The Trust Gap, Vol. 5: the absence rate fell 6.5 points and none of it was new evidence

## Summary

Between Vol. 4 and this volume the share of our published data points marked `no_public_information` fell from **73.37% to 66.86%** — a 6.51-point drop, the largest single-week movement in five volumes.

None of it is new evidence.

On 2026-08-26 a publication floor took effect that demotes any entity carrying fewer than 8 present fields or no scored dimension. It moved 943 entities out of the published set; a further 5 left the same day under a separate editorial withdrawal, for 948 in total. The index did not learn more about the market; it stopped publishing the entities it knew least about. Measured on the identical population Vol. 4 measured, the absence rate over the same week went **up 0.88 points**, from 73.37% to 74.25%.

We are reporting the improvement and the correction together, in that order, because a reader who saw only the first number would draw exactly the wrong conclusion about what changed.

The finding that does survive the population change is the one this publication exists to make. Across an identical 4,656 slots each, on the same 776 entities, the six control-and-governance fields carry **1,951** disclosures and the six risk-transfer fields carry **12**. That ratio — 162.6 to 1 — is immune to which entities we publish, because both halves are measured on the same rows.

---

## 1. The population changed under us

`entities.status` on 2026-08-31:

| Status | Entities | |---|---| | candidate | 2,256 | | published | **776** | | archived | 2 |

Published breaks down as 747 agents and 29 vendors.

Vol. 4 measured 1,704 published entities. To compare anything week over week we first had to reconstruct who those 1,704 were. `entity_overwrite_log` records every status transition with a timestamp, so the published set as of Vol. 4''s run can be recovered by reverse-replaying the log: for each entity, take the `old_status` of the earliest logged change after 2026-08-24 20:10 UTC, or its current status if there is none.

That reconstruction returns **1,704 published** — matching Vol. 4''s reported figure exactly, from a method Vol. 4 did not use. We treat that as validation of the cohort, and every week-over-week number below rests on it.

The flow reconciles completely:

| Movement | Entities | |---|---| | Published at Vol. 4 | 1,704 | | Demoted 2026-08-26 by the publication floor | −943 | | Withdrawn 2026-08-26 under a separate editorial decision | −5 | | Re-promoted since, having cleared the floor | +18 | | Promoted since, not previously published | +2 | | **Published now** | **776** |

The 18 re-promotions are the floor working as designed. It is a waiting room, not a cull: an entity that gathers evidence clears the bar on the next attempt, and 18 already have in five days.

---

## 2. The decomposition

Three measurements. Two are queries against the current database; the third is Vol. 4''s published figure.

| Measurement | Slots | `no_public_information` | Rate | |---|---|---|---| | Vol. 4 cohort (1,704), as Vol. 4 measured it | 100,116 | 73,455 | **73.37%** | | Vol. 4 cohort (1,704), measured today | 100,096 | 74,322 | **74.25%** | | Current cohort (776), measured today | 45,494 | 30,418 | **66.86%** |

Holding the data constant and changing only the population isolates the composition effect. Holding the population constant and changing only the data isolates real editorial movement:

- **Composition: −7.39 points.** 74.25% → 66.86%. - **New evidence: +0.88 points.** 73.37% → 74.25%. - **Net: −6.51 points.**

The two terms sum to the total exactly. The headline improvement is −7.39 points of selection partially offset by +0.88 points of genuine deterioration.

The 20-slot difference between the two 1,704-entity rows is fully accounted for: two entities were reclassified from agent to vendor on 2026-08-28, and vendors are assessed on 49 fields where agents are assessed on 59.

### Why the like-for-like number moved against us

On the Vol. 4 cohort, `present` values fell by 685 over the week while `no_public_information` rose by 867 and never-assessed slots fell by 202.

The largest single contributor is retraction, not discovery. Of the field rows touched on that cohort since Vol. 4, 838 were updated in place to `no_public_information` — 468 of them by the Fact-Checker, 259 by the spark-extractor, 64 and 45 by two further Fact-Checker passes. Only 190 rows were updated in place to `present`.

An index whose absence rate rises because its editors withdrew claims they could no longer stand behind is working correctly. It is still, on the honest measure, a worse week than the previous one.

---

## 3. Why the floor moved the number so far

The demoted entities were not a random sample. They were, by construction, the thinnest records we had.

| Cohort | Entities | Slots | `present` | Absence rate | Present fields per entity | |---|---|---|---|---|---| | Retained (published at both dates) | 774 | 45,376 | 12,607 | 66.83% | **16.29** | | Demoted | 930 | 54,720 | 7,540 | 80.40% | **8.11** | | Newly promoted | 2 | 118 | 25 | 78.81% | 12.50 |

The retained entities carry almost exactly twice the evidence per entity. The floor is a quality gate, and a quality gate applied to a corpus mechanically improves every aggregate computed over the survivors. That is not a criticism of the floor — it is the reason the floor exists — but it disqualifies every cross-volume aggregate that does not control for it.

**A consequence readers should weigh:** the 776 entities are no longer a sample of the agent market. They are a sample of the agent market *conditioned on our having found at least 8 facts and scored one dimension*. Rates computed on them describe documented agents, not agents. Where that distinction matters below, we say so.

---

## 4. The assurance layer, field by field

All 24 assurance fields, on the current 776 published entities. `Applicable` is 776 for fields applying to both types and 29 for vendor-only fields.

| Field | Applicable | Present | `no_public_information` | Never assessed | Present % | |---|---|---|---|---|---| | Service type *(vendor)* | 29 | 24 | 3 | 2 | 82.76% | | Client types *(vendor)* | 29 | 19 | 8 | 2 | 65.52% | | Explainability | 776 | 447 | 329 | 0 | 57.60% | | Audit trail | 776 | 421 | 355 | 0 | 54.25% | | Runtime governance | 776 | 373 | 403 | 0 | 48.07% | | Compliance certifications | 776 | 371 | 405 | 0 | 47.81% | | Permission scopes | 776 | 312 | 464 | 0 | 40.21% | | Regulatory alignment | 776 | 293 | 483 | 0 | 37.76% | | Underwriting backing *(vendor)* | 29 | 9 | 18 | 2 | 31.03% | | SLA terms | 776 | 158 | 618 | 0 | 20.36% | | Evaluation coverage | 776 | 139 | 637 | 0 | 17.91% | | Dispute process | 776 | 116 | 660 | 0 | 14.95% | | Outcome-based pricing | 776 | 112 | 664 | 0 | 14.43% | | Settlement mechanism | 776 | 69 | 707 | 0 | 8.89% | | Self-reported performance | 776 | 41 | 125 | 610 | 5.28% | | Tamper-evident log | 776 | 27 | 749 | 0 | 3.48% | | Regulatory status *(vendor)* | 29 | 1 | 26 | 2 | 3.45% | | Asset custody | 776 | 21 | 147 | 608 | 2.71% | | **Insurance available** | 776 | **4** | 772 | 0 | **0.52%** | | **Indemnification** | 776 | **3** | 773 | 0 | **0.39%** | | **Liability cap** | 776 | **2** | 774 | 0 | **0.26%** | | **Own liability cover** | 776 | **2** | 165 | 609 | **0.26%** | | **Insurance carriers** | 776 | **1** | 775 | 0 | **0.13%** | | **Coverage limits** | 776 | **0** | 776 | 0 | **0.00%** |

**`coverage_limits` is at 0 present for the fifth consecutive volume.** No entity we publish has ever disclosed a policy limit. It remains the only field in a 63-field catalogue that has never once been satisfied.

**772 of 776 published entities (99.48%) have no public evidence of insurance availability.**

### The controls/risk-transfer split, measured on identical denominators

Six control fields against six risk-transfer fields, same 776 entities, 4,656 applicable slots each:

| Block | Applicable slots | Assessed | Present | % of applicable | |---|---|---|---|---| | Controls | 4,656 | 4,656 | **1,951** | 41.90% | | Risk transfer | 4,656 | 4,047 | **12** | 0.26% |

Because the denominators are identical, this comparison is unaffected by the publication floor, by which entities we chose to publish, or by how thin the corpus is. Whatever selection applies, it applies to both halves equally.

**162.6 control disclosures for every risk-transfer disclosure.** That is the finding.

---

## 5. The complete risk-transfer census

Twelve rows is small enough to publish in full rather than sample, so here is every one, read against its stored capture. This is the entire evidentiary basis for our claim that the risk-transfer layer does not exist in public.

**`insurance_available` — 4 rows:**

1. **Coalition** *(vendor)* — a cyber insurer describing its own product: "Cyber Insurance | Active Insurance & Cybersecurity | Coalition Now Available: Active Cyber Insurance for Enterprises". A carrier selling insurance, not an agent carrying it. 2. **ElevenAgents** *(agent)* — "ElevenAgents is also the first agent platform eligible for AI insurance through AIUC, once certified." Our own recorded value reads: *conditional eligibility only, not cover in force.* 3. **ElevenLabs Agents** *(agent)* — the same sentence, from the same URL, citing **the same evidence row** (`evidence_id` 8176). 4. **RentAHuman (workman.rent)** *(agent)* — "Plus requester reputation and incident insurance funded from the fee." A marketplace pool funded from transaction fees.

**Rows 2 and 3 are the same entity, published twice.** `elevenagents` (id 109, first seen 08-01) and `elevenlabs-agents` (id 2879, first seen 08-05) share a homepage URL, a subtype, and a single evidence row.

Deduplicating the field checked, the honest count is **three distinct entities** with any `insurance_available` evidence: one insurer selling cyber cover, one platform stating conditional eligibility that is expressly not cover in force, and one marketplace incident pool.

**Zero published AI agents disclose insurance cover in force from a named carrier.** Not a low rate — none.

**`insurance_carriers` — 1 row:** Coalition again, naming "Allianz''s A+ rated capacity" as backing for its own product.

**`indemnification` — 3 rows,** and one runs backwards:

- **Gemini Code Assist** — genuine vendor-to-customer IP indemnity. - **AgentCard** — "The legal responsibility sits with us, not you." - **ChatGPT for Tinder and Bumble** — "By using the extension, you agree to indemnify its creator from any liability." This is the *user* indemnifying the *vendor*, the opposite of what the field asks. Our stored value says so explicitly, which is the Fact-Checker doing its job, but the field still counts as present.

**Two of 776 published entities disclose a vendor-to-customer indemnity.**

**`liability_cap` — 2 rows:** Agent shield ("Our liability is limited to the engagement fee paid") and NOFireAI, whose terms cap total aggregate liability at **$100**.

**`own_liability_cover` — 2 rows:** Agent shield ("We carry professional liability insurance") and Edwin AI, via LogicMonitor''s trust centre ("Has cyber insurance"). Both are the vendor''s own corporate cover, not cover for the agent''s outputs. Vol. 4 reported zero on this field; these are the first two.

---

## 6. What the scorecards say

Judged `v1` scorecards exist for **69 of 776 published entities (8.89%)**. Trust gap means a score of 0 or 1.

| Dimension | Entities | Trust gap (0–1) | Rate | 0 | 1 | 2 | 3 | |---|---|---|---|---|---|---|---| | `insurance_indemnity` | 69 | 69 | **100.00%** | 65 | 4 | 0 | 0 | | `outcome_settlement` | 69 | 68 | 98.55% | 66 | 2 | 1 | 0 | | `audit_trail` | 69 | 57 | 82.61% | 42 | 15 | 11 | 1 | | `compliance` | 69 | 54 | 78.26% | 39 | 15 | 15 | 0 | | `runtime_governance` | 69 | 44 | 63.77% | 19 | 25 | 25 | 0 |

`insurance_indemnity` is at 100.00% — no judged published entity scores above 1. Vol. 4 reported 97.26% on a larger judged set; the entities that scored 1 and have since been demoted are the difference.

The single 3 in the table is `proofchain` on `audit_trail`, scored 2026-08-24. It remains the only 3 on a published entity in the index''s history. Vol. 3 and earlier claimed the top band was unreachable; it is reachable, and exactly one entity has reached it.

**A defect worth publishing:** the automatic scorer `v1-auto` has still never emitted a 3, on any dimension, on any entity. Its distribution is close to degenerate — on `audit_trail` it scored 374 entities and **351 of them (93.85%) got exactly 1**. Coverage is also lopsided: `runtime_governance` 417 entities, `audit_trail` 374, `compliance` 358, `outcome_settlement` 129, `insurance_indemnity` **7**. It is excluded from every headline figure in this report for the fifth consecutive volume, and the dimension our thesis most depends on is the one it almost never scores.

---

## 7. Where the data undercuts our own position

**The controls half of our thesis is now clearly wrong, and this is the third volume in which it has weakened.**

On the published index, 54.25% of entities document an audit trail, 48.07% document runtime governance, 47.81% document compliance certifications, and 57.60% document explainability. Under the judged rubric, 36.23% of scored entities clear a 2 on `runtime_governance`. "AI agents are ungoverned" is not a statement our own data supports, and we should stop implying it.

Two honest caveats cut in opposite directions. The floor selects for documentation, so these rates overstate the market — an entity is only published if we already found 8 facts about it. But the same floor applies to the risk-transfer fields, which stayed at 0.26% regardless.

Agents with **no** present value on any of the 24 assurance fields fell from 469 of 1,662 (28.22%) at Vol. 4 to **4 of 747 (0.54%)** today. That is almost entirely composition, and we flag it because it is the single most flattering number in this report and the least meaningful.

**What survives, stated as narrowly as the evidence allows:** vendors document how their agents are controlled and are close to silent on who pays when the controls fail. Not "agents are ungoverned" — *agents are governed and uninsured*. Those are different claims and only the second is ours to make.

Vendors also materially outperform agents on assurance disclosure: 7.14 present assurance fields on average against 3.69 for agents, with zero of 29 vendors carrying no assurance evidence at all.

---

## 8. What we cannot measure, and one statistic we are withdrawing

**2,444 of 45,494 slots (5.37%) have never been assessed** — no row exists at all. MIT hand-annotates a 30 × 45 grid so every cell has a verdict; we crawl continuously and do not. We report this as a third category rather than folding it into either column.

It is almost perfectly concentrated: four fields account for **2,436 of 2,444 (99.67%)**. `self_reported_performance` (610), `market_recognition` (609), `own_liability_cover` (609) and `asset_custody` (608) all arrived in migration 0027 on 2026-08-19 and have been assessed on roughly 21% of published entities. The remaining 8 are vendor fields on two entities promoted this week. Every percentage in this report that could be affected is stated against both `applicable` and `assessed` denominators.

### The false-absence correction rate is not reproducible, and we are not restating it

Vol. 4 reported that 85 of 1,044 high-confidence false-absence flags on published entities had been reviewed, and that 59 of those 85 (69.41%) were corrections. Today the table shows, across **all** entities regardless of status: 2,598 flags, 50 reviewed, and **zero** carrying the verdict `value_corrected`.

That is not because the corrections were undone. It is a stock-versus-flow error, and it is ours. `refresh_false_absence_flags()` clears rows whose absence no longer holds — we observed a live run reporting `cleared: 1`. A flag whose verdict is `value_corrected` describes a field that is no longer an absence, so the next refresh removes it. The corrected rows delete themselves, and the surviving population is biased toward `absence_confirmed` by construction.

**A correction rate cannot be computed from this table, and Vol. 4''s 69.41% should not be cited.** We are not asserting the figure was wrong when published — it was measured against rows that were genuinely there — but it is not reproducible, and an unreproducible statistic in a report whose appendix promises reproducibility does not belong in the record.

On the current stock: 955 flags on published entities, 31 reviewed (3.25%), all 31 `absence_confirmed`. The review backlog remains the largest known source of error in everything above, and it is now larger relative to throughput than it was a week ago.

---

## 9. Vol. 4''s handoffs, and what happened to them

**Closed.** Vol. 4 audited all 8 `present` rows on `liability_cap` and found that 7 were misfiled — five blanket liability exclusions, one third-party demo output, and one German corporate form (`UG (haftungsbeschränkt)`) lifted from a copyright footer. All seven were corrected to `no_public_information` by `factchecker-cloud` on 2026-08-25, within a day of publication. Agent shield survived the audit and remains. NOFireAI was added 2026-08-31 as a second genuine cap.

The report changed the data it reported on. That is the loop working.

**Open, and now a volume older.** Vol. 4''s highest-priority handoff was that `lloyds-of-london`, `chaucer-group` and `mosaic-insurance` were published as `no_public_information` across the risk-transfer fields. On `insurance_indemnity` a false absence is indistinguishable from a real 0, so the scorecard prints the opposite of the truth about three of the best-known names in the market we claim to cover.

**New, and small enough to fix today.** `elevenagents` and `elevenlabs-agents` are the same product published twice, sharing one evidence row. It is the only duplicate homepage among 776 published entities (0.13%), so this is not a systemic defect — but of the four rows supporting our single most important field, two are the same row, and the duplicate inflates that field by 25%.

---

## 10. Method, and what this report does not do

Every figure above comes from a query in the appendix, run against `invsbcblyrjsvmygulcz` on 2026-08-31. Nothing is estimated, extrapolated, or rounded from a number we did not compute.

Specifically, we did **not**:

- extrapolate the 3.25% reviewed sample to the 924 unreviewed flags; - restate Vol. 4''s correction rate in either direction; - report the 66.86% headline without its decomposition; - report the fall in zero-assurance agents as market improvement.

Three analyses remain deferred, for the same reason as in Vol. 4 and now with a smaller corpus behind them:

- **Autonomy versus accountability.** `autonomy_level` mixes a numeric scale with free prose across dozens of distinct strings; any ordinal mapping would drive the correlation more than the market would. Fourth deferral. - **Model concentration risk.** `model_provider` and `base_models` are free text. - **The assurance landscape map.** 29 published vendors is enough to list and not enough to map without padding. It was 42 before the floor, and those 13 are the highest-value enrichment targets in the index.

---

## Appendix — reproduction queries

Every query below was run against Supabase project `invsbcblyrjsvmygulcz` on 2026-08-31. They are reproduced verbatim so any figure in this report can be independently recomputed.

**A1 — published population (§1)**

```sql select status, count(*) from entities group by status order by 2 desc; select entity_type, count(*) from entities where status='published' group by entity_type; ```

**A2 — field catalogue shape (§1, §4)**

```sql select category, applies_to, count(*) from field_catalog group by 1,2 order by 1,2; -- 45 fields apply to both types, 14 to agents only, 4 to vendors only. -- An agent is assessed on 59 fields, a vendor on 49. ```

**A3 — reconstructing the Vol. 4 cohort by reverse-replaying the status log (§1)**

```sql with t as (select timestamptz '2026-08-24 20:10:00+00' as cut), first_after as ( select distinct on (l.entity_id) l.entity_id, l.old_status from entity_overwrite_log l, t where l.changed_at > t.cut and l.old_status is distinct from l.new_status order by l.entity_id, l.changed_at asc) select coalesce(fa.old_status, e.status) as status_at_v4, count(*) from entities e left join first_after fa on fa.entity_id = e.id group by 1 order by 2 desc; -- returns published = 1704, matching Vol. 4's independently reported figure ```

**A4 — entity flow, Vol. 4 to now (§1)**

```sql with demoted_0826 as ( select distinct entity_id from entity_overwrite_log where changed_at::date = '2026-08-26' and old_status = 'published' and new_status = 'candidate'), repromoted as ( select distinct l.entity_id from entity_overwrite_log l join demoted_0826 d on d.entity_id = l.entity_id where l.changed_at > '2026-08-26 23:59:59+00' and l.new_status = 'published') select (select count(*) from demoted_0826) as demoted_on_0826, (select count(*) from repromoted) as re_promoted_since, (select count(*) from demoted_0826) - (select count(*) from repromoted) as net_still_demoted, (select count(*) from entities where status='published') as published_now; -- 948, 18, 930, 776 -- of the 948, 943 were demoted by the publication floor and 5 withdrawn -- under a separate editorial decision; split by subtype in the same log ```

**A5 — slot accounting, current published cohort (§2, §8)**

```sql with pub as (select id, entity_type from entities where status='published'), slots as ( select p.id, p.entity_type, f.field_key from pub p join field_catalog f on f.applies_to in ('both', p.entity_type)) select count(*) as applicable_slots, count(*) filter (where ef.value_status='present') as present, count(*) filter (where ef.value_status='no_public_information') as no_public_information, count(*) filter (where ef.value_status='not_applicable') as not_applicable, count(*) filter (where ef.id is null) as never_assessed, round(100.0*count(*) filter (where ef.value_status='no_public_information')/count(*),2) as npi_pct from slots s left join entity_fields ef on ef.entity_id = s.id and ef.field_key = s.field_key; -- 45494 slots | 12632 present | 30418 npi | 0 n/a | 2444 never | 66.86% ```

**A6 — the same accounting on the Vol. 4 cohort, current data (§2)**

```sql with t as (select timestamptz '2026-08-24 20:10:00+00' as cut), first_after as ( select distinct on (l.entity_id) l.entity_id, l.old_status from entity_overwrite_log l, t where l.changed_at > t.cut and l.old_status is distinct from l.new_status order by l.entity_id, l.changed_at asc), v4 as (select e.id, e.entity_type from entities e left join first_after fa on fa.entity_id = e.id where coalesce(fa.old_status, e.status) = 'published'), slots as (select v.id, f.field_key from v4 v join field_catalog f on f.applies_to in ('both', v.entity_type)) select count(*) as applicable_slots, count(*) filter (where ef.value_status='present') as present, count(*) filter (where ef.value_status='no_public_information') as npi, count(*) filter (where ef.id is null) as never_assessed, round(100.0*count(*) filter (where ef.value_status='no_public_information')/count(*),2) as npi_pct from slots s left join entity_fields ef on ef.entity_id = s.id and ef.field_key = s.field_key; -- 100096 slots | 20147 present | 74322 npi | 5627 never | 74.25% -- -- Decomposition: 66.86 - 73.37 = -6.51 total -- 66.86 - 74.25 = -7.39 composition -- 74.25 - 73.37 = +0.88 new evidence ```

**A7 — the 20-slot reconciliation (§2)**

```sql select old_entity_type, new_entity_type, count(*), min(changed_at) from entity_overwrite_log where changed_at > '2026-08-24 20:10:00+00' and old_entity_type is distinct from new_entity_type group by 1,2; -- 2 agent -> vendor on 2026-08-28; 2 x (59-49) = 20 slots ```

**A8 — retraction versus discovery on the Vol. 4 cohort (§2)**

```sql with t as (select timestamptz '2026-08-24 20:10:00+00' as cut), first_after as ( select distinct on (l.entity_id) l.entity_id, l.old_status from entity_overwrite_log l, t where l.changed_at > t.cut and l.old_status is distinct from l.new_status order by l.entity_id, l.changed_at asc), v4 as (select e.id from entities e left join first_after fa on fa.entity_id = e.id where coalesce(fa.old_status, e.status) = 'published') select ef.value_status, ef.extracted_by, count(*) filter (where ef.created_at > '2026-08-24 20:10:00+00') as newly_created, count(*) filter (where ef.created_at <= '2026-08-24 20:10:00+00') as updated_in_place, count(*) as rows_touched from entity_fields ef join v4 on v4.id = ef.entity_id where ef.updated_at > '2026-08-24 20:10:00+00' group by 1,2 order by 5 desc; -- 838 rows updated in place to no_public_information; 190 to present ```

**A9 — retained versus demoted evidence density (§3)**

```sql with t as (select timestamptz '2026-08-24 20:10:00+00' as cut), first_after as ( select distinct on (l.entity_id) l.entity_id, l.old_status from entity_overwrite_log l, t where l.changed_at > t.cut and l.old_status is distinct from l.new_status order by l.entity_id, l.changed_at asc), cohort as (select e.id, e.entity_type, case when coalesce(fa.old_status,e.status)='published' and e.status='published' then 'retained' when coalesce(fa.old_status,e.status)='published' and e.status<>'published' then 'demoted' when coalesce(fa.old_status,e.status)<>'published' and e.status='published' then 'newly_promoted' end as grp from entities e left join first_after fa on fa.entity_id = e.id), slots as (select c.id, c.grp, f.field_key from cohort c join field_catalog f on f.applies_to in ('both', c.entity_type) where c.grp is not null) select s.grp, count(distinct s.id) as entities, count(*) as applicable_slots, count(*) filter (where ef.value_status='present') as present, round(100.0*count(*) filter (where ef.value_status='no_public_information')/count(*),2) as npi_pct, round(count(*) filter (where ef.value_status='present')::numeric/count(distinct s.id),2) as present_fields_per_entity from slots s left join entity_fields ef on ef.entity_id = s.id and ef.field_key = s.field_key group by s.grp order by entities desc; ```

**A10 — the 24 assurance fields (§4)**

```sql with pub as (select id, entity_type from entities where status='published'), slots as (select p.id, f.field_key from pub p join field_catalog f on f.applies_to in ('both', p.entity_type) where f.category = 'assurance') select s.field_key, count(*) as applicable, count(*) filter (where ef.value_status='present') as present, count(*) filter (where ef.value_status='no_public_information') as npi, count(*) filter (where ef.id is null) as never_assessed, round(100.0*count(*) filter (where ef.value_status='present')/count(*),2) as present_pct from slots s left join entity_fields ef on ef.entity_id = s.id and ef.field_key = s.field_key group by s.field_key order by present_pct desc, s.field_key; ```

**A11 — controls versus risk transfer on identical denominators (§4)**

```sql with pub as (select id from entities where status='published'), blocks(block, field_key) as (values ('controls','runtime_governance'),('controls','permission_scopes'),('controls','audit_trail'), ('controls','audit_trail_immutable'),('controls','explainability'),('controls','compliance_certs'), ('risk_transfer','insurance_available'),('risk_transfer','own_liability_cover'), ('risk_transfer','insurance_carriers'),('risk_transfer','coverage_limits'), ('risk_transfer','indemnification'),('risk_transfer','liability_cap')), slots as (select p.id, b.block, b.field_key from pub p cross join blocks b) select s.block, count(*) as applicable_slots, count(*) filter (where ef.id is not null) as assessed_slots, count(*) filter (where ef.value_status='present') as present, round(100.0*count(*) filter (where ef.value_status='present')/count(*),2) as pct_of_applicable from slots s left join entity_fields ef on ef.entity_id = s.id and ef.field_key = s.field_key group by s.block order by s.block; -- controls 4656 / 4656 / 1951 / 41.90% -- risk_transfer 4656 / 4047 / 12 / 0.26% 1951 / 12 = 162.58 ```

**A12 — the complete risk-transfer census (§5)**

```sql select e.slug, e.name, e.entity_type, ef.field_key, ef.value_text, ef.quote, ef.extracted_by, ef.evidence_id, ev.source_url, ev.retrieved_at::date from entity_fields ef join entities e on e.id = ef.entity_id and e.status = 'published' join evidence ev on ev.id = ef.evidence_id where ef.value_status = 'present' and ef.field_key in ('insurance_available','insurance_carriers','coverage_limits', 'liability_cap','indemnification','own_liability_cover') order by ef.field_key, e.slug; -- 12 rows. elevenagents and elevenlabs-agents both cite evidence_id 8176. ```

**A13 — the duplicate check (§5, §9)**

```sql with pub as (select id, slug, lower(regexp_replace(coalesce(homepage_url,''),'/+$','')) as u from entities where status='published' and coalesce(homepage_url,'') <> '') select count(*) as published_with_homepage, count(distinct u) as distinct_homepages, count(*) - count(distinct u) as surplus_rows from pub; -- 776 / 775 / 1

with pub as (select id, slug, lower(regexp_replace(coalesce(homepage_url,''),'/+$','')) as u from entities where status='published' and coalesce(homepage_url,'') <> '') select u, count(*), string_agg(slug, ', ' order by id) from pub group by u having count(*) > 1; -- https://elevenlabs.io/agents | 2 | elevenagents, elevenlabs-agents ```

**A14 — scorecards (§6)**

```sql select sc.rubric_version, sc.dimension, count(*) as n_entities, count(*) filter (where sc.score <= 1) as trust_gap_0_1, round(100.0*count(*) filter (where sc.score <= 1)/count(*),2) as trust_gap_pct, count(*) filter (where sc.score=0) as s0, count(*) filter (where sc.score=1) as s1, count(*) filter (where sc.score=2) as s2, count(*) filter (where sc.score=3) as s3 from scorecards sc join entities e on e.id = sc.entity_id and e.status = 'published' group by 1,2 order by 1, trust_gap_pct desc; ```

**A15 — entities with no assurance evidence at all (§7)**

```sql with pub as (select id, entity_type from entities where status='published'), a as (select p.id, p.entity_type, count(*) filter (where ef.value_status='present') as assurance_present from pub p join field_catalog f on f.applies_to in ('both', p.entity_type) and f.category='assurance' left join entity_fields ef on ef.entity_id = p.id and ef.field_key = f.field_key group by p.id, p.entity_type) select entity_type, count(*) as entities, count(*) filter (where assurance_present = 0) as zero_assurance_evidence, round(avg(assurance_present),2) as mean_assurance_fields_present from a group by entity_type; -- agent 747 / 4 / 3.69 vendor 29 / 0 / 7.14 ```

**A16 — where the never-assessed slots sit (§8)**

```sql with pub as (select id, entity_type from entities where status='published'), slots as (select p.id, f.field_key, f.category from pub p join field_catalog f on f.applies_to in ('both', p.entity_type)) select s.field_key, s.category, count(*) filter (where ef.id is null) as never_assessed from slots s left join entity_fields ef on ef.entity_id = s.id and ef.field_key = s.field_key group by 1,2 having count(*) filter (where ef.id is null) > 0 order by never_assessed desc; -- 4 fields from migration 0027 account for 2436 of 2444 (99.67%) ```

**A17 — the false-absence review stock, and why a correction rate cannot be derived (§8)**

```sql select case when e.status='published' then 'published' else e.status end as status, count(*) as flags, count(*) filter (where f.reviewed_at is not null) as reviewed, count(*) filter (where f.verdict='value_corrected') as corrected, count(*) filter (where f.verdict='absence_confirmed') as confirmed from false_absence_flags f join entities e on e.id = f.entity_id group by 1 order by flags desc; -- published 955 / 31 / 0 / 31 ; candidate 1642 / 19 / 0 / 19 ; archived 1 / 0 / 0 / 0

select job_name, started_at, ok, detail from maintenance_runs where ok is true and detail is not null order by started_at desc limit 8; -- a live run of refresh_false_absence_flags reporting "cleared": 1 — corrected -- rows leave the table, so the surviving stock is biased to absence_confirmed ```

**A18 — Vol. 4''s liability_cap handoff, after the fact (§9)**

```sql select e.slug, e.status, ef.value_status, ef.extracted_by, ef.updated_at::date, coalesce(ef.value_text, ef.quote) as val from entity_fields ef join entities e on e.id = ef.entity_id where ef.field_key = 'liability_cap' and e.slug in ('agent-shield','agentman','913-ai','ai-synth-id-remover','cortx', 'deckdrop-io','free-ai-humanizer','stripe','nofireai') order by e.slug; -- 7 of the 8 rows Vol. 4 audited were corrected to no_public_information -- by factchecker-cloud on 2026-08-25; agent-shield stands; nofireai added 08-31 ```