Public data index › AI promise-surface archive

Which AI policy and terms pages changed between dated captures?

A hash-led record through 2026-09-05 across 186 public pages from 74 companies. Stored page bodies and line-level diffs are withheld.

560 normalized content-hash transitions were observed across 100 surfaces during 57 sealed capture days. A privacy counsel at an AI product company can request the $499/month reviewed change record below, including dates, source links, seals, coverage gaps, and human dispositions.

Mechanical observation, not legal interpretation. A hash transition is a review candidate, not a claim that contractual language changed. Navigation, timestamps, experiments, localization, or template changes can move a normalized hash without changing a policy's substance. Every candidate requires review against the public source.

10,602dated surface observations attempted
8,789normalized text captures stored and hash-verified
1,198blocked outcomes recorded, never treated as absence
57/58calendar dates carrying a complete local seal

Recent transition candidates with positive provenance

Each row names the catalog company and public source page, the two capture dates, and the SHA-256 values of our normalized text before and after. The table does not reproduce the text or claim that the source's contractual meaning moved.

catalog entrycapture windowpublic provenancenormalized text hashes
Runway
terms
2026-09-04 → 2026-09-05source page31808cb2d07e82d13c858b54434c118acc948bb2df3250d4c406a55883ea41d7

31ae7cd5b9f3f31d0dc71240fccbaa9c616d39c96551659dee1fbcde9102a310
Runway
privacy
2026-09-04 → 2026-09-05source paged2c77f7aba834d3a4a556c081d24c8d53a496a1252dc6a0b5883f05fae614bce

a83d22afa7f1457856461a4682c7c80cdb593e544943f9237f7816a4ba3274b0
Openai
deprecation
2026-09-04 → 2026-09-05source page6aca39bf3820b6aac7e732e1be3bcf9b605ad1b44532705239864bc210054f7d

2d02262780208a31f6182848263b6d90bf0dcd8fbbf102087be2705b86b4384e
Microsoft
terms
2026-09-04 → 2026-09-05source page5c4c2bb859ed6812249af3f4c4d0bbb0664b0c18adc092169e2aa1934e9f74ea

3c0b0143c71f7a2e7d90d731a61a82b74e89282f4ff352e54fa99adfc7df3329
Microsoft
privacy
2026-09-04 → 2026-09-05source pagea096901901d7d00327de9e1f420d6898cc88264dfc2bd564403a4c2206e47672

0c6c41e75b4880142863a6d92cbadbb369137f1da21d11fbf369f02318e94934
Microsoft
acceptable use
2026-09-04 → 2026-09-05source page045e2bb9f4e9a7fd156679c6620a6f8b2c8fb0c6d532ecca725a99e597bc0481

84f2e348e535bc5f24b116c822052d4655cdb8b71932d08387954b6bb30d0913
Harvey
terms
2026-09-04 → 2026-09-05source page2d261608ab48af6920c4f67477e0e8e923bd9cb95616c562ddc626ff814ebc4e

6ba58e149a2804db504165413d46095fedf58ffbaf690fbeba92a383b9a99706
Harvey
privacy
2026-09-04 → 2026-09-05source paged385d0cd0aad80d83fddb1de01a2f5f812f709be951aeb7bb1d10f272604a0cd

c0ad68b5f7aae1599ebb7859b39bf783fdabc2b7210a280136af06d6beadee9b
Google Deepmind
data use
2026-09-04 → 2026-09-05source pagea9b307ab5c414f6106fd43f2260895883636327cfa2537e654f85036d4138831

b2fbb7cd988ff0ced76a8645129df34928aaebd8237e485e2dbf1c53d9ee3eef
Cohere
acceptable use
2026-09-04 → 2026-09-05source page14b041f08bc8ef402a761f831f12112cfd9f541629e8045f71c37c8a9231eabf

4ff52db7dcb4fe371db5ea6389b319440dc432435ac3a9ad0bc6b0fc42479bec
Vercel
privacy
2026-09-03 → 2026-09-04source page0d5331de96f167552210798099c84adc25e7b8d4b3d0557ab78bb6219e8e0068

7d4750247bc9bb5e1d592bc6809bac46350790d7e9631cdd3e016c6f4f7341a0
Openai
deprecation
2026-09-03 → 2026-09-04source pageddfa291b8798bc445d592dfb471a378fe17893c5c0b7c290c1555282c3b8d8f9

6aca39bf3820b6aac7e732e1be3bcf9b605ad1b44532705239864bc210054f7d

Monthly AI Promise-Surface Change Record

$499/month

For one privacy counsel at an AI product company: a private dated PDF and CSV covering the frozen 74-company, 186-surface panel, delivered within two business days after month end. Each reviewed candidate includes the source URL, capture window, before/after SHA-256 values, surface type, coverage status, and a concise human disposition separating likely page-template churn from a change worth legal or compliance review.

  • One monthly PDF and CSV over the existing frozen panel.
  • Blocked, dead, and missing captures travel with every issue.
  • No stored page body or unified diff is redistributed.
  • Cancel before the next monthly compilation.

Request a scoped sample →

This is a demand test. No checkout is active; the request form opens a scoped request.

Coverage that constrains every count

The archive spans 2026-07-10 through 2026-09-05: 57 complete seals across 58 calendar days. Missing sealed dates inside that range: 2026-08-31. Newer directories without both a complete manifest and seal are excluded rather than guessed; pending unsealed dates: none.

On the latest sealed date, 155 of 186 catalog pages produced stored normalized text, 21 were blocked by robots or the source, and 10 were dead, empty, non-text, or failed. Across the full window, 615 outcomes were dead or failed. A failed or blocked fetch is never counted as an unchanged page.

Transition mix by page type

surface typecontent-hash transitions
privacy186
terms170
acceptable use79
data use72
deprecation50
enterprise3

The same adjacent-date comparison also recorded 38 newly captured outcomes, 23 capture disappearances, 1 newly blocked outcomes, and 0 unblocked outcomes. Those are observation-state events, not claims that a company launched, withdrew, or restored a policy.

How the seal and comparison work

  1. Fetch the frozen public catalog with an identifying user agent, honor robots policy, and pace requests by host.
  2. Strip executable markup and normalize whitespace; store compressed text privately and hash the uncompressed normalized bytes.
  3. Write one manifest containing every attempted outcome, including blocked and failed rows.
  4. Compute a daily Merkle root over the stored-content hashes in deterministic catalog order.
  5. Compare adjacent complete manifests. Different hashes become review candidates; no body or unified diff crosses the public reader.

Seal limitation: the daily root is a local tamper-evidence commitment. It is not, by itself, an independent timestamp, notarization, legal opinion, security audit, or guarantee that the source text was complete. The public source URL and capture date remain the provenance for each candidate.

Latest Merkle root: eea70487ad041d7fc08cecec87dc850ec5554c506fb8a08a1073b7b4f4cbba83
Manifest SHA-256: be839de07b2b8cb24e7676947ef0bfeab0db26d3762fec61c445163e1f8ef462
Leaves: 155
Sealed at: 2026-09-05T14:48:14.601148+00:00

What is deliberately withheld

This page does not publish stored policy text, copied passages, unified diffs, page titles, fetch-error detail, block-reason text, cookies, response headers, or contacts. It does not assert motive, legal effect, effective date, or materiality. The public record is the smallest reproducible layer: source, dates, hashes, coverage, and method.

Browse other data services · Earlier sealed-evidence experiments · US Tech Automations