The Trust Gap, Vol. 6: 18% of our published values no longer verify against their own source

Every number in this report comes from a query run against the index on 2026-09-07. The queries are reproduced in the appendix. Where a number could not be computed, this report says so rather than estimating it.

## 1. The census

The index holds 3,082 entities. 782 are published — 753 agents and 29 vendors — and 2,298 remain candidates. The field catalogue has 63 fields: 14 apply to agents only, 4 to vendors only, and 45 to both. An agent therefore has 59 applicable fields and a vendor 49.

That gives **45,848 applicable data points** across the published index:

| | Count | Share of all slots | |---|---:|---:| | Present, with a cited source | 12,705 | 27.71% | | `no_public_information` | 30,717 | **67.00%** | | Never assessed | 2,426 | 5.29% | | **Total applicable** | **45,848** | 100% |

**30,717 of 45,848 data points (67.00%) are recorded as having no public information.** Counting never-assessed slots as unknown too, 33,143 of 45,848 (72.29%) carry no published evidence. Of the 43,422 slots we have actually assessed, 70.74% came back empty.

We report this the way the MIT AI Agent Index reports its own gaps, because the absence is the finding. A field that no vendor documents is a fact about the market, not a hole in our sheet.

Assurance fields are emptier than the rest: 2,952 of 15,756 assurance slots are present (18.74%), against 9,753 of 30,092 (32.41%) elsewhere.

## 2. The rate went up, and it was not composition

Vol. 5 reported 66.86% against 776 published entities. It is 67.00% now.

To separate "we publish a different set of entities" from "the evidence changed," we reconstructed the Vol. 5 cohort by reverse-replaying every status change in `entity_overwrite_log` since that report. The reconstruction returns **exactly 776 entities and exactly 45,494 applicable slots** — both matching Vol. 5's published figures by a method Vol. 5 did not use. 776 − 7 demotions + 13 promotions = 782 reconciles to the current count.

Measured today, that same 776-entity cohort is at 67.02%.

| Term | Movement | |---|---:| | Like-for-like (same entities, one week later) | **+0.16pp** | | Composition (which entities we publish) | −0.02pp | | **Net** | **+0.14pp** |

The terms sum to the total exactly. Last week the story was that a 6.5-point improvement was entirely a selection effect. This week there is no improvement to explain away: **on a like-for-like population the index got slightly emptier, and the change of population did almost nothing.**

The reason is that almost nothing happened. Of the 45,494 field slots in the Vol. 5 cohort, **182 rows (0.40%) were touched in the whole week** — 94 landing on `present`, 88 on `no_public_information`. Only 26 rows were created at all, 20 of them absences. Publication continued; enrichment of what is already published very nearly stopped.

## 3. The finding: 18% of our published values no longer verify against their own source

This is the first volume to audit our central claim about ourselves. We beat MIT, we say, because their index is static and ours is continuously re-verified. That claim is testable, and it does not survive the test.

Of the 12,710 published values that carry a source quote, the crawler has since captured a **newer version of the source page for 6,725 of them**. For each, we checked whether the stored quote still appears in that newest capture, using the same whitespace-normalised verbatim test the database itself enforces on write:

| Checkable published values | 6,725 | |---|---:| | Quote still present in the newest capture | 5,506 (81.87%) | | **Quote no longer present** | **1,219 (18.13%)** |

Before reporting that, we ruled out the obvious artefact: a truncated or failed re-crawl would produce a missing quote without anything having changed. It does not explain this. For the 1,219 failures the newest capture is on average **longer** than the one originally cited (10,923 vs 9,848 characters, a ratio of 1.448). Only 259 sit on a materially shorter capture. **960 published values rest on a citation that no longer verifies against a capture that is not smaller than the original** — the conservative floor, 14.28% of everything checkable.

It is worth being precise about what this does and does not show. A vanished quote does not prove the underlying value is now false; a vendor may have reworded a page while the fact stayed true. It proves the **citation** no longer verifies. For an index whose entire promise is that every field carries a verbatim quote from a source document, a citation that no longer verifies is the product failing, whatever the fact underneath is doing.

The problem is systemic rather than concentrated in one corner:

| Category | Checkable | Quote gone | Rate | |---|---:|---:|---:| | Agency | 863 | 185 | 21.44% | | Impact | 985 | 200 | 20.30% | | **Assurance** | **1,567** | **280** | **17.87%** | | Practicality | 1,441 | 256 | 17.77% | | Safety | 664 | 109 | 16.42% | | Ecosystem | 806 | 129 | 16.00% | | Models | 399 | 60 | 15.04% |

280 of them are assurance values — the layer we exist to cover.

The underlying cause is that evidence rows are not re-pointed when a page is re-crawled. **92.05% of published present values cite evidence retrieved more than 30 days ago**, with a median age of 36.7 days against a corpus that only began on 2026-07-31. Meanwhile 497 of the 856 distinct source URLs (58.06%) *have* been re-fetched. The crawler is doing its job. The re-verification never propagates to the published field.

## 4. Why the detector we built for this reported 31

The schema already contains `superseded_quote_flags`, a table whose entire purpose is catching exactly this. It holds **31 rows**, none adjudicated. The refresh job is not broken: it has run **59 times with zero failures**, most recently at 20:05 UTC on the day of this report.

It is scoped wrong. The `superseded_quotes` view only compares a field's cited evidence against **other evidence rows the same entity already cites for the same URL**, gated on `having count(distinct content_hash) > 1`. It never looks at `raw_documents`, where the fresh captures actually live. So if the crawler re-fetched a page but no second evidence row was ever attached to that entity, the entity has one content hash for that URL and drops out of the view entirely.

**769 of 857 published (entity, source URL) pairs — 89.73% — are structurally invisible to it.** The detector is reporting green over roughly a tenth of the index.

This is the same failure mode this organisation has now hit three times: an instrument that returns a clean result for a reason unrelated to the question being asked, and is believed because its answer is the expected one.

One methodological note, because it nearly produced a wrong headline. Our first pass measured supersession by comparing `content_hash` between the cited and newest captures, which reported that 52.91% of published values sat on changed content. That number is discarded and should not be cited. `content_hash` is not a stable content fingerprint: of 1,748 URLs captured more than once, **exactly one** ever produced a repeated hash, and pages of byte-identical length routinely hash differently between captures. The hash detects noise. Only the quote test is sound, which is why every figure in §3 rests on the quote test.

## 5. Risk transfer: still the emptiest layer in the index

| Field | Present | Of 782 | No public evidence | |---|---:|---:|---:| | `insurance_available` | 4 | 0.51% | 99.49% | | `indemnification` | 3 | 0.38% | 99.62% | | `liability_cap` | 2 | 0.26% | 99.74% | | `own_liability_cover` | 2 | 0.26% | 99.74% | | `insurance_carriers` | 1 | 0.13% | 99.87% | | **`coverage_limits`** | **0** | **0.00%** | **100.00%** |

**776 of 782 published entities (99.23%) have no public evidence of insurance availability on any of the four insurance fields.**

`coverage_limits` has now been at 100.00% absence for six consecutive volumes. It remains the only field of 63 that has never once been satisfied by any entity in the index. Nobody publishes what an AI agent policy actually pays.

Because the total is twelve rows, we read all twelve individually again rather than reporting a rate. Doing so still changes the claim:

- **Two of the four `insurance_available` rows are the same product.** `elevenagents` (id 109) and `elevenlabs-agents` (id 2879) share a homepage, a field count, and **the same `evidence_id` 8176**. Three distinct entities, not four. This was Vol. 5's second handoff and is unresolved a second week. - Of those three: one is **Coalition, a cyber insurer describing its own product**; one is **conditional eligibility, expressly not cover in force** — our own stored value says "once certified"; one is a **marketplace incident pool funded from fees**. - **No published agent discloses insurance cover in force from a named carrier.** Not a low rate. None. - `indemnification` still runs backwards on `chatgpt-for-tinder-and-bumble`, where the *user* indemnifies the *vendor*. Two of 782 disclose a real vendor-to-customer indemnity. - `liability_cap`'s second row is NOFireAI capping total aggregate liability at **$100**.

The one comparison immune to every selection story, because the denominators are identical by construction — six risk-transfer fields and six control fields, all applicable to both entity types, across the same 782 entities, **4,692 slots each**:

| Band | Slots | Present | Rate | |---|---:|---:|---:| | Control (pricing, API, integrations, availability, deployment, support) | 4,692 | 2,540 | 54.13% | | Risk transfer | 4,692 | **12** | **0.26%** |

**211.7 to 1.** The same vendors who document their pricing and their integration surface in the majority of cases document who pays when the agent causes a loss in a quarter of one percent of cases. The risk-transfer count has not moved off 12 since Vol. 5.

## 6. What the scores say, once the two rubrics are separated

Every published entity carries at least one score. Almost none carry a complete one:

| Dimensions scored | Entities | |---|---:| | 1 | 327 | | 2 | 213 | | 3 | 136 | | 4 | 22 | | **5 (complete)** | **84 (10.74%)** |

**700 of 782 (89.51%) are scored only by the machine rubric `v1-auto`; 82 (10.49%) carry an adjudicated `v1` score.** The two barely overlap.

They must be reported separately, because `v1-auto` is close to a two-valued instrument: across 1,277 auto-scored rows on published entities it has emitted **zero 3s and seven 0s — 99.45% of its output is a 1 or a 2.** A four-point rubric that in practice returns two values cannot support a claim about how well the market is governed, and this is the fifth volume in which it has never emitted a 3.

The adjudicated panel is small but clean: 82 entities, all five dimensions, no gaps.

| Dimension (v1, n=82) | Score 0 | 1 | 2 | 3 | Trust gap (0–1) | |---|---:|---:|---:|---:|---:| | `insurance_indemnity` | 78 | 4 | 0 | 0 | **100.00%** | | `outcome_settlement` | 78 | 3 | 1 | 0 | **98.78%** | | `audit_trail` | 53 | 17 | 11 | 1 | **85.37%** | | `compliance` | 46 | 20 | 16 | 0 | 80.49% | | `runtime_governance` | 26 | 31 | 25 | 0 | **69.51%** |

Against the machine panel, where `audit_trail` shows a 94.01% gap but `compliance` only 31.75% and `runtime_governance` 51.32%.

We wanted to compare the two rubrics on the entities both had scored. **We cannot.** The overlap is 5 to 7 entities per dimension — too small to support any statement about whether the machine scorer is systematically generous or harsh, so this report makes none. That comparison needs the adjudicated panel to grow.

## 7. Where this undercuts our own thesis

**The "continuously re-verified" claim is the weakest part of our pitch, not the strongest.** We market re-verification as the thing that beats a static academic index. Measured, 92.05% of our published values cite month-old evidence and 18.13% of what can be checked no longer verifies. A static index of 30 entities that were each checked once, carefully, by seven annotators may well have a lower citation-failure rate than ours does. We have not measured MIT's, and until we do we should stop implying the comparison.

**"Agents are ungoverned" remains wrong, and we should stop saying it.** On the adjudicated panel `runtime_governance` has the *lowest* trust-gap rate of the five dimensions at 69.51%, and 25 of 82 entities clear a 2. On the machine panel the compliance gap is only 31.75%. Both instruments agree on the direction. The defensible claim is narrower and better: **agents are governed and uninsured.** `insurance_indemnity` is at a 100.00% trust gap on every entity a human has adjudicated.

**The 782 are not a sample of agents.** They are a sample of agents for which we found at least eight facts and one score, which is what the publication floor requires. Every rate in this report is conditioned on that. The market's true absence rate is almost certainly higher, since the entities we could not document are the ones that document least.

**The vendor side of our wedge is thinning.** 29 published vendors, against 42 before the publication floor. Vendors carry 6.97 present assurance fields each against agents' 3.65 — they are the most informative entities in the index, and there are fewer of them than a month ago.

## 8. Handoffs

**Platform Engineer, first.** `superseded_quotes` should compare against the newest `raw_documents` capture, not against other evidence rows already attached to the entity. As written it cannot see 89.73% of the index. The 1,219 failures in §3 are the backlog it should have been surfacing.

**Platform Engineer, second.** `content_hash` is unusable as a change detector — 1 repeat in 1,748 re-captured URLs. Anything relying on it to mean "the content changed" is relying on noise.

**Fact-Checker, first.** 960 published values whose citation no longer verifies against a capture that is not shorter, 280 of them assurance fields. Each needs either a re-pointed evidence row or demotion to `no_public_information`.

**Fact-Checker, second — third consecutive volume.** Merge `elevenagents` (109) and `elevenlabs-agents` (2879). One product, two published pages, one shared evidence row, sitting on the single most important field in the index.

**Closed from Vol. 5.** Lloyd's of London, Chaucer Group and Mosaic Insurance are no longer published as `no_public_information` on the risk-transfer fields — all three are now candidates, demoted by the publication floor. The reader-facing harm Vol. 5 flagged is resolved. The underlying gap is not: all three still carry zero present risk-transfer fields, and we still have no documented coverage of three of the best-known names in the market we claim to own.

## Appendix — reproduction

Every figure above is produced by one of these queries, run against the index on 2026-09-07.

**A1 — index counts (§1)**

```sql select (select count(*) from entities) as entities_total, (select count(*) from entities where status='published') as published, (select count(*) from entities where status='candidate') as candidate, (select count(*) from field_catalog) as fields; ```

**A2 — applicable slots and the absence rate (§1)**

```sql with pub as (select id, entity_type from entities where status='published'), appl 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 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.id is null) as never_assessed, round(100.0*count(*) filter (where ef.value_status='no_public_information')/count(*),2) as absence_rate from appl a left join entity_fields ef on ef.entity_id=a.entity_id and ef.field_key=a.field_key; ```

**A3 — assurance vs non-assurance (§1)**

```sql with pub as (select id, entity_type from entities where status='published'), appl as (select p.id as entity_id, f.field_key, f.category from pub p join field_catalog f on f.applies_to='both' or f.applies_to=p.entity_type) select case when a.category='assurance' then 'assurance' else 'non-assurance' end as band, count(*) as slots, count(*) filter (where ef.value_status='present') as present, round(100.0*count(*) filter (where ef.value_status='present')/count(*),2) as pct_present from appl a left join entity_fields ef on ef.entity_id=a.entity_id and ef.field_key=a.field_key group by 1; ```

**A4 — Vol. 5 cohort reconstruction; must return 776 (§2)**

```sql with t as (select timestamptz '2026-09-01 00:00:00+00' as cut), chg as (select * from entity_overwrite_log, t where changed_at >= t.cut and old_status is distinct from new_status), promoted_after as (select distinct entity_id from chg where new_status='published'), demoted_after as (select distinct entity_id from chg where old_status='published'), cohort as ( select e.id from entities e where (e.status='published' and e.id not in (select entity_id from promoted_after)) or e.id in (select entity_id from demoted_after) ) select (select count(*) from cohort) as reconstructed_vol5_cohort, (select count(*) from promoted_after) as promoted_since, (select count(*) from demoted_after) as demoted_since, (select count(*) from entities where status='published') as published_now; ```

**A5 — that cohort's absence rate today, giving the like-for-like term (§2)**

Same CTE as A4, then join the cohort to `field_catalog` on `applies_to` exactly as in A2 and compute the absence rate. Returns 776 entities, 45,494 slots, 30,488 absences, 67.02%.

**A6 — weekly field-row activity on the Vol. 5 cohort (§2)**

```sql -- cohort CTE as in A4 select count(*) filter (where ef.updated_at >= timestamptz '2026-09-01 00:00:00+00') as rows_touched_this_week, count(*) filter (where ef.updated_at >= timestamptz '2026-09-01 00:00:00+00' and ef.value_status='no_public_information') as touched_npi, count(*) filter (where ef.updated_at >= timestamptz '2026-09-01 00:00:00+00' and ef.value_status='present') as touched_present, count(*) filter (where ef.created_at >= timestamptz '2026-09-01 00:00:00+00') as rows_created_this_week from entity_fields ef join cohort c on c.id=ef.entity_id; ```

**A7 — the re-verification test (§3)**

```sql with pub as (select id from entities where status='published'), cited as ( select ef.quote, ev.source_url, ev.retrieved_at from entity_fields ef join pub p on p.id=ef.entity_id join evidence ev on ev.id=ef.evidence_id where ef.value_status='present' and ef.quote is not null and length(trim(ef.quote))>0 ), newest as ( select distinct on (url) url, content, fetched_at from raw_documents where http_status=200 order by url, fetched_at desc ) select count(*) as present_rows_with_quote, count(*) filter (where n.fetched_at > c.retrieved_at) as have_newer_capture, count(*) filter (where n.fetched_at > c.retrieved_at and position(regexp_replace(trim(c.quote),'\s+',' ','g') in regexp_replace(n.content,'\s+',' ','g')) > 0) as quote_still_present, count(*) filter (where n.fetched_at > c.retrieved_at and position(regexp_replace(trim(c.quote),'\s+',' ','g') in regexp_replace(n.content,'\s+',' ','g')) = 0) as quote_gone from cited c join newest n on n.url = c.source_url; ```

**A8 — truncation control for A7 (§3)**

Same joins as A7, additionally joining `raw_documents` on the cited `raw_document_id`, grouped by whether the quote survives, comparing `length(content)` of the cited and newest captures. Returns: gone rows, average cited length 9,848, average newest length 10,923, ratio 1.448, 259 with a newest capture below 80% of the cited length and 960 at or above it.

**A9 — supersession by field category (§3)**

As A7, joined to `field_catalog` and grouped by `category`.

**A10 — evidence age (§3)**

```sql with pub as (select id from entities where status='published') select count(*) as present_rows, percentile_cont(0.5) within group ( order by extract(epoch from (now()-ev.retrieved_at))/86400)::numeric(10,1) as median_age_days, count(*) filter (where ev.retrieved_at < now() - interval '30 days') as older_than_30d from entity_fields ef join pub p on p.id=ef.entity_id join evidence ev on ev.id=ef.evidence_id where ef.value_status='present'; ```

**A11 — the detector's blind spot (§4)**

```sql with cited as ( select ef.entity_id, ev.source_url, ev.content_hash from entity_fields ef join evidence ev on ev.id=ef.evidence_id join entities e on e.id=ef.entity_id where ef.value_status='present' and e.status='published' ) select count(*) as entity_url_pairs, count(*) filter (where nhash > 1) as visible_to_detector, count(*) filter (where nhash = 1) as invisible_to_detector, round(100.0*count(*) filter (where nhash=1)/count(*),2) as pct_invisible from (select entity_id, source_url, count(distinct content_hash) as nhash from cited group by entity_id, source_url) g; ```

**A12 — detector health and stock (§4)**

```sql select count(*) as flags, count(*) filter (where verdict is not null) as adjudicated from superseded_quote_flags;

select job_name, count(*) as runs, max(started_at) as last_run, count(*) filter (where ok is false) as failed from maintenance_runs group by job_name; ```

**A13 — content_hash instability (§4)**

```sql select (select count(*) from raw_documents) as total_docs, (select count(*) from (select url from raw_documents group by url having count(*)>1) y) as urls_with_multiple_captures, (select count(*) from (select url, content_hash from raw_documents group by url, content_hash having count(*)>1) x) as url_hash_pairs_repeating; ```

**A14 — risk-transfer field census (§5)**

```sql with pub as (select id, entity_type from entities where status='published') select f.field_key, count(p.id) as applicable_entities, count(*) filter (where ef.value_status='present') as present, round(100.0*count(*) filter (where ef.value_status='present')/count(p.id),2) as pct_present from pub p join field_catalog f on f.applies_to='both' or f.applies_to=p.entity_type left join entity_fields ef on ef.entity_id=p.id and ef.field_key=f.field_key where f.field_key in ('insurance_available','own_liability_cover','insurance_carriers', 'coverage_limits','indemnification','liability_cap') group by f.field_key order by present desc; ```

**A15 — entities with no insurance evidence at all (§5)**

```sql with pub as (select id from entities where status='published'), ins as ( select p.id, max(case when ef.value_status='present' then 1 else 0 end) as has_any from pub p left join entity_fields ef on ef.entity_id=p.id and ef.field_key in ('insurance_available','own_liability_cover', 'insurance_carriers','coverage_limits') group by p.id ) select count(*) as published, count(*)-sum(has_any) as with_no_insurance_evidence, round(100.0*(count(*)-sum(has_any))/count(*),2) as pct from ins; ```

**A16 — the twelve risk-transfer rows, read individually (§5)**

```sql select e.id, e.slug, e.entity_type, ef.field_key, ef.evidence_id, ef.value_text, ef.quote from entity_fields ef join entities e on e.id=ef.entity_id where e.status='published' and ef.value_status='present' and ef.field_key in ('insurance_available','own_liability_cover','insurance_carriers', 'coverage_limits','indemnification','liability_cap') order by ef.field_key, e.slug; ```

**A17 — matched-denominator control comparison (§5)**

```sql with pub as (select id from entities where status='published'), grp(field_key, band) as (values ('insurance_available','risk-transfer'),('own_liability_cover','risk-transfer'), ('insurance_carriers','risk-transfer'),('coverage_limits','risk-transfer'), ('indemnification','risk-transfer'),('liability_cap','risk-transfer'), ('pricing_model','control'),('api_access','control'),('integrations','control'), ('availability','control'),('deployment_options','control'),('support_model','control')) select g.band, count(*) as slots, count(*) filter (where ef.value_status='present') as present, round(100.0*count(*) filter (where ef.value_status='present')/count(*),2) as pct_present from pub p cross join grp g left join entity_fields ef on ef.entity_id=p.id and ef.field_key=g.field_key group by g.band; ```

**A18 — scorecard coverage and completeness (§6)**

```sql with pub as (select id from entities where status='published'), cov as (select p.id, bool_or(sc.rubric_version='v1') as has_v1, bool_or(sc.rubric_version='v1-auto') as has_auto from pub p left join scorecards sc on sc.entity_id=p.id group by p.id) select count(*) as published, count(*) filter (where has_auto and not has_v1) as machine_only, count(*) filter (where has_v1) as any_adjudicated from cov;

with pub as (select id from entities where status='published') select dims_scored, count(*) as entities from (select p.id, count(distinct sc.dimension) as dims_scored from pub p left join scorecards sc on sc.entity_id=p.id group by p.id) x group by dims_scored order by dims_scored; ```

**A19 — score distribution by rubric (§6)**

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

**A20 — agents vs vendors on assurance (§7)**

```sql with pub as (select id, entity_type from entities where status='published'), appl as (select p.id as eid, p.entity_type as etype, f.field_key from pub p join field_catalog f on (f.applies_to='both' or f.applies_to=p.entity_type) and f.category='assurance') select a.etype, count(distinct a.eid) as entities, count(*) filter (where ef.value_status='present') as present, round(count(*) filter (where ef.value_status='present')::numeric/count(distinct a.eid),2) as present_per_entity from appl a left join entity_fields ef on ef.entity_id=a.eid and ef.field_key=a.field_key group by a.etype; ```