From 275fbd4b177145da106e8924ecef729b75af31b1 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kov=C3=A1cs=20=C3=81d=C3=A1m?= Date: Mon, 7 Sep 2026 12:48:05 +0200 Subject: [PATCH] docs: mark shipped roadmap items (#85, #50) --- PUBLIC_ROADMAP.md | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) diff --git a/PUBLIC_ROADMAP.md b/PUBLIC_ROADMAP.md index 81e009a..63259a6 100644 --- a/PUBLIC_ROADMAP.md +++ b/PUBLIC_ROADMAP.md @@ -11,19 +11,17 @@ Open: - [Real batched inference](https://github.com/KRLabsOrg/LettuceDetect/issues/23): `predict_prompt_batch` currently loops one sample at a time; the plan is padded-batch tokenization with a single forward pass. - [Lighter installs](https://github.com/KRLabsOrg/LettuceDetect/issues/69): lazy top-level imports so the LLM detector never touches torch. How to slim the default install without breaking existing users is an open design question; input welcome on the issue. - [Define evidence aggregation for chunked inference](https://github.com/KRLabsOrg/LettuceDetect/issues/74): benchmark whether evidence in any chunk or every chunk is required before changing the current conservative behavior. -- [Map repeated LLM span strings to distinct answer occurrences](https://github.com/KRLabsOrg/LettuceDetect/issues/85) -Shipped in this milestone so far: the network-free unit suite as a merge gate ([#73](https://github.com/KRLabsOrg/LettuceDetect/issues/73)), the `lettucedetect` CLI ([#47](https://github.com/KRLabsOrg/LettuceDetect/issues/47)), the detector latency/throughput benchmark ([#58](https://github.com/KRLabsOrg/LettuceDetect/issues/58)), token output from the LLM detector ([#65](https://github.com/KRLabsOrg/LettuceDetect/issues/65)), and the `min_confidence` documentation ([#64](https://github.com/KRLabsOrg/LettuceDetect/issues/64)). +Shipped in this milestone so far: the network-free unit suite as a merge gate ([#73](https://github.com/KRLabsOrg/LettuceDetect/issues/73)), the `lettucedetect` CLI ([#47](https://github.com/KRLabsOrg/LettuceDetect/issues/47)), the detector latency/throughput benchmark ([#58](https://github.com/KRLabsOrg/LettuceDetect/issues/58)), token output from the LLM detector ([#65](https://github.com/KRLabsOrg/LettuceDetect/issues/65)), the `min_confidence` documentation ([#64](https://github.com/KRLabsOrg/LettuceDetect/issues/64)), and deterministic localization of repeated LLM spans ([#85](https://github.com/KRLabsOrg/LettuceDetect/issues/85)). ## Next: [Validation — Typed-span workflows](https://github.com/KRLabsOrg/LettuceDetect/milestone/3) The detectors already emit typed spans (category and subcategory per detected span). This milestone tests whether that localization improves real debugging, evaluation, monitoring, and agent workflows. - [Dataset-level hallucination-rate evaluation and reporting](https://github.com/KRLabsOrg/LettuceDetect/issues/56) -- [A Claude Code hook that flags hallucinations in agent answers](https://github.com/KRLabsOrg/LettuceDetect/issues/50) - [A stable HTTP contract for typed spans and detector selection](https://github.com/KRLabsOrg/LettuceDetect/issues/75) -Shipped: typed spans in the Streamlit demo ([#57](https://github.com/KRLabsOrg/LettuceDetect/issues/57)). +Shipped: typed spans in the Streamlit demo ([#57](https://github.com/KRLabsOrg/LettuceDetect/issues/57)) and the Claude Code Stop hook ([#50](https://github.com/KRLabsOrg/LettuceDetect/issues/50)). ## Exploring: [transferable span detection](https://github.com/KRLabsOrg/LettuceDetect/milestone/2)