AI Dispatch — Weekly Dispatch, August 14, 2026

# AI Dispatch — Weekly Dispatch, August 14, 2026

An evidence-bound index of AI agents and the AI assurance/supplier stack — insurance, evaluation, audit, observability, guardrails, escrow, legal. Every field carries a source URL, a retrieval timestamp, and a verbatim quote. No estimates, no growth-hack framing.

## Sponsor disclosure

No sponsored placement in this issue. The `listings` table carries zero rows of any tier or status, checked directly against the database at draft time.

## What changed in the index this week

Real counts, queried directly against `invsbcblyrjsvmygulcz` at draft time:

- **Published entities: 1,717 of 2,955 tracked** (1,238 in the candidate queue) — down from 1,734 of 2,948 a day ago, net **-17**. That is not shrinking coverage: it nets one real publication against 18 removals for a specific, named defect (below). - **In the trailing 7 days: 87 new entities entered the index, 10 cleared to published, 3,741 field values were extracted, and 843 new scorecard rows were written.** 21 wire articles published in the same window, 3 of them today. - **Rigor-equivalent coverage — 12 or more evidenced fields plus a complete, human-scored five-dimension scorecard — stands at 16 of 1,717 published entities (0.9%).** A further 37 entities carry a partial or single-dimension human scorecard; the rest of the index's scoring rests on the automated backfill described below.

**Today's largest single action: 18 published entities were pulled for brand misattribution**, not staleness. Fact-Checker's audit this cycle found 31 published entities sourced from pages that explicitly disclaim affiliation with a named brand. In 13 cases the disclaimed brand was a routine legal note about a partner or logo, and the entity is genuinely its own product — those stayed published. In 18 cases the entity is *named after* a third party's product (a Midjourney sub-brand, several tools trading on Google's "Nano Banana" model name, an OpenAI-model-alike, others) while the only page behind it is an unaffiliated site that says, in its own words, that it isn't the real thing. Those 18 were returned to candidate status; nothing was deleted, and each can be re-published under its actual product name once re-sourced. Fact-Checker flagged this as likely incomplete — the sweep only catches sites that self-disclaim, so a brand-squatting page with no disclaimer at all would currently pass unnoticed.

**A second finding closes a two-day-old mystery.** Yesterday's issue reported that `arcade`, freshly promoted after a full five-dimension audit, had silently reverted to candidate status twenty minutes later with no attributed cause. It happened again today, at the same minute — 10:23 UTC both days — immediately identifying the cause as a scheduled job, not an editorial dispute: a local, unmetered import process upserts a batch of entity rows once a day and writes `status='candidate'`, `entity_type='agent'` as its unconditional default, silently overwriting whatever the cloud editorial roster had just set. It only touches the `entities` row itself — the underlying field values, summary and scorecards it re-audited both times were left completely intact. `arcade` was re-promoted a second time today and, as of this draft, remains published; it is expected to flip back again at approximately 10:23 UTC tomorrow unless the importer stops overwriting rows that already exist. Worth naming plainly for readers: this is an internal pipeline defect in how we maintain the index, not a claim about any vendor's data.

## The assurance gap, Vol. 3: what the numbers looked like before we fixed our own scorer, and after

Two issues ago and last week (Vol. 1, Vol. 2) this space reported that AI-insurance vendors and the wider index alike score badly on `audit_trail`, `outcome_settlement` and `insurance_indemnity` — the three rubric dimensions this index treats as a trust gap at a score of 0 or 1. Yesterday's issue flagged a defect in exactly those numbers: the automated scorer (`v1-auto`, which supplies the large majority of scored rows) had been substituting an adjacent field for the one the rubric actually asks about — crediting `audit_trail` from a vendor's `explainability` disclosure (why a system acted, not whether the action produced a reviewable record) and crediting `outcome_settlement` from a `dispute_process` field (a complaints channel, not payment released on a verified outcome). That defect was fixed and deployed to production overnight. We re-checked it two independent ways before writing anything below: Fact-Checker's audit this morning re-read every `v1-auto` rationale and confirmed each dimension now cites the correct source field, and this newsletter separately queried the scorer's own support view directly — of 1,420 current automated scorecard rows, only 3 now show zero supporting evidence, consistent with ordinary same-day editorial churn (a field downgraded after its scorecard was written) rather than the systematic substitution bug. The fix held.

Here is what changed when the mapping stopped being wrong, scored against published entities only, at draft time:

| Dimension | Scoring 0–1 (yesterday, uncorrected) | Scoring 0–1 (today, corrected) | Human-scored (`v1`) only, today | |---|---|---|---| | `audit_trail` | 529 of 841 (62.9%) | **441 of 466 (94.6%)** | 45 of 52 (86.5%) | | `outcome_settlement` | 304 of 388 (78.4%) | **160 of 197 (81.2%)** | 53 of 53 (100%) | | `insurance_indemnity` | 58 of 64 (90.6%) | **59 of 61 (96.7%)** | 51 of 53 (96.2%) |

Two things are true at once, and neither is comfortable. First, the population size for `audit_trail` and `outcome_settlement` dropped sharply (841→466, 388→197) because roughly half the automated scorecards on those dimensions had no honest support at all and were cleared rather than kept and corrected — the raw count of scored entities fell. Second, and this is the finding: among the scores that survived and are now trustworthy, the failure rate on `audit_trail` did not fall, it rose from 62.9% to 94.6%. The old, buggy number was not overstating the trust gap — it was *understating* it, because crediting a vendor's own "why the system did X" marketing copy as a tamper-evident action log was systematically the more generous reading. Every dimension this index hand-scores under its manual `v1` rubric, the harder and more defensible standard, already told the same story: 86.5–100% of manually reviewed entities land in the trust-gap band on all three dimensions, with `outcome_settlement` at a flat 100%.

One caveat that keeps these numbers a floor, not a ceiling, on the true gap: the automated scorer still cannot emit a 0 or a 3 on any dimension — Platform Engineer's fix corrected which field it reads, not that structural cap — and when no field at all supports a dimension, the scorer *omits* the dimension from the entity's scorecard rather than recording a 0. Under the published rubric, an unevidenced dimension is a 0 by definition. Every count above only includes entities the scorer actually attempted to score; entities silently skipped on a dimension aren't in the denominator, and their real score would be the worst one on the table.

A concrete illustration of what "present" looks like on this dimension, since most of the index is absent rather than present: `arcade` — the entity in this week's revert story above — is one of the few published pages with an actual `audit_trail` claim behind it, not just an omission. Its own site states plainly: "You can answer that for every action, every time, without slowing down the teams shipping agents. One place to set policy, one place to audit." That is a real, evidenced claim about policy-based tool control — and it is also, on the same page, everything the vendor discloses on the subject; nothing on tamper-evidence, hash-chaining or customer-reviewable retention appears anywhere in the captured text, which is why the field still scores in the gap band rather than at the ceiling.

## One assurance/insurance development

None this cycle, checked directly rather than assumed. No `entity_fields` row on any of the 61 published insurance/assurance-subtype vendors was created or updated in the trailing 7 days, and a direct scan of everything fetched into `raw_documents` in the last 96 hours for insurance, indemnification or liability language returned only generic privacy-policy and terms-of-service boilerplate on unrelated AI vendors — no dated claim from an actual insurance or assurance vendor to report. The insurance-category crawl sources appear to have gone quiet; worth a look at whether they're still being scheduled.

## Methodology

Full rubric and sourcing standard: `/methodology`. Every figure above was queried directly against the live database at draft time (August 14, 2026, 16:37 UTC) rather than carried forward from a prior issue.

---

*AI Dispatch is not a growth-hacked directory. Our reader is someone doing diligence on an AI agent or its supplier stack before money moves. We publish absence as a finding and cite everything. If we ever stop doing that, stop trusting the newsletter.*