# None of the Pages We Could Read Exposed a Machine-Readable AI Training Opt-Out

**FreePolicyTools original research.** Collected 19–21 November 2026 from a USA cloud path. Corpus frozen 19 November 2026 at 14:51:45Z. Protocol SHA-256 `edf01912678fadd972988c49e4f44afb7b67e6d64268820f0153f076f91d6b5d`. Corpus SHA-256 `763c75fd270695952b0c47ae7134b3e969c0b6e0413be2d62c42d52e035951fa`.

**Suggested citation:** FreePolicyTools. “None of the Pages We Could Read Exposed a Machine-Readable AI Training Opt-Out.” Original research collected 19–21 November 2026. Evidence: `/research/ai-opt-out-signals-3day/`. Article: `/blog/machine-readable-ai-opt-out/`.

## Result

On 24 public sites, three times in three days, this study found **zero observed** TDM Reservations, **zero** robots `noai` or `noimageai` rules, and **zero** root `ai.txt` documents.

“Zero observed” is the true count. It is narrower than “every site lacks the signal.” Three sites had at least one carrier we could not read, so those headlines stay **unknown**:

| Slot | When (UTC) | Late | TDM reservation true / false / unknown | robots `noai` or `noimageai` | root `ai.txt` |
| --- | --- | --- | --- | --- | --- |
| day0 | 2026-11-19 14:56:02–14:57:36 | no | 0 / 22 / 2 | 0 / 23 / 1 | 0 / 23 / 1 |
| day1 | 2026-11-20 06:01:27–06:03:11 | no | 0 / 21 / 3 | 0 / 22 / 2 | 0 / 23 / 1 |
| day2 | 2026-11-21 06:01:19–06:03:00 | no | 0 / 21 / 3 | 0 / 22 / 2 | 0 / 23 / 1 |

`tdm_protocol_present` matched the TDM reservation counts on every day. No site returned a valid `tdmrep.json`, and no observed carrier was `0` or `1`. A reservation of `0` would have counted as the protocol being present and would still have left `exposes_tdm_reservation` false. We never reached that case.

Unknown sites:

- **All three days, TDM:** NordVPN, ExpressVPN.
- **All three days, robots and `ai.txt`:** NordVPN.
- **20 and 21 November, TDM and robots:** Amazon, which had been false (observed, no signal) on 19 November.

The only scored signal change in the study is Amazon’s move from false to unknown. Day1 to day2 has an empty `signal_changes` list. File bytes moved on other hosts. Those moves are body drift or HTML size drift. They did not add a reservation, a `noai` token, or an `ai.txt` document.

## Question

Which sites in the frozen privacy-policy observatory corpus expose a [TDM Reservation](https://www.w3.org/community/reports/tdmrep/CG-FINAL-tdmrep-20240510/), a robots `noai` or `noimageai` rule, or a root [`ai.txt`](https://site.spawning.ai/spawning-ai-txt), and do those signals change across 19, 20, and 21 November 2026?

The observatory at [`/blog/privacy-policy-observatory/`](/blog/privacy-policy-observatory/) supplies the site list only. This study does not re-score the 12-dimension rubric, and it does not read the retention, sharing, or arbitration clauses tracked by the privacy-policy mutation week.

A companion on what a notice should say in prose is [`what a website privacy policy should include`](/blog/what-should-a-website-privacy-policy-include/). Placement of that notice is a separate question, covered in [`where to display legal policies`](/blog/where-should-you-display-your-legal-policies-on-a-website/). Operators who need a starting document can use the [`privacy policy generator`](/tools/privacy-policy/). Common prose failures are listed in [`privacy policy mistakes to avoid`](/blog/privacy-policy-mistakes-avoid/). None of those pages answers the machine-readable question in this study.

## What counted

The definitions were frozen in `protocol.md` before the first crawl.

**TDM Reservation.** The W3C TDM Reservation Protocol Community Group Report of 10 May 2024 uses `tdm-reservation: 1` for rights reserved and `0` for not reserved. Any other value is a protocol error and is stored, then treated as unset for the effective value. Carriers, in order, for one URL: a matching rule in `/.well-known/tdmrep.json`, then a `tdm-reservation` response header, then an HTML `<meta name="tdm-reservation">`, then PDF XMP `tdm:reservation`. A later observed `0` or `1` overrides an earlier one. A missing carrier does not erase a value already set. `tdm-policy` URLs are stored and are not fetched.

Path matching uses the URL path only. `*` matches any sequence, a trailing `$` anchors the end, and anything else is a prefix. The most specific pattern wins (count of non-star characters). Ties keep the earliest rule. Matching is case-sensitive. Amazon’s help URL keeps its query string on the request, and the node id in that query is outside the match. No `tdmrep.json` was valid in this corpus, so the matcher never selected a rule.

Headline `exposes_tdm_reservation` is true only when some observed carrier or file rule is `1`. It is false when every applicable carrier was observed and none is `1`. It is unknown when none is `1` and at least one applicable carrier was blocked, errored, or (for HTML meta) unread because the status was not a 2xx HTML document. Headers on a 403 are observed. HTML meta and PDF body on a 403 are not, because that body is a challenge page. A 200 HTML document at the well-known URL is `html_soft`. It is not the protocol file.

**Robots `noai` / `noimageai`.** A rule is a `robots.txt` directive of that name, a whole token in a field value, a `User-agent` of exactly that name, an HTML robots meta (`name=robots`, a name ending in `bot`, or `http-equiv=x-robots-tag`), or an `X-Robots-Tag` header. `noimageai` is a separate token. `Disallow: /noai/` is a path, and a comment is a comment. [RFC 9309](https://www.rfc-editor.org/rfc/rfc9309) is the robots exclusion protocol this parser sits on; `noai` itself is a community token, described in practice by publishers such as [DeviantArt’s opt-out note](https://www.deviantart.com/team/journal/UPDATE-All-Deviations-Are-Opted-Out-of-AI-Datasets-934500371). The headline is true if either token appears on an observed carrier. False means every applicable carrier was read and neither token appeared. Unknown means a carrier was blocked or errored and no token was seen elsewhere. A 200 HTML body at `/robots.txt` is an observed response with no robots rule in it.

**Root `ai.txt`.** The headline location is `https://{site origin}/ai.txt`, the Spawning root file. It counts only on HTTP 200, a non-HTML body, and at least one `User-agent`, `Allow`, or `Disallow` line. 404 and 410 are absent (false). 200 HTML is `html_soft` (false), including a soft-404 shell. 200 non-HTML without those directives is `unstructured` (false), and the hash is kept. 401, 403, 429, 503, timeout, TLS failure, or DNS failure is unknown. `/.well-known/ai.txt` uses the same classifier and is secondary. It is not part of the headline.

The European text-and-data-mining exception that the W3C report is written against is Directive (EU) 2019/790, Article 4, on [EUR-Lex](https://eur-lex.europa.eu/eli/dir/2019/790/oj). This study records whether the machine-readable reservation was present on these URLs. It does not decide whether any operator, model, or jurisdiction is inside that exception.

## Method

**Corpus.** N=24, copied from the observatory sample frozen 17 November 2026 (sample SHA-256 `a268213114958277b2dc5480abec29cda50fff367039d9651020be7eee8b95d2`). Strata: VPN 7, proxy 5, privacy tools 4, mainstream platforms 4, legal-tech vendors 4. Ids, names, strata, and observatory URLs were not edited.

**Origin rule.** `site_origin` is `https://` plus the host of the observatory URL. `www` is not folded into the apex. Google is `policies.google.com`. Meta is `www.facebook.com`. Microsoft is `privacy.microsoft.com`.

**One pre-registered exception.** NordVPN’s observatory document is a PDF on `sb.nordcdn.com`. The site origin is `https://nordvpn.com`. The CDN host’s `robots.txt`, `ai.txt`, and `tdmrep.json` were not fetched and cannot satisfy the site-level signals. The PDF was fetched for resource-level TDM headers and XMP.

**Fetches.** Each site, each slot: `GET /`, `GET /robots.txt`, `GET /ai.txt`, `GET /.well-known/tdmrep.json`, `GET /.well-known/ai.txt`, and the document URL with the fragment removed. All 24 documents had a path, so every site was document-applicable. That is 6 URLs × 24 sites × 3 slots = **432 GET requests**. Every request returned an HTTP status. Status mix: 211 × 200, 195 × 404, 20 × 403, 6 × 410. No transport error.

**Client.** Python 3.12 `urllib`, TLS verification on, an empty proxy handler so `HTTP_PROXY` and `HTTPS_PROXY` cannot move the path, user-agent `FreePolicyToolsResearch/1.0 (+https://freepolicytools.com/research; ai-opt-out-signals-3day)`, timeout 20 seconds, at most 5 redirects, `Accept-Encoding: identity` with gzip recovery when a server compressed anyway. About 0.2 seconds between URLs on one host and 1.0–1.6 seconds between hosts. No cookies, no form posts, no login, no script execution. A 401, 403, 429, or 503 was recorded as blocked. It was not retried with a browser user-agent.

**Caps.** 512 KiB for `robots.txt`, `ai.txt`, and `tdmrep.json`. 1.5 MiB for HTML. 2 MiB for PDF. Truncation is recorded. Raw signal bodies stay in untracked `private/`. Public JSON stores status, final URL, hashes, extracted values, and short redacted excerpts. Homepage and PDF HTML are not saved.

**Path.** Cloud-direct from a Cursor-hosted USA datacenter. ProxySeller HTTP, ProxySeller SOCKS, and CometVPN are `NOT_RUN` on every slot. CometVPN is VPN-only and is also a corpus member, fetched as a public website. Environment names that matched the proxy/VPN pattern were recorded as names only.

**Schedule.** day0 is the first measurement, on 19 November, as soon as the protocol and corpus were frozen. day1 is 20 November 06:00–06:30 UTC. day2 is 21 November 06:00–06:30 UTC. All three slots started inside the window (`late: false`). Collection stopped after 21 November 2026, 06:30:00 UTC. There is no 22 November slot.

**Change.** A signal change is any difference in the scored tuple: headline tri-states, effective TDM values, the `tdmrep.json` rule list, robots rule excerpts, the `ai.txt` status class, and the `ai.txt` SHA-256 when that class is `document`. A body drift is a SHA-256 change of `robots.txt`, `tdmrep.json`, or an `ai.txt` body that leaves the scored tuple alone. `Content-Signal`, `Content-Usage`, `/.well-known/ai.txt`, and named full-site crawler disallows are diffed separately and do not flip the headline.

## Every site on 21 November

“Not observed” means false: the applicable carriers were read and the signal was absent. “Unknown” means at least one applicable carrier was blocked or unread, and no positive signal was seen on the carriers that were read. Amazon’s row is the 21 November state. On 19 November Amazon was “not observed” for TDM and robots as well. Every other row was stable across the three days.

| Site | Stratum | TDM reservation | robots `noai` | root `ai.txt` |
| --- | --- | --- | --- | --- |
| Proton VPN | VPN | not observed | not observed | absent |
| Mullvad VPN | VPN | not observed | not observed | absent |
| NordVPN | VPN | unknown | unknown | blocked |
| Surfshark | VPN | not observed | not observed | HTML page |
| ExpressVPN | VPN | unknown | not observed | absent |
| IVPN | VPN | not observed | not observed | absent |
| CometVPN | VPN | not observed | not observed | absent |
| Proxy-Seller | Proxy | not observed | not observed | absent |
| Bright Data | Proxy | not observed | not observed | absent |
| Oxylabs | Proxy | not observed | not observed | absent |
| IPRoyal | Proxy | not observed | not observed | absent |
| Webshare | Proxy | not observed | not observed | absent |
| Bitwarden | Privacy tool | not observed | not observed | absent |
| 1Password | Privacy tool | not observed | not observed | absent |
| DuckDuckGo | Privacy tool | not observed | not observed | HTML page |
| Signal | Privacy tool | not observed | not observed | absent |
| Google | Mainstream | not observed | not observed | absent |
| Meta | Mainstream | not observed | not observed | HTML page |
| Amazon | Mainstream | unknown | unknown | absent |
| Microsoft | Mainstream | not observed | not observed | absent |
| Termly | Legal tech | not observed | not observed | absent |
| iubenda | Legal tech | not observed | not observed | absent |
| CookieYes | Legal tech | not observed | not observed | absent |
| Didomi | Legal tech | not observed | not observed | absent |

No stratum had a true. On 21 November the unknown cells are VPN (NordVPN for all three headlines, ExpressVPN for TDM) and mainstream (Amazon for TDM and robots). Proxy, privacy-tool, and legal-tech rows are false on every headline.

`/.well-known/ai.txt` was never a document. On 21 November, 22 sites were absent and 2 were blocked (NordVPN and ExpressVPN). That secondary file is not in the headline, including for ExpressVPN, whose root `ai.txt` was a 404.

## The unknown cases, carrier by carrier

### NordVPN — unknown on all three headlines, all three days

Identical on 19, 20, and 21 November:

| URL | Status | What was scored |
| --- | --- | --- |
| `https://nordvpn.com/` | 403, 5,626 bytes | Homepage meta unread. Headers on the 403 had no reservation and no `noai` token. |
| `https://nordvpn.com/robots.txt` | 200, `text/plain`, 15,951 bytes | No `noai` or `noimageai` rule. The hash did not change. |
| `https://nordvpn.com/ai.txt` | 403 | Headline `ai.txt` unknown (`blocked`). |
| `https://nordvpn.com/.well-known/ai.txt` | 403 | Secondary, blocked. |
| `https://nordvpn.com/.well-known/tdmrep.json` | 403 | File state `blocked`. |
| Privacy PDF on `sb.nordcdn.com` | 200, `application/pdf`, 107,311 bytes | XMP `tdm-reservation` unset. Document meta not applicable because the body is a PDF. |

The robots file we fetched has no `noai` token. The site-level robots headline stays unknown because the homepage was a 403 and HTML meta was not taken from a challenge page. The PDF was read and carried no reservation. The site-level TDM headline stays unknown because the homepage and `tdmrep.json` were blocked.

### ExpressVPN — TDM unknown, robots and `ai.txt` false, all three days

Homepage and the privacy-policy HTML were 200 on every crawl, and both had meta unset (no reservation, no `noai`). `robots.txt` was 200, `text/plain`, 1,136 bytes, with no `noai` rule, and the hash did not change. Root `ai.txt` was 404, so the headline is false (`absent`). `/.well-known/ai.txt` and `/.well-known/tdmrep.json` were 403, `application/xml`, 111 bytes, on every crawl. The blocked well-known TDM file is why TDM is unknown even though the HTML we did read had no reservation. The blocked well-known `ai.txt` does not move the headline, because that URL is secondary and the root file was an observed 404.

Homepage HTML moved from 265,222 bytes to 265,237 bytes on 21 November. The policy HTML moved from 157,572 bytes on 19 November to 158,754 bytes on 20 November and stayed there. Neither move changed a scored field.

### Amazon — false on 19 November, unknown for TDM and robots on 20 and 21 November

| Carrier | 19 November | 20 and 21 November |
| --- | --- | --- |
| Homepage | 200, 882,108 bytes, meta unset | 200, 3,781 bytes, meta unset |
| `robots.txt` | 200, `text/plain`, 7,947 bytes, no `noai` rule | Same status, size, and hash |
| `/ai.txt` | 404 | 404, headline stays false |
| `tdmrep.json` | 404, file absent | 404, file absent |
| Policy document | 200, `text/html`, 379,320 bytes, meta unset | 403, `text/html`, 2,319 bytes, meta unread |

The scored tuple changed on six fields: `effective_document`, `exposes_noai`, `exposes_noimageai`, `exposes_robots_noai`, `exposes_tdm_reservation`, and `tdm_protocol_present`, each from an observed negative to unknown. `exposes_ai_txt` stayed false. The 3,781-byte homepage is still HTTP 200 HTML, so its meta was scored as unset. The unknown comes from the policy URL, whose HTML meta could not be read on a 403. The robots file itself still had no `noai` token. This is a loss of observation, not a new opt-out.

The same `robots.txt` still contained full-site `Disallow` rules for 27 named crawlers. That list did not change. It is unscored, and it is not a `noai` rule. See below.

## Bytes that moved without a new signal

**`ai.txt` hashes, status class unchanged (`html_soft`).** Surfshark, DuckDuckGo, and Meta returned HTTP 200 `text/html` at `/ai.txt` every day. The bytes changed between every pair of days. A 200 HTML body is not a Spawning `ai.txt`. DuckDuckGo’s page is an HTML search response. It contains a script parameter named `safe_noai`. That parameter is not a robots directive, not a robots meta token, and not an `ai.txt` rule, so the classifier left it out. The same host returned HTTP 410 HTML (144,097 bytes) for `/.well-known/ai.txt` and `/.well-known/tdmrep.json` every day. 410 is known absence.

**`robots.txt` hashes, scored rules unchanged.** 21 of 24 robots files had the same SHA-256 on all three days. Signal (11,109 bytes) and Google (1,652 bytes) changed hash every day at a constant length. Bright Data stayed 609 bytes on 19 and 20 November and grew to 945 bytes on 21 November. For all three, the extracted `noai` list stayed empty, and the separate diff of `Content-Signal` and named crawler disallows was empty (`secondary_or_unscored_ids` is `[]` on both diffs). The public package records the hash and the extracted fields. The raw robots bodies stay private.

**HTML sizes.** Homepage or document byte lengths moved, with status still 200, for Surfshark, ExpressVPN, CometVPN, Proxy-Seller, Bright Data, Oxylabs, Bitwarden, 1Password, Google, Meta, Termly, iubenda, and CookieYes. HTML page size is not a scored field. Amazon is the only host whose document status changed, and that status change is the signal change above.

No `tdmrep.json` hash entered a drift record, because no site returned a 200 protocol file to hash.

## What the files contained instead

These fields were recorded and do not answer the question. They were stable: the unscored diff is empty from day0 to day1 and from day1 to day2.

**`Content-Signal`.** Five sites sent the same triple: `ai-input=yes`, `ai-train=yes`, `search=yes`.

| Site | Copies stored | Named full-site disallow in the frozen crawler list |
| --- | --- | --- |
| IVPN | 1 | Bytespider |
| Bright Data | 1 | none |
| Oxylabs | 1 | none |
| IPRoyal | 1 | none |
| CookieYes | 13 copies of that same triple | PetalBot |

`ai-train=yes` is a machine-readable permission. It is the other direction from a training opt-out. It was not scored as a TDM Reservation, a `noai` rule, or `ai.txt`. `Content-Usage` was empty on every site.

CookieYes’s 13 copies are repeated lines across the robots file and response headers, not 13 different policies. CookieYes’s 404 bodies for `/ai.txt`, `/.well-known/ai.txt`, and `tdmrep.json` were truncated at 524,288 bytes. The status was already 404, so the file is absent. The missing tail of an error page cannot create a reservation.

**Named full-site disallows.** The frozen list is `unscored-crawlers.json` (32 agent names, not extended during the study). A hit is a group whose user-agent is on that list, with `Disallow: /` or `Disallow: /*`, and without `Allow: /` or `Allow: /*`.

- **Meta** (`www.facebook.com`): Amazonbot, Applebot-Extended, ClaudeBot, Google-Extended, GPTBot, PerplexityBot, PetalBot (7).
- **Amazon**: GPTBot, CCBot, PerplexityBot, Google-Extended, ClaudeBot, meta-externalagent, Bytespider, PetalBot, omgili, AI2Bot, Diffbot, DuckAssistBot, Ai2Bot-Dolma, Google-CloudVertexBot, meta-externalfetcher, cohere-ai, YouBot, Claude-User, Claude-SearchBot, Perplexity-User, ChatGPT-User, OAI-SearchBot, Timpibot, Claude-Web, GoogleOther, ICC-Crawler, ImagesiftBot (27).
- **IVPN:** Bytespider.
- **CookieYes:** PetalBot.

Meta and Amazon published long bot-specific `Disallow` lists and still had no `noai` token, no `noimageai` token, no TDM Reservation, and no root `ai.txt` on the responses we could read. Amazon’s robots headline is unknown on 20 and 21 November because the policy HTML was blocked, not because `robots.txt` gained a `noai` line. The robots file stayed 7,947 bytes and the 27-name list stayed put.

## Microsoft’s redirect

Every Microsoft request started at `privacy.microsoft.com` and finished on `www.microsoft.com` (`cross_host: true`), all three days:

| Request | Final URL | Status |
| --- | --- | --- |
| Homepage | `https://www.microsoft.com/en-us/privacy` | 200, 173,753 bytes |
| Document | `https://www.microsoft.com/en-us/privacy/privacystatement` | 200, 349,076 bytes |
| `robots.txt` | `https://www.microsoft.com/en-us/privacy/robots.txt` | 404, 76,859-byte HTML |
| `ai.txt` | `https://www.microsoft.com/en-us/privacy/ai.txt` | 404, same HTML length |
| `/.well-known/ai.txt` | `https://www.microsoft.com/en-us/privacy/.well-known/ai.txt` | 404 |
| `tdmrep.json` | `https://www.microsoft.com/en-us/privacy/.well-known/tdmrep.json` | 404 |

The final HTML pages had no reservation and no `noai` token, and the three signal URLs were 404, so all three headlines are false. The collector followed the redirect of `privacy.microsoft.com/robots.txt`. It did not also request `https://www.microsoft.com/robots.txt`. A robots file at the destination host’s root would be a different URL.

## Limits

- **Research user-agent.** Blocks and thin 200 responses can be aimed at this client. NordVPN’s 403s and Amazon’s collapse from an 882,108-byte homepage to a 3,781-byte 200 are consistent with a bot response. The protocol forbids a second fetch with a browser user-agent, so those gaps stay unknown.
- **One path.** USA cloud-direct only. ProxySeller HTTP, ProxySeller SOCKS, and CometVPN were not used as exits. A residential path or a browser could see a different status.
- **Purposive N=24.** The set is the observatory list. It is not a sample of the web, and a count here is not a prevalence estimate.
- **Hosts as named.** `www` and the apex are different origins. Google, Meta, and Microsoft are scored on the policy host the observatory named, not on every property those companies run.
- **Path-only TDM matching.** Query strings, including Amazon’s help-page node id, are not part of `tdmrep.json` location matching.
- **Root `ai.txt` only.** A file at `/.well-known/ai.txt` cannot flip the headline. None was a document anyway.
- **No browser rendering.** Meta tags inserted by JavaScript were not seen. The fetch is the bytes on the wire.
- **A missing file is an observation of this fetch.** It does not record permission to train, and it does not record a reservation. `Content-Signal: ai-train=yes` is a separate, unscored permission on five sites.
- **Caps.** CookieYes 404 bodies were cut at 512 KiB. Status 404 already classifies those URLs as absent. No scored 200 signal file was truncated.
- **NordVPN CDN.** Signals on `sb.nordcdn.com` other than the privacy PDF’s own headers and XMP were out of scope by the origin rule.
- **Short window.** Three fetches in three days can show a change that happened inside the window, as Amazon’s document status did. They cannot show what the same URLs did in October or will do in December.

## Evidence

Public package: [`/research/ai-opt-out-signals-3day/`](/research/ai-opt-out-signals-3day/).

| File | What it is |
| --- | --- |
| `protocol.md` | Frozen definitions |
| `corpus.json` | 24 origins and document URLs |
| `schedule.json` | Slot windows and the stop time |
| `unscored-crawlers.json` | Frozen agent list |
| `slots/day0.json`, `day1.json`, `day2.json` | Per-site statuses, hashes, scored tuple, unscored context |
| `diffs/day0-day1.json`, `day1-day2.json` | Signal changes and body drifts |
| `summary.json` | Counts and the same diffs |
| `collection-status.json` | Slots complete, article published |
| `manifest.sha256` | Checksums of the public package |
| `research-article.md` | This article |

Screenshots are omitted on purpose. The evidence is status codes, headers, and file bytes. A picture of a homepage would not show `robots.txt`.

## Disclosure

This article links to [ProxySeller](https://freepolicytools.com/proxyseller/) and [CometVPN](https://freepolicytools.com/cometvpn/). FreePolicyTools may earn a commission if you buy through those links. The links are affiliate placements. They are not measurements.

ProxySeller HTTP and SOCKS were `NOT_RUN`. CometVPN was `NOT_RUN` as a tunnel. Both companies appear in the corpus as public websites (Proxy-Seller in the proxy stratum, CometVPN in the VPN stratum) and were fetched with the same cloud-direct client as the other 22 sites. Affiliates did not design the protocol, choose the corpus, or see the results before publication. The corpus is the observatory list.

## Reproduce

From the repository root, after 21 November 2026, the collector refuses another scheduled slot. The commands that produced the three files were:

```bash
python3 research/ai-opt-out-signals-3day/scripts/collect.py --self-test
python3 research/ai-opt-out-signals-3day/scripts/collect.py --slot day0
python3 research/ai-opt-out-signals-3day/scripts/collect.py --slot day1
python3 research/ai-opt-out-signals-3day/scripts/collect.py --slot day2
```

The collector refuses a slot before its window unless `--allow-early` is passed, and it refuses to overwrite a finished slot unless `--force` is passed. Neither flag was used. Re-running a slot would be a new measurement, not a replay of these hashes.
