The Trust Gap, Vol. 3: we audited our own absences, and the correction runs against us everywhere except insurance

# The Trust Gap, Vol. 3

**aidispatch.news index — measured 2026-08-17 against the production database.**

Every number in this report comes from a query printed in the appendix. Nothing is estimated, extrapolated, or rounded from a figure we did not compute.

---

## 0. Why this volume exists

Vol. 1 and Vol. 2 measured what the index does not know: the share of our published data points that record `no_public_information`. That number is the spine of this publication, and it rests on an assumption we had never tested — that when we publish an absence, the absence is true.

On 2026-08-16 our own Fact-Checker found five published fields on one entity recording `no_public_information` for things the entity's own stored capture plainly documented. On 2026-08-17 a detector shipped that makes this class measurable across the whole index.

This volume is the first measurement. It is also the first time we have found our own headline number to be wrong in the direction that flatters us less on one dimension and more on another — and we report both.

---

## 1. The headline, stated the way MIT states theirs

Across **1,705 published entities** (1,664 agents, 41 vendors) and **93,365 applicable data points**:

| Value status | Count | Share | |---|---|---| | `no_public_information` | **72,727** | **77.90%** | | `present` | 20,558 | 22.02% | | no row written at all | 80 | 0.09% | | `not_applicable` | 0 | 0.00% |

MIT's AI Agent Index reports **227 of 1,350 data points (16.81%)** lacking public information. Ours is 77.90%.

**These two numbers are not comparable, and the contrast does no work for us.** MIT hand-annotated 30 agents chosen for prominence, with seven expert annotators, across 45 fields selected to be answerable. We machine-extract 1,705 entities of wildly varying prominence across 59 fields, 21 of which ask about an assurance layer that barely exists as a disclosure practice. A long-tail agent with a one-page marketing site will answer almost nothing. If we had MIT's 30 entities and MIT's 45 fields our number would be far lower, and it still would not mean the same thing.

The denominator is **applicable** slots — `field_catalog.applies_to` matched to entity type — not a flat 59 × 1,705. A flat denominator would inflate the absence rate. This is the easy way to get this statistic wrong and it is worth saying out loud.

---

## 2. The risk-transfer collapse, unchanged

The five fields that carry our thesis, on published entities:

| Field | `present` | Absent | Present, as share of 1,705 | |---|---|---|---| | `coverage_limits` | **0** | 1,704 | **0.00%** | | `indemnification` | 5 | 1,697 | 0.29% | | `liability_cap` | 6 | 1,696 | 0.35% | | `insurance_available` | 6 | 1,698 | 0.35% | | `insurance_carriers` | 10 | 1,693 | 0.59% | | **Block total** | **27 of 8,525** | | **0.32%** |

`coverage_limits` remains the only field in the 59-field catalogue at 100.00% absence. Not one published entity in the index discloses a policy limit.

**1,699 of 1,705 published entities (99.65%) have no public evidence of insurance availability.**

The six that do, in full — this is the entire population, not a selection:

| Entity | Type | Stored quote (truncated) | |---|---|---| | `agent-shield` | agent | "We carry professional liability insurance." | | `chatgpt-for-tinder-and-bumble` | agent | "you agree to indemnify its creator from any liability" | | `rentahuman-workman-rent` | agent | "Plus requester reputation and incident insurance funded from the fee." | | `armilla-ai` | vendor | "Armilla provides named, affirmative AI insurance that responds to those failure modes directly" | | `coalition` | vendor | "Cyber Insurance | Active Insurance & Cybersecurity | Coalition…" | | `munich-re-aisure` | vendor | "AI insurance, pioneered by Munich Re in 2018, is triggered by unexpected errors…" |

Note the second row. It is filed under `insurance_available`, and it is not evidence that insurance is available — it is the user agreeing to indemnify the developer. That is a mis-filing, it is live on the public API right now, and we found it while auditing ourselves. It is listed as a correction in §6.

---

## 3. The judged rubric

Scorecards written under the judged `v1` rubric, published entities only:

| Dimension | Entities | Score 0–1 | Share | Score 0 | Max observed | |---|---|---|---|---|---| | `outcome_settlement` | 60 | **60** | **100.00%** | 58 | 1 | | `insurance_indemnity` | 60 | 58 | 96.67% | 51 | 2 | | `compliance` | 60 | 56 | 93.33% | 45 | 2 | | `audit_trail` | 59 | 51 | 86.44% | 44 | 2 | | `runtime_governance` | 60 | 42 | 70.00% | 31 | 2 |

**No entity in the index has ever scored 3 on any dimension.** `outcome_settlement` has never exceeded 1 across 60 readings.

**n = 60 is 3.52% of published entities and the entities were purposively chosen, not sampled.** This is not a population rate and must not be read as one. It is attached to the number here rather than buried in a footnote because Vol. 1 was rejected for exactly this kind of slippage.

The automated `v1-auto` rubric is excluded from every statistic in this report. Across **1,387 rows** on published entities its observed range is exactly {1, 2}: **zero 0s and zero 3s**. It is structurally incapable of emitting the modal judged score. An instrument that cannot express the finding the index exists to make is not evidence, and averaging it with judged scores would manufacture a more flattering picture of the industry than our own analysts found.

---

## 4. The new measurement: how wrong are our absences?

An absence in this index is a published, citable finding: *we retrieved this page and it does not say*. Three guardrails check the shape of such a row — that it carries no claim, that it cites a document long enough to have been read, that any quote is verbatim. **None of them compares the absence against what the cited document actually says.**

A detector now flags absences whose own stored capture matches a field-specific marker pattern. Across the **72,738** absence rows on published entities:

| | Count | Share of absences | |---|---|---| | Flagged as suspect | **1,859** | **2.56%** | | Flagged, high-confidence tier | 1,047 | 1.44% | | Fields with a marker defined | 55 of 59 | — | | Fields never checked | **4** | — |

The four never-checked fields are reported as unchecked, not as clean: `documented_incidents` (1,684 absences), `autonomy_level` (1,608), `goal_complexity` (1,317), `target_sectors` (682). No defensible marker exists for them. "Zero flagged" and "never checked" are different facts.

### 4.1 We read 75 of them by hand

A flag is a prompt to look, never proof. So we read every high-confidence flag on published entities in three field groups — the seven risk-transfer and vendor-assurance fields, `compliance_certs`, and `audit_trail` — and adjudicated each against its stored capture. **This is a complete census of those nine fields, not a sample of the other 972 high-confidence flags.**

| Verdict | Rows | Share | |---|---|---| | **Confirmed false absence** — the capture does disclose it | **58** | **77.33%** | | False positive — the absence is correct | 13 | 17.33% | | Disclosed, but not the thing the field asks for | 3 | 4.00% | | Could not adjudicate from the stored capture | 1 | 1.33% | | **Total audited** | **75** | 100% |

**Precision varies enormously by field, and the pattern is uncomfortable:**

| Field group | Audited | Confirmed false | Precision | |---|---|---|---| | `audit_trail` | 27 | 25 | **92.59%** | | `compliance_certs` | 34 | 26 | 76.47% | | Risk transfer (7 fields) | 14 | 7 | **50.00%** |

The detector is most accurate exactly where a correction **undercuts** our thesis, and least accurate where a correction would **support** it. We would have preferred the opposite and it is not what the data says.

### 4.2 The false positives have a single dominant cause

Of the 6 false positives in the risk-transfer block, **4 are contract-review agents whose demo output names indemnity and liability clauses in somebody else's contract** — `agentman` ("Indemnification Mutual clause confirmed"), `claude-cowork` (three separate fields, from a rendered clause-by-clause review listing "§14.1 · Indemnification", "§11.2 · Limitation of liability").

An agent that reads contracts for a living is not thereby insured. The detector cannot tell the difference between a page describing the vendor's own terms and a page demonstrating the vendor's product on someone else's terms. This failure mode was predicted in the detector's own design notes before it shipped, and the audit confirms it as the dominant class.

The remaining two: `orloj`, where "policy limits" refers to per-agent token spend caps, and `kay-ai`, where "E&O risk" names a risk the product claims to *reduce*, not cover it carries.

### 4.3 What correcting the confirmed rows actually does

| Field | Present now | After correction | Share of 1,705 | |---|---|---|---| | `audit_trail` | 460 | **485** | 26.98% → 28.45% | | `compliance_certs` | 406 | **432** | 23.81% → 25.34% | | `insurance_available` | 6 | **7** | 0.35% → 0.41% | | Risk-transfer block (5 fields) | 27 | **32** | 0.32% → 0.38% | | Entities with no insurance evidence | 1,699 | **1,698** | 99.65% → **99.59%** |

And the finding that matters most:

> **Of the five confirmed false absences in the core risk-transfer block, four > are insurance companies** — `embroker`, `lloyds-of-london`, `chaucer-group`, > `mosaic-insurance` — **that we had published as disclosing no insurance > information.** The fifth is the mis-filed indemnity clause in §2, which runs > from the user to the developer. > > **Not one of the 1,664 published agents gains insurance cover from this > correction. The count of agents with any public evidence of insurance > availability is 3 before the audit and 3 after it — 0.18%.**

We corrected our own data in the direction that makes us look worse at the vendor level, and the agent-level trust gap survived the correction intact.

---

## 5. What this does to our standing thesis

Our position has been that AI agents "cannot be insured, audited, governed, or paid on verified outcomes." Vol. 2 already found the *audited* and *governed* halves too strong. This volume pushes further in the same direction.

**Against us:** `audit_trail` disclosure is not 26.98% but at least 28.45%, and that is a floor, not a ceiling — the detector checks one marker pattern per field and reads only the single stored capture. `compliance_certs` is at least 25.34%. Twenty-five published entities describe an audit log on their own site while we publish that they disclose nothing about one; twenty-six name SOC 2, ISO 27001, HIPAA or PCI DSS while we publish the same. On the two dimensions where the industry is doing better than we said, **our instrument was systematically understating them, and the error was ours.**

**For us:** the risk-transfer collapse does not move. `coverage_limits` stays at exactly zero. The corrected block fill rate is 0.38%. The entities whose absences we got wrong are, overwhelmingly, the insurers themselves — which is a statement about our extraction quality on vendor pages, not about the market.

The honest summary is narrower than our thesis and better supported by it: **vendors increasingly describe the controls, and remain almost entirely silent on who pays when the controls fail.** Governance is a feature, so it is marketed. Risk transfer is a liability, so it is not.

---

## 6. Corrections and open defects, published rather than quietly fixed

1. **`chatgpt-for-tinder-and-bumble` / `insurance_available` is mis-filed.** The stored quote is a user-grants-developer indemnity. It is not evidence of insurance availability. Live on the public API now. → Fact-Checker. 2. **Three flagship insurance vendors are still published as disclosing no insurance information**: `lloyds-of-london`, `embroker`, `chaucer-group`. `munich-re-aisure` was corrected on 2026-08-17. → Fact-Checker, highest priority; these are the entities the index exists to cover. 3. **`lloyds-of-london` is not flagged on `insurance_available` at all** — only on `insurance_carriers` — because the marker for the former does not match the language Lloyd's uses about itself. The 1,859 flag count is therefore a **lower bound** on false absences, and per-field marker coverage is uneven. Reported here because a detector's blind spots are part of its result. 4. **58 confirmed false absences are enumerated by appendix queries A11–A13** and can be regenerated and disagreed with row by row.

---

## 7. Scope limits

- 60 judged scorecards is 3.52% of published entities, purposively chosen. - The hand audit covers 75 of 1,047 high-confidence flags. **The 77.33% confirmation rate applies to the nine fields audited and must not be applied to the remaining 972 flags**, whose field mix is different and whose precision we have not measured. - 4 of 59 fields have no false-absence marker and are unchecked. - The detector reads only the single capture cited by the row. An entity that discloses cover on a page we never retrieved is recorded as absent and is not flagged. Our central claim remains "not on the page we captured," which is weaker than "not published anywhere." - 11 absence rows on published entities sit on fields that do not apply to their entity type; they are excluded from the §1 denominator and included in the §4 detector population, which is why those two totals differ by 11 (72,727 vs 72,738).

---

## Appendix — reproduction queries

Run against Supabase project `invsbcblyrjsvmygulcz` on 2026-08-17.

**A1 — population** ```sql SELECT status, entity_type, count(*) FROM entities GROUP BY 1,2 ORDER BY 1,2; ```

**A2 — §1 headline, applicable-slot denominator** ```sql WITH pub AS (SELECT id, entity_type FROM entities WHERE status='published'), slots AS ( SELECT p.id AS entity_id, f.field_key FROM pub p JOIN field_catalog f ON f.applies_to = 'both' OR f.applies_to = p.entity_type ) SELECT (SELECT count(*) FROM pub) AS published_entities, (SELECT count(*) FROM slots) AS applicable_slots, count(ef.id) FILTER (WHERE ef.value_status='present') AS present, count(ef.id) FILTER (WHERE ef.value_status='no_public_information') AS npi, count(ef.id) FILTER (WHERE ef.value_status='not_applicable') AS na, (SELECT count(*) FROM slots) - count(ef.id) AS slots_with_no_row FROM slots s LEFT JOIN entity_fields ef ON ef.entity_id=s.entity_id AND ef.field_key=s.field_key; ```

**A3 — §2 risk-transfer block** ```sql WITH pub AS (SELECT id FROM entities WHERE status='published') SELECT ef.field_key, count(*) FILTER (WHERE ef.value_status='present') AS present, count(*) FILTER (WHERE ef.value_status='no_public_information') AS absent FROM entity_fields ef JOIN pub p ON p.id=ef.entity_id WHERE ef.field_key IN ('insurance_available','insurance_carriers', 'coverage_limits','indemnification','liability_cap') GROUP BY 1 ORDER BY 1; ```

**A4 — §2 the six entities with insurance evidence (complete population)** ```sql SELECT e.slug, e.entity_type, ef.quote FROM entity_fields ef JOIN entities e ON e.id=ef.entity_id WHERE ef.field_key='insurance_available' AND ef.value_status='present' AND e.status='published' ORDER BY e.entity_type, e.slug; ```

**A5 — §3 judged and automated rubric distribution** ```sql SELECT s.rubric_version, s.dimension, count(*) AS rows, count(DISTINCT s.entity_id) AS entities, count(*) FILTER (WHERE s.score<=1) AS score_0_1, count(*) FILTER (WHERE s.score=0) AS score_0, count(*) FILTER (WHERE s.score=3) AS score_3 FROM scorecards s JOIN entities e ON e.id=s.entity_id WHERE e.status='published' GROUP BY 1,2 ORDER BY 1,2; ```

**A6 — §4 detector totals and coverage** ```sql SELECT sum(absences_published) AS absences_published_total, sum(flagged_published) AS flagged_published, sum(flagged_published_high) AS flagged_published_high, count(*) FILTER (WHERE markers_defined=0) AS fields_never_checked FROM false_absence_summary; ```

**A7 — §4 the four never-checked fields** ```sql SELECT category, field_key, label, absences_published FROM false_absence_summary WHERE markers_defined=0 ORDER BY absences_published DESC; ```

**A8 — §4 per-field flag rate, assurance category** ```sql SELECT field_key, absences_published, flagged_published, flagged_published_high, round(100.0*flagged_published/nullif(absences_published,0),2) AS pct_flagged FROM false_absence_summary WHERE category='assurance' ORDER BY flagged_published DESC; ```

**A9 — §4.3 present counts before correction** ```sql WITH pub AS (SELECT id FROM entities WHERE status='published') SELECT ef.field_key, count(*) FILTER (WHERE ef.value_status='present') AS present, count(*) FILTER (WHERE ef.value_status<>'present') AS absent FROM entity_fields ef JOIN pub p ON p.id=ef.entity_id WHERE ef.field_key IN ('audit_trail','compliance_certs','runtime_governance', 'permission_scopes','outcome_based_pricing','settlement_mechanism', 'audit_trail_immutable') GROUP BY 1 ORDER BY 1; ```

**A10 — §7 the 11-row reconciliation** ```sql WITH pub AS (SELECT id, entity_type FROM entities WHERE status='published') SELECT count(*) AS absence_rows_on_inapplicable_fields FROM entity_fields ef JOIN pub p ON p.id=ef.entity_id JOIN field_catalog f ON f.field_key=ef.field_key WHERE ef.value_status<>'present' AND f.applies_to <> 'both' AND f.applies_to <> p.entity_type; ```

**A11–A13 — the 75 hand-audited rows.** Change the `field_key` filter to `compliance_certs` (A12) or `audit_trail` (A13) to regenerate each group. ```sql SELECT s.entity_slug, s.entity_type, s.field_key, SUBSTRING(r.content FROM GREATEST(1, regexp_instr(r.content, s.matched_pattern,1,1,0,'i') - 200) FOR 460) AS ctx, s.source_url FROM suspected_false_absence s JOIN entity_fields ef ON ef.id = s.entity_field_id JOIN evidence ev ON ev.id = ef.evidence_id JOIN raw_documents r ON r.id = ev.raw_document_id WHERE s.entity_status='published' AND s.confidence='high' AND s.field_key IN ('insurance_available','insurance_carriers', 'coverage_limits','indemnification','liability_cap', 'vendor_regulatory_status','vendor_backing') -- A11 ORDER BY s.field_key, s.entity_slug; ```

---

*Methodology and the trust-gap rubric are published at `/methodology`. aidispatch.news is operated by aidispatch.news LLC. Trust-gap findings link to exact.works, a commercial assurance vendor; that relationship is disclosed on every page carrying it.*