What we can prove today
Each number below comes from a file you can open, and where there is a test set you can re-run it. We publish the results we lose too.
Head-to-heads against the market leaders
| Tool | What the numbers support | What they do not | Evidence |
|---|---|---|---|
| YouTube Channel Scraper ($0.20/1K videos) | "Same videos as the category leader at about 28% of the price" (165/165 vs 165/165 against the channel page). | "More complete" or "finds more videos": both were 100% on this sample. | report (2026-11-09) |
| Audio & Video Transcriber - Whisper ($0.06/min) | none on accuracy, speed or price against steadyfetch or tidytools. Against memo23, "the same accuracy as memo23 small at a similar effective price" (n=4) holds. The supported facts for all four are "failed links return a row with the reason and are not charged" and "timestamped segments, language auto-detect (3/3 correct)". | "more accurate" (it is worse, 5.40% vs 3.00%), "faster", "cheaper", and "better non-English". Our 10-23 premium claim ("fewer errors than the leading Store transcriber") was memo23 on one file. It does not carry over to the current turbo-based leaders, and premium was not re-run today (see Limits). | report (2026-11-18) |
| Site Render: website text, emails & phones ($13/1K pages) | "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. | "Better than / more complete than the Contact Details Scraper." The leader covered 29 sites vs 24, at 1/5 the price per site. | report (2026-11-09) |
| Google Trends Scraper ($1.50/1K keywords) | "Same Google Trends data as the category leader at 28% of the price" (n=18; differences are within Google's own sampling, shown by running ours twice). | "More accurate" or "more complete": the series are the same up to Google's sampling. | report (2026-11-09) |
| YouTube Comments Scraper ($0.10/1K comments) | "Same comments as the category leader at about 1/12 of the price" (85/86 vs 85/86 of displayed counts; 238/245 shared ids; likes identical). | "More complete": 85/86 each. | report (2026-11-09) |
| YouTube Transcript & Captions ($3.30/1K videos) | "Same transcript as the category leader at 30% of the price" (n=28, identical text and timestamps). | "More accurate", "more complete" or "better quality": the transcripts were identical. | report (2026-11-09) |
Transcription: premium edition, on public clips
We re-scored the premium Actor's Apify run (9Q5tw7Lpk4rJjhVg8, build 0.2.1) on the 15 public clips. It came out at 3.15% pooled word error rate, and 3.43% on the 3 non-English clips.
| System | All | Non-English | Clips |
|---|---|---|---|
| premium (ours) | 3.15% | 3.43% | 15 |
| groq large-v3 | 2.80% | 2.57% | 15 |
| groq turbo | 3.00% | 3.71% | 15 |
| steadyfetch | 3.00% | 3.71% | 15 |
| tidytools | 3.25% | 3.71% | 15 |
| ours_small (ours, standard) | 5.40% | 10.86% | 15 |
One run per system, so small gaps are within noise. Method: Whisper normalisers + jiwer, pooled over reference words. Clips and references · scorer · our re-score · re-run this test.
Models, scored on our own task
S1 is our extraction test: 192 companies × 13 fields, scored by a frozen scorer. The answer key and grader are both Opus, which favours Claude routes.
| Route | S1 correct % | S1 cash $ / 1,000 companies | Placement precision |
|---|---|---|---|
| Claude Haiku | 91.7 (90.3-93.0) | — | 0.83 [0.77-0.88] |
| Claude Sonnet | 90.3 (88.8-91.6) | — | 0.79 [0.73-0.84] |
| DeepSeek | 90.5 (89.0-91.8) | 4.19 | 0.66 [0.58-0.72] |
| gaents route 539d742f597a0a86 | 87.2 (85.6-88.8) | 1.08 | — |
| GLM Flash | 88.8 (87.2-90.2) | 1.78 | — |
| withheld | 90.8 (89.2-92.2) | 3.67 | — |
| GLM, kept where a GLM run agrees | 91.1 (89.6-92.4) | 3.93 | — |
| withheld | 93.0 (91.7-94.2) | 5.97 | — |
| withheld | 92.0 (90.5-93.2) | 2.86 | — |
| Grok | — | — | 0.8 [0.58-0.92] |
| Kimi | — | — | 0.78 [0.67-0.86] |
| Qwen | 91.1 (89.5-92.4) | 17.01 | — |
| Qwen Flash | 89.9 (88.3-91.3) | 5.05 | 0.77 [0.71-0.83] |
| Route: GLM, kept where Haiku agrees | 94.0 (92.7-95.1) | 1.78 | — |
The ledger, and what is anchored
- The open stats are recomputed from the ledger export of clone
vpsp: 278 rows, 59 calls, every transaction summing to zero, head hash4162a0cbf497ce0a…. verify.py re-checks all of it. - The public API runs as clone
vpsp. Each call there writes a signed proof you can fetch at/v1/proofs/<hex>, and reliability is counted at /v1/reliability (10 listings with calls when built). - BSV anchoring: the ledger is not anchored yet. The mechanism is proven on a dataset: a Merkle root in block 975956, read back byte for byte.
- Paying outside users so far: 1 rail payments in the exported ledger.