# affterms — Sourced Affiliate Terms for Dev and SaaS Tools Affiliate programme terms for developer and SaaS tools, every figure sourced and dated. Jev, a calibrated model, picks which pages a person re-reads. Live tool, by dave8172. Page: https://quirkyagents.com/wizard/projects/affterms/ ## Overview **In plain terms:** if you promote developer tools, the headline commission is the least useful number in the deal. Whether it recurs, how long the cookie lasts and how much you must earn before a payout decide what you get paid. Most lists give you a name and a percentage, with no date and no source. **affterms** is a directory of those terms — commission, recurring or one-time, cookie window, minimum payout, payout method and network — for **471 programmes** (as of 24 Sept 2026). Every figure was read off the vendor's own page, links back to it, and carries the date it was checked. Where a vendor doesn't publish something, the site says *not stated* rather than guessing. 50 programmes are free on the site; the full directory is a paid CSV + PDF. → [Visit affiliateprogramterms.com](https://affiliateprogramterms.com) --- ## The one rule **No figure goes in that wasn't read off the vendor's own page.** One invented number destroys a directory whose only claim is that its numbers were checked. So the question for every piece of automation here is not "is it good?" but "can it put a figure in a record?" — and the answer is built to be no. --- ## **How the model is used, and fenced** **Jev** (TypeSafe's System One model) is a second reader. It gets a page code has already fetched and answers yes/no questions about it: *is this the vendor's affiliate page? does it still say the cookie lasts 30 days?* That's all. It never fetches, converts units, compares dates, writes a note or supplies a value. **It decides whether a person looks, never what the record says.** The refresh re-reads every source, runs a free deterministic check that each published figure is still on the page, and asks Jev one yes/no per stated field — the question string matching can't ask. A person reads only what gets flagged. The one thing the tool may write is the `checkedOn` date, and only on a record where every check passed. **Why not let it extract?** It was measured doing exactly that. Picking values from candidate spans, it beat the regex extractor — right on 49% of records against 42% — and was still wrong on 7% of what it asserted at ≥0.9 confidence. The rule is no unsourced facts, and 7% is not zero. **Recurrence is asked in pairs.** A single *"does commission repeat?"* flagged 6 of 48 true records. Asking the claim and its opposite together, and flagging only when the opposite is affirmed: 0 of 48 true records flagged, 41–42 of 48 planted flips caught. --- ## Calibrated, not trusted Every threshold was measured against **68 pages a person had already read**, plus 16 vendor homepages as negatives, with planted false claims to catch a reader that just says yes. The model is pinned to `jev-1.13.0`, because an alias that moves would move every threshold silently. `calibrate_jev.py` re-runs the whole battery and **exits non-zero if a gate fails** — so a reworded question can't slip through. Some of what that measurement caught: - *"Commission **rate**"* was read literally, so a flat "$100 per referral" came back "not stated". - Adding "including saying it never expires" to the cookie question dropped agreement from 29/29 to 3/29. - A drop line of 0.2 would have thrown out a real referral page (Jobber, 0.19). It sits at 0.1. **A full re-check costs about four cents** — ~947k input tokens for 417 records, around five minutes, and the wall time is the fetches, not the model. --- ## The pipeline **harvest → triage → intake → refresh.** - **harvest** sweeps candidate domains for programme pages. 79% of its 23,166 requests were 404 path probes, so there was no page for a model to judge — Jev can't make the sweep faster, and doesn't pretend to. - **triage** asks two questions of each page that did come back — *is it a programme page? does it state a term?* — and routes it `drop`, `unsure`, `extract` or `no-terms`, so a page that never was one doesn't reach a browser read. - **intake** is the only way in, for a sweep and a public submission alike. A submitted figure that isn't on the page is rejected automatically, and **every figure a record carries is matched against the stored page text before it can merge** — whoever wrote it, agent included. - **refresh** re-reads everything and sends a person to what changed. On 24 Sept it re-read 413 records: 373 passed, 23 were read by hand, 6 were corrected. --- ## In production * **Site:** static, zero-dependency Node build on Cloudflare Pages. The build **aborts and deletes its output** if a paid record's figure reaches a page, or if a count was typed in rather than computed from the data. * **Payments:** Dodo, live since 22 Sept 2026. A webhook grants the download; a claim endpoint verifies the payment against Dodo's own API, so a late or lost webhook leaves nobody charged and empty-handed. * **Delivery:** files published once per edition to Cloudflare KV, served instantly and emailed through Resend. * **Tooling:** the authoring tools are Python and the build never imports them. Without a Jev key everything still runs — it just does less. --- ## Where it fits Same idea as [tally-aiagent](https://quirkyagents.com/wizard/projects/tally-aiagent/) and [doceval](https://quirkyagents.com/wizard/projects/doceval/): a model's answer is not evidence. Here it is allowed to point a person at a page, and nothing more.