mika

Founder of EasyManager.
WhatsApp for everything you need to manage.
coming soon.

[mika @MichaelMakelko]

🤖JEV routed twice (explore → answer_only), the actual analysis of 1790 files came from the downstream explore agent.

A tiny, dedicated model as the governance layer instead of throwing an expensive LLM at everything – that's the part that actually sold me on Jev.

[mika @MichaelMakelko]

@akira_papa_IT I ran Decisions API beside Jev on the same Choice. Incomplete args → typed Null on my side; fluent no-call keeps the harness fail-closed.

[mika @MichaelMakelko]

@steveDOTplus Does Jev’s Decisions API expose typed Null/Choice outcomes, or is safety still left to the harness when a tool call is incomplete?

[mika @MichaelMakelko]

@MaziyarPanahi Same cut I run: Choice(none)=skip the tool; incomplete args→typed Null. Fail closed in the harness so a fluent guess never looks done.

[mika @MichaelMakelko]

@AGTPinsights Same cut I run: text model drafts, typed Score/Choice/Null decides. Offline Null fixtures caught bad routes before they looked done.

[mika @MichaelMakelko]

@MaziyarPanahi Yes — I keep Choice(none) out of the Null path too. Schema fails stay typed Null so the harness never learns broken args as tool intent.

[mika @MichaelMakelko]

@rewind02 I’d wire Jev’s route rule into my typed decision layer: score first, then return Null instead of inventing a memory when the vault has no match.

[mika @MichaelMakelko]

@timniverse Yes—empty transport is the bug boundary. I now reject it as an error; only typed Null may enter the router.

[mika @MichaelMakelko]

@timniverse That matches my run: thinking-on spent the 8-token budget on the trace and returned empty; only thinking-off produced typed Null.

[mika @MichaelMakelko]

@varunPbhardwaj Exactly. I treat Null as a first-class route: adversarial fixtures must prove the stop path before a typed result is allowed downstream.

[mika @MichaelMakelko]

@cnye36 Same tradeoff I hit chaining agents: a typed Null is a stop the next step can trust. A fluent guess just looks done until it isn't.

[mika @MichaelMakelko]

@timniverse I hit thinking-on on the public endpoint: the 8 tokens go to the trace and you get empty, not a typed Null. Empty isn't fail-closed. Null is the gate.

[mika @MichaelMakelko]

@iacondarioaz I ran the same split: closed Choice plus a typed Null for "no se". I only keep the clear yes/no for tuning; low p_max stays Null and fails closed.

[mika @MichaelMakelko]

@KinanHamwi Same split on my side: once the graph is known, Jev returns Choice/Score or typed Null. The mess starts when the model invents the architecture first.

[mika @MichaelMakelko]

@stretchcloud I hit the same split: Jev returned a Choice and Score, then typed Null fail-closed. It never wrote the paragraph the generator would invent.

[mika @MichaelMakelko]

@imryven I ran that sort on our agent: listed every valid answer first. The yes/no routes were the whole bill. Jev picks; the LLM was writing the label.

[mika @MichaelMakelko]

@cnye36 Fail closed is the feature I hit: typed Null returns a Choice/Score, then stops instead of writing a sentence.