The model learns from what actually happens
Astris HR does not stop at the score. Every time someone corrects the model or a real hiring result comes in, that information changes the next prediction — it is not filed away and forgotten. Real placement results adjust how much weight the model gives to its retention predictions. Corrections to parsed details and translations improve how the model reads similar documents next time.
Four kinds of feedback
Each kind is a clear signal the product collects, and each one improves a specific part of the model.
A reviewer rates a match. That rating helps the model refine how it ranks candidates and how it weighs the eight things it compares — skills, experience, logistics, retention, language, education, culture fit, and interview.
A real result — for example, whether a hire stayed for the first 90 days — feeds two improvements: it becomes a new training example that sharpens the retention prediction, and it adjusts how much weight that prediction should carry going forward.
A reviewer fixes a detail the model pulled from a document. That correction is saved and used as a worked example, so the model reads similar documents more accurately next time.
A corrected translation is saved the same way as a parse correction, so the next translation of that detail benefits from the fix.
What each kind of feedback improves
When feedback comes in, the model applies it to a specific part of the system — and reports back exactly what it changed.
| Match rating | Improves how the model ranks candidates and how it weighs the eight things it compares: skills, experience, logistics, retention, language, education, culture fit, and interview. |
| Placement outcome | Does two things. It adds a real example — the candidate's details plus what happened at 90 days — that sharpens the retention prediction, and it adjusts how much weight that prediction should carry, based on how well it has predicted real results. |
| Parse correction | Is saved as a worked example. The model draws on these saved corrections when reading new documents, so a fix measurably improves how it reads similar ones. |
| Translation correction | Is saved by source language, so the corrected text improves later translations of the same detail and related languages. |
How the loop closes
A correction or a real result is not just recorded — it is used by the very next prediction.
The key point is that every piece of feedback actually changes the model. Real placement results adjust how much weight the retention prediction carries — so the same weights the model uses to make a prediction are the weights that feedback produced. Corrections are added to the same set of examples the model already relies on, so the next reading of a similar document draws on the fix.
Today the model is trained and tested on synthetic (simulated) data; real placement and correction volume is still to come. The way feedback flows and updates the model is already working, but any accuracy figures elsewhere on this site come from internal testing on simulated data, not real-world results.
See Astris HR in your workflow
Run it against your candidate base or a sample JD. We'll set up a tenant in under a day and show you the matches your existing tools are missing.
Request a demo