Receipt · March–July 2026
HikrLink, built by the loop
It is our own product. We built it with our own platform. Here is every number, including the ones that don't flatter us.
HikrLink is our own SaaS, built and maintained by an orchestrated agent platform under quality gates. We own it and we ship it, so this is a dogfooding receipt — not a client outcome. We would rather lead with that than have you find it out on your own.
Most AI-delivery claims are demos. This is four months of a real product with paying-grade infrastructure, and every execution scored by a grader we publish.
The numbers
Four HikrLink repos, 11 March to 12 July 2026, 8 model variants:
- 122 quality-scored agent executions
- 29 sprints
- 0.655 mean quality score
- 81% passed the 0.7 acceptance threshold
Every execution is scored by EQS, our open-source grader, on a 0–1 scale. 0.7 is the bar: below it, work goes back. The score is not the model marking its own homework.
Is 0.655 good? It's honest. The average execution lands below our own acceptance bar and gets reworked. Nearly a fifth fail outright. The point of a gate is that this is visible and caught, rather than shipped and discovered by a user.
The breakdown, including the bad parts
An average hides things. Here is every repo:
| Repo | Execs | Sprints | Mean | Pass |
|---|---|---|---|---|
| hk.ink (legacy core) Four months of continuous work. The real evidence. | 94 | 17 | 0.675 | 88% |
| HikrLink Website One day, one well-scoped job. Our best score, and our smallest sample. | 15 | 5 | 0.833 | 80% |
| HikrLink New repo, July. Our worst result. | 7 | 3 | 0.286 | 29% |
| HikrLink App (multi-repo) Cross-repo context. The platform handles it badly. | 6 | 4 | 0.333 | 33% |
Our two newest repos are our two worst — 0.286 and 0.333, passing 29% and 33%. Both from July. New repos give the loop no established patterns to follow, and the multi-repo case adds cross-repository context our platform handles badly today. That is the clearest weakness in this data, and burying it under a four-month average would make this page worthless.
Note also the shape of the good number: the website repo's 0.833 is our best mean and our smallest meaningful sample — one day, one tightly scoped job. Small samples flatter. The four-month, 94-execution legacy core at 0.675 / 88% is the number that actually carries weight, because it survived time.
We corrected this page
The first version of this receipt claimed 109 executions, 0.72 mean, 91% pass. Those numbers did not reproduce.
They summed two of the four repos and dropped the other two — which happened to be the newest and the worst. Nobody chose that to flatter the result; the query was never written down, only described, so it couldn't be checked. That is how a number drifts: not by lying, but by being unverifiable.
The corrected figures are lower. The claim now carries the SQL that produces it, so anyone can re-run it and get the same answer. We are telling you this because a receipt you can't audit is just a nicer-looking advert.
What this does and doesn't prove
- It proves the loop maintains, not just demos. Four months, one live product, 29 sprints.
- It doesn't prove it works on your codebase. This is our product, our conventions, our platform. Yours is different.
- It doesn't prove agents are better than your team. We never measured that and we're not implying it.
- It's a dogfooding receipt. We chose the work, the bar, and the grader. Weigh it accordingly.
Questions
Is HikrLink a client project?
No. HikrLink is our own SaaS. We own it, we ship it, and we built it with our own platform. That makes this a dogfooding receipt, not a client outcome, and we would rather say so at the top than have you discover it later.
What do the numbers cover?
122 quality-scored agent executions across 29 sprints and four HikrLink repos, between 11 March and 12 July 2026, using 8 model variants. Mean quality score 0.655; 81% passed the 0.7 acceptance threshold. Our analytics product, Hikr, is a separate codebase and is deliberately excluded.
What is a quality score here?
Every execution is scored by EQS, our open-source grader, on a 0 to 1 scale. 0.7 is the acceptance threshold: below it, the work goes back. The score is not a self-assessment by the model that did the work.
Is 0.655 good?
It is honest. It means the average execution lands below our own acceptance bar and gets reworked, and that 19% of executions failed outright. What matters is that failures are caught by the gate rather than shipped.
Why are your newest repos your worst?
The two most recent HikrLink repos score 0.286 and 0.333, with 29% and 33% pass rates. New repos have no established patterns for the loop to follow, and the multi-repo one adds cross-repository context the platform handles badly today. It is the clearest weakness in this data and we are not going to hide it under a four-month average.
Did you correct this page?
Yes. The first version of this receipt said 109 executions, 0.72 mean, 91% pass. Those figures did not reproduce: they summed only two of the four repos, and the two omitted were the newest and the worst. The corrected numbers are lower. The claim now carries the SQL that produces it.
Every figure traces to our claim registry and re-derives from a query we publish with the claim. Where we got it wrong, we say so. Related: The Checklists That Did Nothing · HikrLink: server-side tracking