# Head-to-head: mathtec/site-render (ours) vs vdrmota/contact-info-scraper (Store leader), 2026-11-09

Bead: hustle-qtk. Purpose: find out whether the new listing ("contact details scraper for company websites that block plain requests") can honestly claim anything against the category leader, and if so what.

## Outcome first

- **On this test, the numbers do not support "better than the leader" overall.** The leader covered more sites, ran about 20x faster and cost about 5x less per site. Ours found more of the contacts on the pages it did load, and it alone decoded Cloudflare-protected emails.
- **Coverage: leader ahead.** Both Actors got the same 60 URLs (30 sites × homepage + contact page). The leader, on its defaults, scraped **54/60 pages (29/30 sites)** in one 45 s run. Ours on its defaults (1024 MB, 300 s timeout) **TIMED OUT after 13/60 pages**. The log shows memory at 95% and CPU overloaded, so concurrency stayed at 1 and each page took about 20 s. At 4096 MB (the leader's default) and a 1800 s timeout, ours returned **28/60 usable pages**. The other 26 pages were 20 navigation timeouts at 45 s, 7 HTTP 403s, and 6 pages that were charged but unusable (see below). Re-running only the 26 failures at concurrency 3 recovered 13 more, so the total was **40/60 pages (24/30 sites)** across two runs.
- **Contacts found: ours ahead where pages loaded.** Across the 30 sites, with ours' two 4 GB runs combined: valid emails **28 ours vs 20 leader**, with email recall against the reference of **0.79 vs 0.59** (n=34 reference emails). Valid phones were **47 vs 33**, with phone recall of **0.58 vs 0.41** (n=81 reference phones). Sites with ≥1 valid email: **18 vs 17**. Sites with ≥1 valid phone: **20 vs 16**.
- **Cloudflare email protection: ours only (n=8 emails on 3 sites).** pumpsupplies.co.uk (5), banda.co.uk (2) and valvesonline.co.uk (1) hide their emails behind Cloudflare's `data-cfemail` obfuscation. The leader returned none of the 8. Ours renders the page, so the decoding script runs, and across its runs it returned all 8 (pumpsupplies in the 4 GB run, banda and valvesonline in the 1 GB run). No single ours run got all 8, because banda rendered as an empty challenge page in the 4 GB run. The leader also missed 5 plain-text emails on ibstock.co.uk.
- **Precision: both clean on emails; the leader's phones have a noisy second field.** Emails: 0 junk from either Actor (ours 28/28, leader 20/20). None of the expected junk (`x@2x.png`, `sentry@`, `example@`) appeared on these 60 pages. Phones: ours 47/47 valid (UK-only regex). The leader's `phones` field was 33/34 valid; the junk one was a `tel:8884567890` template placeholder. Its `phonesUncertain` field held 151 values: 92 matched reference phones, 41 were not phones (company numbers like `05125479`, years like `1991-2026`), and 18 had a valid phone shape but were not on the reference (VAT-like digit groups).
- **UK vs non-UK phones.** Ours returned 0 non-UK numbers, by design (UK regex). The reference has 12 non-UK numbers (11 on extech.com's contact page, 1 Irish number on pumpsupplies). The leader returned 1 (the Irish number).
- **Sites that block plain requests (n=10): no win for ours.** Usable pages were ours 11/20 vs leader 15/20 on defaults, 15/20 with residential proxy ($0.064), and 16/20 in browser mode ($0.084). Seven of these 10 sites blocked plain fetches from our VPS (datacenter IP). All 10 passed a plain curl from withheld (residential IP). So the blocking here is mostly by IP, not "needs JavaScript". The leader's plain HTTP through Apify's datacenter proxy got through on most of them. Ours' residential option ($0.1782/page on this tier, so $3.56 for 20 pages) was not run, because it is over this test's cap.
- **Failure handling: mixed.** The leader writes a row for every URL (60/60). Its 6 failed rows are empty and carry **no reason field**; the reason ("Request blocked - received 403") is only in the run log. Ours writes a row with the error text (`Request blocked - received 403 status code.`, `page.goto: Timeout 45000ms exceeded`) when the run completes. In the default-settings run, which timed out, **47/60 pages got no row at all**. Ours also returned **7 charged pages with nothing usable**: 6 in the 4 GB run (5 bot-challenge interstitials titled "Loading https://…" with 0 text chars and empty `challenge_signals`, plus 1 HTTP 520) and 1 in the 1 GB run. Its event description says "Pages that fail to load are not charged". For challenge interstitials and 5xx pages, that is not what happened.
- **Price: leader much cheaper.** On this STARTER (BRONZE) account: leader **$0.112 for 60 pages = $0.0037 per site, $0.0021 per valid contact**. Ours (4 GB + retry) **$0.550 = $0.0183 per site, $0.0073 per valid contact**. List prices per page are ours $0.0117 vs leader $0.002 + $0.001/GB start.
- **Speed: leader much faster.** Run seconds: leader 45 s for 60 pages, with an average of 4.4 s per finished request. Ours at 4 GB took 496 s for 60 pages (19.8 s per finished request, 38.2 s per failed one), plus 478 s for the 26-page retry. The default 1 GB run averaged 20.5 s per page at concurrency 1. Per-site medians cannot be read from batch runs; see Limits.
- **Extra fields.** The leader returns social links (LinkedIn, X, Facebook, Instagram, YouTube and more) on 24/30 sites, plus `phonesUncertain`. Ours returns rendered page text, title, final URL, HTTP status, challenge signals and optional rendered HTML. The leader returns no page text.

**Claims the numbers support for the listing:**
- "Reads emails hidden by Cloudflare email protection" (8/8 across runs on 3 sites where the leader returned 0/8). This is narrow, but it is real.
- "Returns the page text with the contacts" (a field the leader lacks).
- "Failed pages come back as a row with the error", but only when the run completes. It is not true for runs that hit the timeout.
- "Phones are UK-format only" (a scope statement, not a quality claim).

**Claims the numbers do NOT support:**
- "Better than / more complete than the Contact Details Scraper." The leader covered 29 sites vs 24, at 1/5 the price per site.
- "Works on sites that block plain requests." On the 10 such sites, the leader's plain HTTP via Apify proxy loaded more pages (15/20) than ours (11/20).
- "Faster" or "cheaper": both are false by large margins.
- "Pages that fail to load are not charged": 7 unusable pages were charged.
- "No junk": the email side was clean on this sample, but no `x@2x.png` / `sentry@` cases appeared, so no comparison was possible.

**What the test says to fix (facts for whoever owns the Actor):** at defaults (1024 MB, 300 s, concurrency 10), 60 pages time out with 47 pages and no rows. Challenge interstitials titled "Loading …" with 0 text chars are not caught by `challenge_signals` and are charged. On 4 GB, concurrency 10, 20 of 60 pages hit the 45 s navigation timeout, and half of those also failed again at concurrency 3.

## Method

1. **Inputs (30 sites).** Each site was chosen by a written rule from org artefacts and probed with a plain fetch from this VPS on 2026-11-09 (`scripts/probe.py`, `ref/candidate-probe.jsonl`; 632 candidates probed).
   - *easy (10):* tier-1 proven sites (`alternatives/data/site-text/2026-10-23-tier1-batch1/proven-sites.tsv`), `.co.uk`, in rank order. Each had a plain fetch of 200, at least 1,500 visible chars, at most 20 `<script>` tags and a same-site link with "contact" in the path. trade.evo-stik.co.uk was dropped because its only "contact" link was a product page (`contact-adhesives`), and the next eligible site was used instead.
   - *mixed (10):* the 10 companies in Lin's `reports/for-lin/li-employees-valves-2026-11-09.tsv`, with sites from Scout's `alternatives/data/site-text/2026-10-20-valves/sites.tsv`.
   - *hard (10):* from Scout/Pixel's js-fallback targets (`alternatives/data/site-render/batches/2026-10-26-js-fallback/targets-why.tsv`, the sites that were 403 or thin on plain fetch on 10-26). These were re-probed on 11-09, and 10 were chosen to cover different block types: Cloudflare "Just a moment", empty 403, a WAF "Server Error" 403, Apache 403, Imunify "One moment, please", "Verifying Your Connection", and one splash page. Universities, gov.uk and non-company hosts were excluded.
   - *Second page:* the contact page, found from the plain-fetch homepage link. Where the homepage blocked both plain fetch and a local headless render from the VPS, the link came from a plain curl from withheld (residential IP). hobbsvalve.com's "Contact" is a Wix dialog with no page, so its about page (`/who-are-we`) was used.
2. **Same pages, no crawling.** Both Actors got exactly the same 60 URLs, and link-following was turned off in both. This tests extraction and fetching on identical pages. It does **not** test how each one picks links on its own.
3. **Reference.** I fetched the same 60 URLs three ways, none of them through either Actor (`scripts/ref_fetch.py`):
   - plain urllib from this VPS (datacenter IP; 49/60 were 200);
   - plain curl from withheld (residential UK IP; 60/60 were 200, but awmidgley returned a challenge page);
   - local headless Chromium from this VPS (47/60 were 200).

   The reference set per site is the union over the three fetches (`scripts/contacts.py`, `scripts/build_ref.py`). It includes `mailto:`/`tel:` hrefs, Cloudflare `data-cfemail` decoding, and emails/phones in visible text. Phones are normalised to UK national form (0…) or +international. Junk rules (asset filenames, sentry/wixpress/example domains, placeholder local parts, hex ids) apply to both the reference and the Actors. Result: **34 emails and 81 phones** on 30 sites. Five sites have no email on either page (form-only), and achech has no phone.
4. **Manual check (10 sites).** I read the page text around every contact on wincanton, warburtons, riddingtonspayroll, pumpsupplies, krohne, extech, stuart-turner, banda, achech and awmidgley (`scripts/ctx.py`). Findings:
   - wincanton and warburtons are form-only, with no email or phone on either page.
   - riddingtons has a hidden `tel:8884567890` placeholder, which was removed from the reference and counted as junk when returned.
   - pumpsupplies has 5 Cloudflare-protected emails, 8 UK numbers and 1 Irish number.
   - krohne has no contact in its static HTML.
   - extech lists 11 non-UK office numbers.
   - stuart-turner also lists an Irish freephone "1800 4567 01" with no prefix, which is not in the reference.
   - banda's emails are Cloudflare-protected.
   - achech shows 2 emails and no phone.
   - awmidgley returns a challenge page to withheld but rendered from the VPS.

   One accept was recorded: the leader's browser run on krohne was geo-redirected to the US page and returned `info@krohne.com` and +1 800 356 9464. Both are real KROHNE contacts, but from a page other than the one requested. All hand judgements are in `ref/manual.json`.
5. **Runs** (Hustle org `mathtechustle`, STARTER/BRONZE; the org does not own `mathtec/site-render`, so it pays the public price). Every run was async (`POST /v2/acts/<id>/runs`, then poll) with `maxTotalChargeUsd`. No Actor settings or prices were changed.

   | label | Actor | input (exact, plus the 60 or 20 `startUrls`) | run options | cap |
   |---|---|---|---|---|
   | ours | mathtec/site-render | `followLinkWords: []`, `maxPagesPerSite: 2`; the rest default (datacenter proxy, navigationTimeoutSecs 45, waitAfterLoadMillis 1500, maxConcurrency 10) | defaults: 1024 MB, 300 s | $0.75 |
   | ours_4g | same | same | memory 4096, timeout 1800 | $0.75 |
   | ours_retry | same | the 26 URLs that ours_4g returned as error rows, plus `maxConcurrency: 3` | 4096 MB, 1800 s | $0.35 |
   | leader | vdrmota/contact-info-scraper | `maxDepth: 0`, `maxRequestsPerStartUrl: 1`, `maxRequests: 60`, `mergeContacts: true`, `proxyConfig: {useApifyProxy: true}`; the rest default (`useBrowser: false`, `considerChildFrames: true`, no leads, no social enrichment) | defaults: 4096 MB | $0.50 (its Store minimum) |
   | leader_browser | same, 10 hard sites | as leader, plus `useBrowser: true`, `maxRequests: 20` | defaults | $0.50 |
   | leader_resi | same, 10 hard sites | as leader, plus `apifyProxyGroups: ["RESIDENTIAL"]`, `maxRequests: 20` | defaults | $0.50 |

6. **Metrics.**
   - A page is *usable* if, for ours, there is no error, status 200, text chars > 0 and no challenge signal; for the leader, `scrapedUrls` is non-empty.
   - Contacts are counted per site as the union over its 2 pages.
   - *Valid* means in the reference or hand-accepted. *Junk* means a junk rule, a non-phone shape or a hand reject. Nothing was left unjudged.
   - Precision = valid / returned. Recall = valid ∩ reference / reference.
   - The leader is scored on `phones`; `phonesUncertain` is reported separately.
   - $ is `usageTotalUsd` per run, after the runs settled.

## Inputs (30)

Columns: homepage; contact (second) page; why chosen; reference emails/phones (n).

| # | group | homepage | second page | why chosen | ref E/P |
|---|---|---|---|---|---|
| 1 | easy | wincanton.co.uk | /contact-us/ | tier-1 rank order; plain 200, 11.6k chars | 0/0 |
| 2 | easy | warburtons.co.uk | /contact/ | same rule | 0/0 |
| 3 | easy | ibstock.co.uk | /contact-us | same rule | 5/9 |
| 4 | easy | haystravel.co.uk | /contact-us/ | same rule | 0/4 |
| 5 | easy | volkerwessels.co.uk | /en/contact-us | same rule | 2/2 |
| 6 | easy | arlafoods.co.uk | /contact-and-help/ | same rule | 0/3 |
| 7 | easy | holcim.co.uk | /contact | same rule | 1/9 |
| 8 | easy | candopersonnel.co.uk | /contact | same rule | 1/1 |
| 9 | easy | bullough.co.uk | /contact.html | same rule | 1/1 |
| 10 | easy | riddingtonspayroll.co.uk | /contact | same rule (replaces evo-stik) | 1/1 |
| 11 | mixed | valves.co.uk (Oliver Valves) | /contact-us/ | Lin valves TSV | 1/1 |
| 12 | mixed | pumpsupplies.co.uk | /contact-us/ | Lin valves TSV | 5/9 |
| 13 | mixed | hobbsvalve.com | /who-are-we (Wix contact is a dialog) | Lin valves TSV | 1/1 |
| 14 | mixed | valvesonline.co.uk | /contact | Lin valves TSV | 2/2 |
| 15 | mixed | colson.co.uk | /contact/ | Lin valves TSV | 1/1 |
| 16 | mixed | ttpumps.com | /contact/ | Lin valves TSV | 1/1 |
| 17 | mixed | flowtech.org.uk | /contact-us | Lin valves TSV | 2/8 |
| 18 | mixed | achech.co.uk | /contactus | Lin valves TSV | 2/0 |
| 19 | mixed | halehamilton.com | /contact-us/ | Lin valves TSV | 1/1 |
| 20 | mixed | leengatevalves.co.uk | /contact/ | Lin valves TSV | 1/2 |
| 21 | hard | extech.com | www.extech.com/contact-us/ | VPS plain 403, Cloudflare "Just a moment" | 0/13 |
| 22 | hard | uk.krohne.com | www.krohne.com/en-uk/contact | VPS plain and render 403, Cloudflare | 0/0 |
| 23 | hard | kalair.co.uk | www.kalair.uk/eform/contact | VPS plain 403, empty body | 0/2 |
| 24 | hard | hatchercomponents.com | hatchercomp.co.uk/contact-us/ | VPS plain 403 "Server Error" WAF | 1/1 |
| 25 | hard | serfilco.co.uk | /contact/ | VPS plain 403 "Server Error" WAF | 1/1 |
| 26 | hard | talbotdesigns.co.uk | /contact-us/ | VPS plain 403 Forbidden | 1/2 |
| 27 | hard | stuart-turner.co.uk | /contact-us/ | "One moment, please…" challenge to plain fetch | 0/2 |
| 28 | hard | banda.co.uk | /contact/ | "One moment, please…" challenge | 2/1 |
| 29 | hard | awmidgley.co.uk | /contact-us/ | "Verifying Your Connection" (Anubis) | 0/2 |
| 30 | hard | foscos.co.uk | /contact.php | thin splash page (122 visible chars) | 1/1 |

## Results

| run | $ | run s | usable pages | sites ≥1 usable page | sites ≥1 email | sites ≥1 phone | valid emails (prec / recall) | valid phones (prec / recall) | non-UK phones | $ per site | $ per valid contact |
|---|---|---|---|---|---|---|---|---|---|---|---|
| ours, defaults (TIMED-OUT) | 0.152 | 300 | 12/60 | 7 | 5 | 7 | 7 (1.00 / 0.21) | 11 (1.00 / 0.14) | 0 | 0.0051 | 0.0085 |
| ours_4g | 0.398 | 496 | 28/60 | 18 | 13 | 14 | 14 (1.00 / 0.38) | 27 (1.00 / 0.33) | 0 | 0.0133 | 0.0097 |
| **ours_4g + retry** | **0.550** | **974** | **40/60** | **24** | **18** | **20** | **28 (1.00 / 0.79)** | **47 (1.00 / 0.58)** | 0 | **0.0183** | **0.0073** |
| **leader, defaults** | **0.112** | **45** | **54/60** | **29** | **17** | **16** | **20 (1.00 / 0.59)** | **33 (0.97 / 0.41)** | 1 | **0.0037** | **0.0021** |
| leader_browser (10 hard) | 0.084 | 203 | 16/20 | 9 | 5 | 5 | 6 (1.00 / 0.83) | 6 (1.00 / 0.20) | 1 | 0.0084 | 0.0070 |
| leader_resi (10 hard) | 0.064 | 49 | 15/20 | 9 | 4 | 4 | 4 (1.00 / 0.67) | 5 (1.00 / 0.20) | 0 | 0.0064 | 0.0071 |

By group, ours_4g+retry vs leader (defaults); $ and seconds are per run, so they are not split by group:

| group | usable pages ours / leader | sites ≥1 email ours / leader | sites ≥1 phone ours / leader | email recall ours / leader | phone recall ours / leader |
|---|---|---|---|---|---|
| easy (10) | 19 / 20 | 6 / 5 | 8 / 7 | 1.00 / 0.55 (n=11) | 0.83 / 0.57 (n=30) |
| mixed (10) | 10 / 19 | 7 / 8 | 7 / 5 | 0.71 / 0.59 (n=17) | 0.58 / 0.42 (n=26) |
| hard (10) | 11 / 15 | 5 / 4 | 5 / 4 | 0.67 / 0.67 (n=6) | 0.28 / 0.20 (n=25) |

Per site (ours_4g+retry vs leader): usable pages, then valid emails/phones.

| # | site | ref E/P | ours pages | ours E/P | leader pages | leader E/P |
|---|---|---|---|---|---|---|
| 1 | wincanton.co.uk | 0/0 | 1 | 0/0 | 2 | 0/0 |
| 2 | warburtons.co.uk | 0/0 | 2 | 0/0 | 2 | 0/0 |
| 3 | ibstock.co.uk | 5/9 | 2 | 5/5 | 2 | 0/7 |
| 4 | haystravel.co.uk | 0/4 | 2 | 0/3 | 2 | 0/4 |
| 5 | volkerwessels.co.uk | 2/2 | 2 | 2/2 | 2 | 2/2 |
| 6 | arlafoods.co.uk | 0/3 | 2 | 0/3 | 2 | 0/1 |
| 7 | holcim.co.uk | 1/9 | 2 | 1/9 | 2 | 1/0 |
| 8 | candopersonnel.co.uk | 1/1 | 2 | 1/1 | 2 | 1/1 |
| 9 | bullough.co.uk | 1/1 | 2 | 1/1 | 2 | 1/1 |
| 10 | riddingtonspayroll.co.uk | 1/1 | 2 | 1/1 | 2 | 1/1 (+1 junk) |
| 11 | valves.co.uk | 1/1 | 2 | 1/1 | 2 | 1/0 |
| 12 | pumpsupplies.co.uk | 5/9 | 1 (home: HTTP 520, charged) | 5/8 | 2 | 0/5 |
| 13 | hobbsvalve.com | 1/1 | 2 | 1/1 | 2 | 1/0 |
| 14 | valvesonline.co.uk | 2/2 | 0 (timeout; contact page flagged captcha) | 1/2 | 2 | 0/0 |
| 15 | colson.co.uk | 1/1 | 1 | 1/1 | 2 | 1/1 |
| 16 | ttpumps.com | 1/1 | 0 (403, timeout) | 0/0 | 1 | 1/1 |
| 17 | flowtech.org.uk | 2/8 | 0 (timeout ×2) | 0/0 | 2 | 2/3 |
| 18 | achech.co.uk | 2/0 | 1 | 2/0 | 2 | 2/0 |
| 19 | halehamilton.com | 1/1 | 2 | 1/1 | 2 | 1/0 |
| 20 | leengatevalves.co.uk | 1/2 | 1 | 0/1 | 2 | 1/1 |
| 21 | extech.com | 0/13 | 0 (timeout ×2) | 0/0 | 1 (empty row) | 0/0 |
| 22 | uk.krohne.com | 0/0 | 2 | 1/0 (accepted) | 0 (empty rows) | 0/0 |
| 23 | kalair.co.uk | 0/2 | 2 | 0/2 | 2 | 0/0 |
| 24 | hatchercomponents.com | 1/1 | 1 (403) | 1/1 | 1 (empty row) | 1/0 |
| 25 | serfilco.co.uk | 1/1 | 1 (403) | 1/1 | 1 (empty row) | 1/0 |
| 26 | talbotdesigns.co.uk | 1/2 | 2 | 1/2 | 2 | 1/1 |
| 27 | stuart-turner.co.uk | 0/2 | 0 (2 charged "Loading" pages, 0 chars) | 0/0 | 2 | 0/2 |
| 28 | banda.co.uk | 2/1 | 0 (2 charged "Loading" pages, 0 chars) | 0/0 | 2 | 0/1 |
| 29 | awmidgley.co.uk | 0/2 | 1 (home charged, 0 chars) | 0/0 | 2 | 0/0 |
| 30 | foscos.co.uk | 1/1 | 2 | 1/1 | 2 | 1/1 |

Ours' 1 GB default run did reach stuart-turner (0/2), banda (2/1), awmidgley (0/2), talbotdesigns, foscos, hobbsvalve and valvesonline (2/2) before it timed out. So the same site rendered in one ours run and came back as an empty challenge page in the next.

## Spend

- **This test: $0.9625**, attributed per run: ours $0.1521 + $0.3978 + $0.1521 = $0.702; leader $0.112 + $0.084 + $0.064 = $0.260. The hard cap was $1.50.
- **`/v2/users/me/limits`:** `monthlyUsageUsd` was 6.5739 before and 7.5364 after, a delta of **$0.9625**. That equals the sum of the six run ids, so no other org runs fell in the window (09:04–09:28Z).
- Every run passed `maxTotalChargeUsd`. The leader enforces a **$0.50 minimum** cap per run (`minimalMaxTotalChargeUsd`), so its real spend was bounded by `maxRequests`.

## Limits of this test

- One pass per configuration, at one time of day, on 30 sites. The ours results varied between runs on the same sites (challenge pages, timeouts), so expect noise.
- The reference was built partly with a headless browser and visible-text extraction, which is the same approach ours uses. That favours ours on recall. The leader's misses that are not artefacts of this are the Cloudflare-protected emails (8) and ibstock's 5 plain-text emails.
- Per-site medians are not available: both Actors ran each configuration as one batch run. Seconds are run time and the crawler's average per-request duration.
- Link selection was not tested; both got explicit URLs. Ours' `followLinkWords` and the leader's depth crawl may pick different pages in normal use.
- Ours' residential option was not run (price). The leader's residential and browser modes were run only on the 10 hard sites.
- "Junk" emails such as `x@2x.png` and `sentry@` did not occur on these pages for either Actor, so this sample does not test that claim.

## Files

All under `/home/a/Projects/swarm/agents/hustle/reports/market/h2h-site-render/`:
- `inputs.json`: the 30 sites, pages, why chosen, source and probe.
- `ref/candidates.tsv`, `ref/candidate-probe.jsonl`: the 632-candidate pool and its plain-fetch probe.
- `ref/html/{vps-plain,withheld-plain,vps-render}/NN.html` (+ `.txt` for render), `ref/fetch-meta.json`: the independent fetches.
- `ref/reference.json`, `ref/manual.json`, `ref/unjudged.json` (empty at the end): reference sets and hand judgements.
- `runs/{ours,ours_4g,ours_retry,leader,leader_browser,leader_resi}.json`: input, cap, settled run object, dataset items and log.
- `per-site.json`, `summary.json`: computed metrics.
- `actors/*.json`: each Actor's metadata, pricing and latest-build input schema at test time.
- `limits-before.json`, `limits-before-pass2.json`, `limits-after.json`: account usage snapshots.
- `scripts/`: `apify.py`, `getact.py`, `probe.py`, `select_inputs.py`, `finalize_inputs.py`, `ref_fetch.py`, `contacts.py`, `build_ref.py`, `ctx.py`, `runner.py`, `analyse.py`. The token is read from `.secrets/apify-org.env` at run time and is never written to disk (scanned: no token in outputs).
