Hafiza OS
The integration narrows model judgment to candidate selection over source-checked records, with privacy checks, input and question budgets, caching, deadlines, and a local fallback around the call.
Added to Jevfast
How it uses Jev
Hafiza uses Jev as an optional advisor over verified knowledge cards and memory-catalog candidates. Given a query, facets, and eligible candidate records, it asks typed questions about candidate relevance and returns scores, distributions, or choices. Local retrieval and source checks remain in code; Jev scores can gate or rank candidate IDs, and records still require local validity checks. Jev does not write canonical memory or grant user approval.
What Jev decides
Example: score a candidate memory against a queryExample answers · not a recorded Jev response · Source ↗
Question 1 · f0 c0
YOUR APP
INSTRUCTION
Does candidate 0 directly support facet 0 of the current request, within its scope, domain, conditions, and exceptions?
STATE
Source-defined request state contains query, facets, candidate cards (id, title, statement, scope, domains), and optional string-valued previous_user/cwd/project fields. This illustrative request has one facet and one candidate, therefore one question key f0_c0.
JEV · SCORE
0
App workflow
- Prepare eligible candidates
Local retrieval builds candidate cards from records that pass existing scope, source, and validity rules.
- Package a bounded request
The client places the query, facets, candidates, and permitted optional state fields in one shared state, then creates one typed question per facet and candidate.
- Use the result under local policy
The caller consumes candidate scores or distributions only in configured modes; it rechecks records and falls back to local retrieval when Jev is off, degraded, or results are unusable.
Why it is interesting
The integration narrows model judgment to candidate selection over source-checked records, with privacy checks, input and question budgets, caching, deadlines, and a local fallback around the call. It separates retrieval advice from ownership of stored facts.