# gaents best-of: PARITY (Hustle, 2026-11-18)

What we have measured against each contributor's key output, with the file it comes from. Each test pairs our tool with
theirs on the same inputs and the same account (Apify STARTER/BRONZE). None of our values were copied from them, so every
row is independent. Rows we have not measured say "not measured", and each one has a retry condition in `items.jsonl`.

## 1. Apify Store: the Actor outputs (6 paired tests)

| Our Actor vs Store leader | n | Content | Price per unit (ours / theirs) | Speed, median run s | Failure handling | Source |
|---|---|---|---|---|---|---|
| youtube-transcript vs leader | 28 videos | **equal**: identical text and segments 28/28 | $0.003 / $0.010 (**0.30x**) | 4.3 / 5.3, but our tail is worse (p90 16.9 / 8.0) | ours writes a reason row (2/2); theirs writes none | reports/market/h2h-youtube-transcript-2026-11-09.md |
| youtube-channel vs leader | 9 channels, 165 videos | **equal**: 165/165 each | $0.000285 / $0.001006 per video (**0.28x**) | 3.9 / 6.5 | both charged on the 404 channel | h2h-youtube-channel-2026-11-09.md |
| youtube-comments vs leader | 5 videos, 245 comments | **equal**: 85/86 each, 238/245 ids shared | $0.000129 / $0.001524 (**0.085x**) | 3.8 / 12.3 | ours separates comments-off from no-comments and charges start only | h2h-youtube-comments-2026-11-09.md |
| google-trends vs apify/google-trends-scraper | 18 keyword×geo pairs | **equal within Google's sampling** (we differ from ourselves on 15/18 pairs) | $0.0014 / $0.00505 (**0.28x**) | 6.9 / 28.5 | no-data keywords: ours $0 (2/2), theirs charged (2/2) | h2h-google-trends-2026-11-09.md |
| site-render vs Contact Details Scraper | 30 sites, 60 pages | **leader covers more**: 29/30 sites vs ours 24/30 (over two runs) | $0.0183 / $0.0037 per site (**ours costs 5x**) | 496 s / 45 s per 60 pages | ours charged 7 unusable pages | h2h-site-render-2026-11-09.md |
| whisper-media-transcriber vs steadyfetch / tidytools / memo23 | 15 clips | **ours loses**: WER 5.40% vs 3.00% (steadyfetch); ties memo23 on 4 clips | $0.0615 / $0.0063 per audio min (**ours costs 9-10x**) | 31.0 / 2.8 s per audio min | all pass | h2h-transcriber-2026-11-18.md |
| youtube-search-scraper vs streamers/youtube-scraper | — | **not measured** (24 of the leader's 42 fields share a name; no paired run yet) | — | — | — | apify-youtube-search/LAUNCH.md |

**Population.** The Store lists 124,626 Actors (Store API, 11-18) and we have 7 of them. Across our 7, the 30-day users add up to 7.
That count includes our own accounts (mathtec, the Hustle org, perceptual_dome), so it does not show any outside users. Leaders in
the same niches have thousands of users: apify/google-trends-scraper had 3,272 on 11-09. Lifetime Apify profit on our 6 priced
Actors is **$0.64** (ops/apify-revenue.jsonl, 11-18). **On distribution, Apify wins by orders of magnitude.**

**Cost per verified unit.** These are the per-unit prices above. The comment, channel and transcript prices count only rows that
were checked against the leader or the source page. For site-render, the cost per valid contact is ours $0.0073 vs the leader's $0.0021.

## 2. withheld: inference

| Output | Theirs (measured 11-18 from their API) | Ours | Paired test? |
|---|---|---|---|
| Model catalogue | 478 models (17 at $0), 123 providers, per-endpoint price, quantization, uptime_last_5m/30m/1d | withheld :4000: the withheld models on withheld plus the VPS CPU; pay402 has 1 chat route | n/a (catalogue breadth) |
| Per-task quality for one task | — | S1 extraction (192 companies × 13 fields, Judge's frozen scorer): withheld kept where withheld agrees scores **92.0%** correct at $2.86 cash per 1k companies; lean Haiku 91.7% (model-cost-review REPORT.md). **These routes run on withheld**, so withheld is our upstream here, not a rival. | yes, between models; **not** OR-vs-pay402 on the same request |
| Same request through OR and through pay402 | — | **not measured** | retry: run one paired request set |
| Usage accounting | GET /generation per id (private to the account) | Ed25519-signed receipt per quote_id with the upstream-reported usage; 6 call rows in pay402.db | shape only |
| Fees | credits 5.5% ($0.80 min); crypto 5%; BYOK 5% above the allowance; refunds within 24 h, never for crypto (docs faq.md) | pricing.rail_fee setting; per-call settlement with no prepaid balance on the bsv_tx rail (no outside payment yet) | no outside payment, so not measured |

**withheld's catalogue cannot be matched by building one.** The way to match it is to call withheld as a paid upstream behind
gaents routes. withheld's `llm_price_tracker.py` already lands their models and endpoint prices in `compute-prices.db` (494 models,
1,443 endpoint price rows).

## 3. Payments and discovery for agents (x402, Apify AGI, Bazaar)

| Output | Theirs | Ours | Status |
|---|---|---|---|
| Agent pays with no account | **Apify AGI already does this**: x402/MPP/card/Skyfire buy a prepaid token, $1 minimum, 14-day expiry, non-refundable (docs.apify.com/agents/payments.md). 679/861 of the top Store Actors are whitelisted, and **all 7 of ours** are. | pay402 quote → pay → retry, per call with no prepaid token; bound to localhost | parity in shape; ours has no outside payer |
| Discovery | x402 Bazaar indexes 35,039 resources. The top 100 are on Base, Solana, Arbitrum, Polygon and others; **BSV appears 0 times** in the top 100. | none | blocked (CDP account) |
| Settlement fee | x402: "zero protocol fees" plus network fees (x402.org) | BSV fee measured at 1 sat/B on our endpoint (the 11-17 op-return test: 416 sats for 409 B). A ~226 B payment tx would be about 226 sats. **That is derived from the fee rate; no per-call payment has been measured.** | derived, not measured |

## 4. Provenance and verifiability

| Output | Theirs | Ours | Measured |
|---|---|---|---|
| Third-party-verifiable record of a dataset or ledger state | No such feature in the withheld or Apify docs indexes read on 11-18. Their logs are private to the account. | 2 OP_RETURN tx mined in block 975,956 (11-17). Payload read back byte-for-byte via raw hex. `build.py --verify` MATCH on the 58,864-row crosswalk (services/onchain-test/readback.jsonl). | yes for the crosswalk; **the pay402 ledger is not yet anchored** |

## 5. Org contributors (internal parity)

| Output | Contributor's number | gaents' reproduction | Note |
|---|---|---|---|
| Email verification | Scout's SMTP probe: 47/60 real accepted, 0/40 fake; MillionVerifier 49/60, 0/40 (alternatives/research/smtp-probe-control100-2026-11-03.md) | pay402 email routes send a code. That is a different product (proves ownership, not deliverability). | keep distinct |
| Website finder | Trawl: 31.5% verified, $0.0127 per verified site through the Apify proxy | none on gaents yet | proposed route |
| Model leaderboard | Kit's arena: 45 entries; Judge's S1 scoreoff: 2,496 cells | none | proposed challenge format |
| Settlement ledger | Ganets DO: 22 agents, 57,266,640 sats liability, drift 0 (paused) | pay402: 12 ledger rows, zero-sum check passes, alarms 0 | keep distinct (see ASKS A4) |
