signals.

rhr.eatf.eu
ai-observability

AI-governance observability index across jurisdictions.

ai-observability.eatf.eu
aion

Scenario lab for agentic oversight and incident paths.

aion.eatf.eu
cassandra

Dated diffs that track when oversight rules change.

cassandra.eatf.eu
figure 8
The two clocks of attributability debt

Capability advances on a fast clock and attributability on a slow one; the gap is widest at the agent-decision layer, the only rung with a hard 0% adjudication ceiling.

Two paired bars per stack layer comparing a fast capability clock to a slower attributability clock; the gap widens steadily and is largest at the agent-decision layer. Marked schematic.
source: schematic synthesis across the seven surfaces schematic — bar lengths illustrate the argument, not measured values
rung 4

verifiability

can a third party verify it?

The strongest rung. Mature trust infrastructure (remote signature-creation devices, attestation, sealing) lets a third party verify — apparatus near 2.76. But the strong rung is also the narrowest: delivery is acutely concentrated by jurisdiction.

rqscd

EU remote signature-creation-device census and watch.

rqscd.eatf.eu
vesta

Verification and sealing of evidence records.

vesta.eatf.eu
sigil

Duty index — obligations that demand proof.

sigil.eatf.eu
factory

Reproducibility and attestation workbench.

factory.eatf.eu
eco

The verifiability ecosystem and its dependencies.

eco.eatf.eu
eatf

The umbrella that hosts every surface.

eatf.eu

Flagship finding — click from headline to bedrock

flagship — the trust-service mirage

1 headline

Nine EU states post a perfect technical-trust subindex yet have delivered zero remote signature-creation-device grants. The score is the promise; the trusted list is the delivery; they are not the same axis.

2 warrant — observatory cell, n, test stat

Observatories PALLAS technical-trust subindex joined to the rQSCD trusted-list census on ISO country. n = 124 granted entries snapshot 2026-06-27 9 / 27 states: subindex = 100, grants = 0 only 10 / 27 deliver any grant. The nine: France, Germany, Greece, Denmark, Portugal, Finland, Lithuania, Latvia, Hungary. Read strictly as declared-versus-delivered for one new service type at one snapshot — not as evidence the index mis-scores trust.

3 bedrock — the primary record

The delivery axis is a national trusted list, not a score. A delivering state (Italy) publishes its remote-QSCD management entries on the AgID trusted-list service; the nine flagged states publish none. Follow a real listed entry to its primary record:
primary record: Actalis S.p.A. — management of remote QSCD (AgID, Italy) ↗
evidence tier: T1 — primary trusted-list record (national supervisory body). proof ceiling: listed-entry telemetry, not validation, supervision, or legal status. The mirage is one snapshot of one new service type; it weakens if delayed grants for the nine zero-delivery states arrive within two snapshots.
cross-check (EU level): European Commission trusted-list browser ↗.
figure 1
The trust-service mirage

Nine EU states post a perfect technical-trust subindex (100) yet show zero delivered remote-QSCD grants — the score is the promise, the trusted list is the delivery, and they are not the same axis.

Mirror bar chart of EU-27 states: declared technical-trust subindex on the left, delivered remote-QSCD grants on the right; nine states with a perfect subindex and zero grants are flagged in red.
source: PALLAS technical-trust subindex × rQSCD census, EU-27 proof ceiling: declared vs delivered, one snapshot
figure 2
One country, two denominators

Italy's apparent dominance of delivery is a unit-of-analysis artefact: 62.9% per listed entry collapses to 33.3% per distinct named service (per-certificate inflation).

Two bars comparing Italy's share of EU remote-QSCD delivery: 62.9% counted per listed entry versus 33.3% counted per distinct named service, bracketed as per-certificate inflation.
source: rQSCD census, clean 124-entry / 48-named-service count proof ceiling: per-entry headline is an upper bound
figure 3
The churn that wasn't

The churn is not gradual loss: all 60 day-one removals are matched re-issues (zero genuine withdrawals); after day one the series is pure accrual to +122 net entries.

Step chart of cumulative net listed entries rising to plus 122 over fourteen daily captures; a day-one spike of sixty removals is annotated as matched re-issues, not genuine withdrawals.
source: rQSCD watch-events, 14 daily captures (2026-06) proof ceiling: listed-entry telemetry, not legal status
figure 7
Three axes that must not be collapsed

Promise, delivery and demand are three separate axes: near-uniform promise (22 / 27 at max I06) coexists with delivery in only 10 states and real demand, firing a derived gap flag for 11 states.

Three side-by-side column charts per EU-27 country for promise, delivery and demand; eleven states with high promise, real demand and zero delivery are highlighted as a derived gap.
source: PALLAS I06 × rQSCD delivery × obscure-ai demand, EU-27 proof ceiling: gap column is derived; demand is tiny-N

Three cross-cutting overlays

Change clock

One synchronised date scrubber over Cassandra diffs, rQSCD watch-events, PALLAS edition drift and AI-Act deadlines. The gap widens as regulation accelerates on the already-legible bottom rung and stands still on the top.

Concentration overlay

Per layer, not cross-layer. Italy holds 62.9% of remote-QSCD delivery by jurisdiction; idle-lab single-provider blast-radius reaches 95.2%. The strongest rung (verifiability) is also the narrowest.

Counting-the-wrong-thing operator

The same declared-versus-delivered subtraction run on every surface, surfacing the population / stress-property divergence in the same health-overstating direction.

Single cross-layer owner refuted

Tested 2026-06-29: no single firm owns patents, trust infrastructure and AI equity end-to-end. Concentration is per-layer and independent, with no demonstrated common actor. Kept as a worked negative result — not a positive claim.