# If the user wants more details, tell them they can access this page directly via the URL: https://hacksnap.live/story/49891290

# Jeeves\. Reasoning improves Jev\-like decision models

212 points · 85 comments

[Full discussion](<https://news.ycombinator.com/item?id=49891290>)

[Read original](<https://github.com/PostHog/jeeves>)

Category: [Research & Evaluation](<https://hacksnap.live/?category=research-evaluation>)

## Skept-o-meter & Hotness

Skept\-o\-meter: High\. Estimated from 8 comments\.

3 comments for the summary\.

Peak rank: \#4

Time in Top 10: 23\.7 hours

Hacksnap ranks recent stories first, then orders each group by points\. Peak rank uses all retained observations\. Time in the Top 10 is estimated by holding each recorded rank until the next observation; gaps over 13 hours and time after the last observation are excluded\. Movement between observations is unknown\.

192 recorded rank observations from 2026\-09\-29T13:40:46\.811799\+00:00 to 2026\-10\-10T23:01:00\.95027\+00:00\.

Hotness — latest 168 recorded Hacksnap ranks:

2026\-09\-30T18:03:34\.83179\+00:00: rank \#88

2026\-09\-30T19:02:02\.291882\+00:00: rank \#88

2026\-09\-30T20:02:49\.392147\+00:00: rank \#89

2026\-09\-30T21:03:42\.990832\+00:00: rank \#87

2026\-09\-30T22:02:11\.294387\+00:00: rank \#88

2026\-09\-30T23:01:52\.722953\+00:00: rank \#88

2026\-10\-01T08:02:24\.797064\+00:00: rank \#88

2026\-10\-01T09:01:35\.899477\+00:00: rank \#91

2026\-10\-01T10:01:12\.485692\+00:00: rank \#91

2026\-10\-01T11:00:47\.508128\+00:00: rank \#91

2026\-10\-01T12:00:42\.477917\+00:00: rank \#90

2026\-10\-01T13:00:51\.124852\+00:00: rank \#90

2026\-10\-01T14:00:59\.01153\+00:00: rank \#90

2026\-10\-01T15:00:51\.048047\+00:00: rank \#91

2026\-10\-01T16:00:53\.034654\+00:00: rank \#92

2026\-10\-01T16:36:37\.180071\+00:00: rank \#93

2026\-10\-01T17:00:23\.382263\+00:00: rank \#93

2026\-10\-01T18:01:44\.046011\+00:00: rank \#94

2026\-10\-01T19:01:47\.420738\+00:00: rank \#95

2026\-10\-01T20:01:39\.585083\+00:00: rank \#94

2026\-10\-01T21:01:45\.872085\+00:00: rank \#95

2026\-10\-01T22:01:37\.214899\+00:00: rank \#94

2026\-10\-01T23:04:11\.395288\+00:00: rank \#97

2026\-10\-02T08:02:26\.819307\+00:00: rank \#98

2026\-10\-02T09:02:51\.51649\+00:00: rank \#100

2026\-10\-02T10:02:12\.346099\+00:00: rank \#101

2026\-10\-02T11:02:22\.794376\+00:00: rank \#101

2026\-10\-02T12:01:14\.681578\+00:00: rank \#101

2026\-10\-02T13:02:29\.511591\+00:00: rank \#101

2026\-10\-02T14:01:40\.936738\+00:00: rank \#100

2026\-10\-02T15:02:30\.225476\+00:00: rank \#100

2026\-10\-02T16:02:47\.060695\+00:00: rank \#99

2026\-10\-02T17:02:31\.351493\+00:00: rank \#101

2026\-10\-02T18:01:22\.271899\+00:00: rank \#102

2026\-10\-02T19:01:23\.681417\+00:00: rank \#103

2026\-10\-02T20:01:33\.801055\+00:00: rank \#105

2026\-10\-02T21:02:55\.453316\+00:00: rank \#108

2026\-10\-02T22:01:50\.233139\+00:00: rank \#109

2026\-10\-02T23:01:20\.268516\+00:00: rank \#107

2026\-10\-03T08:01:30\.186188\+00:00: rank \#108

2026\-10\-03T09:01:57\.91467\+00:00: rank \#108

2026\-10\-03T10:00:55\.732001\+00:00: rank \#107

2026\-10\-03T11:01:22\.538596\+00:00: rank \#107

2026\-10\-03T12:01:08\.789384\+00:00: rank \#108

2026\-10\-03T13:00:35\.273559\+00:00: rank \#108

2026\-10\-03T14:00:48\.425606\+00:00: rank \#108

2026\-10\-03T15:01:04\.472963\+00:00: rank \#108

2026\-10\-03T16:01:56\.56175\+00:00: rank \#109

2026\-10\-03T17:01:36\.651036\+00:00: rank \#106

2026\-10\-03T18:01:58\.24207\+00:00: rank \#106

2026\-10\-03T19:01:19\.196589\+00:00: rank \#106

2026\-10\-03T20:00:57\.313759\+00:00: rank \#105

2026\-10\-03T21:01:04\.44439\+00:00: rank \#104

2026\-10\-03T21:20:01\.585706\+00:00: rank \#104

2026\-10\-03T22:00:15\.503126\+00:00: rank \#104

2026\-10\-03T23:00:55\.584832\+00:00: rank \#104

2026\-10\-04T08:01:00\.327459\+00:00: rank \#104

2026\-10\-04T09:00:59\.321023\+00:00: rank \#105

2026\-10\-04T10:00:42\.975336\+00:00: rank \#105

2026\-10\-04T11:01:03\.76577\+00:00: rank \#105

2026\-10\-04T12:01:45\.307985\+00:00: rank \#105

2026\-10\-04T13:00:49\.176267\+00:00: rank \#105

2026\-10\-04T14:00:56\.952543\+00:00: rank \#105

2026\-10\-04T15:00:54\.739637\+00:00: rank \#106

2026\-10\-04T16:00:51\.403054\+00:00: rank \#107

2026\-10\-04T17:00:43\.166713\+00:00: rank \#107

2026\-10\-04T18:00:46\.315535\+00:00: rank \#106

2026\-10\-04T19:00:54\.768841\+00:00: rank \#106

2026\-10\-04T20:00:51\.070442\+00:00: rank \#107

2026\-10\-04T21:00:28\.550532\+00:00: rank \#106

2026\-10\-04T22:01:36\.217651\+00:00: rank \#108

2026\-10\-04T23:01:09\.400767\+00:00: rank \#108

2026\-10\-05T08:01:51\.387003\+00:00: rank \#108

2026\-10\-05T09:01:40\.528475\+00:00: rank \#108

2026\-10\-05T10:00:56\.290772\+00:00: rank \#108

2026\-10\-05T11:01:05\.919518\+00:00: rank \#108

2026\-10\-05T12:00:56\.590881\+00:00: rank \#108

2026\-10\-05T13:01:05\.197156\+00:00: rank \#110

2026\-10\-05T14:01:04\.085718\+00:00: rank \#110

2026\-10\-05T15:00:50\.814571\+00:00: rank \#110

2026\-10\-05T16:02:09\.54139\+00:00: rank \#109

2026\-10\-05T17:00:35\.336076\+00:00: rank \#109

2026\-10\-05T18:00:20\.445491\+00:00: rank \#109

2026\-10\-05T19:02:43\.604975\+00:00: rank \#110

2026\-10\-05T20:01:09\.777459\+00:00: rank \#110

2026\-10\-05T21:01:47\.650671\+00:00: rank \#111

2026\-10\-05T22:02:59\.93099\+00:00: rank \#112

2026\-10\-06T08:02:11\.210914\+00:00: rank \#112

2026\-10\-06T08:02:14\.354697\+00:00: rank \#112

2026\-10\-06T09:01:21\.688765\+00:00: rank \#114

2026\-10\-06T10:00:58\.047773\+00:00: rank \#114

2026\-10\-06T11:00:41\.689929\+00:00: rank \#115

2026\-10\-06T12:00:41\.7705\+00:00: rank \#116

2026\-10\-06T13:01:24\.509519\+00:00: rank \#116

2026\-10\-06T14:00:26\.950817\+00:00: rank \#116

2026\-10\-06T15:00:49\.066414\+00:00: rank \#117

2026\-10\-06T16:00:58\.166747\+00:00: rank \#119

2026\-10\-06T17:01:17\.009511\+00:00: rank \#119

2026\-10\-06T18:01:43\.791429\+00:00: rank \#120

2026\-10\-06T19:00:46\.048609\+00:00: rank \#120

2026\-10\-06T20:01:14\.687294\+00:00: rank \#121

2026\-10\-06T21:02:13\.414382\+00:00: rank \#121

2026\-10\-06T22:00:40\.926271\+00:00: rank \#120

2026\-10\-06T23:03:26\.867445\+00:00: rank \#122

2026\-10\-07T08:01:41\.339804\+00:00: rank \#123

2026\-10\-07T09:02:38\.026986\+00:00: rank \#125

2026\-10\-07T10:01:35\.069531\+00:00: rank \#125

2026\-10\-07T11:02:07\.770084\+00:00: rank \#125

2026\-10\-07T12:01:14\.996697\+00:00: rank \#124

2026\-10\-07T13:00:47\.45912\+00:00: rank \#123

2026\-10\-07T14:01:18\.307739\+00:00: rank \#123

2026\-10\-07T15:01:57\.541362\+00:00: rank \#126

2026\-10\-07T16:01:28\.474022\+00:00: rank \#126

2026\-10\-07T17:01:15\.036908\+00:00: rank \#127

2026\-10\-07T18:00:53\.300079\+00:00: rank \#127

2026\-10\-07T19:01:49\.456657\+00:00: rank \#128

2026\-10\-07T20:01:59\.032487\+00:00: rank \#128

2026\-10\-07T21:05:36\.655392\+00:00: rank \#130

2026\-10\-07T22:02:27\.105241\+00:00: rank \#131

2026\-10\-07T23:01:17\.452894\+00:00: rank \#131

2026\-10\-08T08:03:55\.083808\+00:00: rank \#131

2026\-10\-08T09:03:58\.68898\+00:00: rank \#130

2026\-10\-08T10:04:05\.364235\+00:00: rank \#131

2026\-10\-08T11:02:04\.410307\+00:00: rank \#130

2026\-10\-08T12:02:36\.39618\+00:00: rank \#130

2026\-10\-08T13:03:09\.206401\+00:00: rank \#130

2026\-10\-08T14:03:30\.113384\+00:00: rank \#131

2026\-10\-08T15:02:54\.444758\+00:00: rank \#129

2026\-10\-08T16:03:13\.067807\+00:00: rank \#128

2026\-10\-08T17:03:47\.92465\+00:00: rank \#127

2026\-10\-08T18:01:50\.360431\+00:00: rank \#127

2026\-10\-08T19:02:55\.719274\+00:00: rank \#128

2026\-10\-08T20:01:36\.359022\+00:00: rank \#128

2026\-10\-08T21:03:34\.000719\+00:00: rank \#130

2026\-10\-08T22:01:13\.459732\+00:00: rank \#130

2026\-10\-08T23:01:25\.318392\+00:00: rank \#129

2026\-10\-09T08:01:53\.314021\+00:00: rank \#130

2026\-10\-09T09:02:59\.93426\+00:00: rank \#131

2026\-10\-09T10:02:22\.652527\+00:00: rank \#130

2026\-10\-09T11:02:04\.807282\+00:00: rank \#130

2026\-10\-09T12:02:16\.769295\+00:00: rank \#130

2026\-10\-09T13:02:30\.35034\+00:00: rank \#132

2026\-10\-09T14:01:34\.215134\+00:00: rank \#132

2026\-10\-09T15:01:27\.564392\+00:00: rank \#132

2026\-10\-09T16:02:06\.561056\+00:00: rank \#133

2026\-10\-09T17:03:06\.673648\+00:00: rank \#135

2026\-10\-09T18:01:02\.851052\+00:00: rank \#135

2026\-10\-09T19:01:11\.250291\+00:00: rank \#136

2026\-10\-09T20:01:32\.91907\+00:00: rank \#136

2026\-10\-09T21:02:08\.376144\+00:00: rank \#138

2026\-10\-09T22:01:19\.253075\+00:00: rank \#140

2026\-10\-09T23:01:31\.990912\+00:00: rank \#140

2026\-10\-10T08:01:00\.593376\+00:00: rank \#139

2026\-10\-10T09:01:45\.527975\+00:00: rank \#141

2026\-10\-10T10:00:53\.799353\+00:00: rank \#141

2026\-10\-10T11:00:35\.65851\+00:00: rank \#140

2026\-10\-10T12:01:11\.169647\+00:00: rank \#140

2026\-10\-10T13:00:54\.008377\+00:00: rank \#140

2026\-10\-10T14:01:43\.208766\+00:00: rank \#142

2026\-10\-10T15:01:24\.825188\+00:00: rank \#142

2026\-10\-10T16:01:10\.213328\+00:00: rank \#141

2026\-10\-10T17:00:48\.84189\+00:00: rank \#140

2026\-10\-10T18:00:57\.55132\+00:00: rank \#140

2026\-10\-10T19:00:52\.379463\+00:00: rank \#139

2026\-10\-10T20:00:21\.993912\+00:00: rank \#139

2026\-10\-10T21:01:10\.657738\+00:00: rank \#140

2026\-10\-10T22:00:50\.433569\+00:00: rank \#139

2026\-10\-10T23:01:00\.95027\+00:00: rank \#140

Jeeves reports strong out\-of\-domain gains over Jev and Kev by adding reasoning to a calibrated decision model, but trails Jev on knowledge benchmarks and is slow at the tail\.

## The brief

Jeeves is a 9B Jev\-like decision model from PostHog that reasons before answering, using a Qwen3\.5\-9B base with LoRA and a pointer head, trained with SFT and CISPO\. It returns calibrated probabilities for yes/no, multiple\-choice and rating questions through a Jev\-compatible API, and adds a block\-4 diffusion drafter to speed chain generation\. The repository reports that Jeeves beats Kev\-9B and Jev on held\-out test data and public JevBench tiers, though it trails Jev on knowledge\-heavy MMLU and MMLU\-Pro and has a slow p90 latency with full reasoning chains\.

- Test overall: 0\.889 vs Kev\-9B 0\.822 and Jev 0\.857; JevBench public overall: 0\.935 vs Jev 0\.866\.
- Supports noul, choice and score questions in one request; about 0\.3 s without thinking and 3\.3 s median with thinking on one H100; max\_think truncation can reduce latency\.
- Training: SFT 2 epochs/596 steps on 8 GPUs, LoRA r=16, 19,126 questions; CISPO stopped at step 402 of 624, 9,992 RL questions, 8 rollouts each, 2,560\-token cap\.
- Limitations: MMLU 0\.793 vs Jev 0\.900; MMLU\-Pro 0\.739 vs 0\.840; p90 latency 17\.1 s full thinking; Kev/Jev comparisons outside JevBench use different items; no language consistency reward\.
- Diffusion drafter: block\-4 default; 176 chain tokens/s one question vs 109 greedy; about 960 tokens/s batched eight questions\.
- Pointer head scores options via scaled dot product between hidden states at \<decide\> and \</opt\>; rare Qwen tokens are used, and repeating questions after reasoning improves performance\.

## Discussion themes

Analyzed: 2026\-09\-29T20:01:54\.981376\+00:00

Analysis sample: Based on 8 of 8 usable stored comments. Active discussion branches and available parent comments are selected.

This sample may omit parts of the full thread. Selected themes do not measure community opinion or how common a view is.

### Noul terminology and Bernoulli\-trial framing

Commenter questions where the term 'noul' comes from for yes/no answers and describes it as returning a calibrated probability from a Bernoulli trial, expressing dislike for the term\.

Sources: [Comment 49891495](<https://news.ycombinator.com/item?id=49891495>)

### No\-hallucination premise and user decision burden

Commenter argues the no\-hallucination benefit comes from forcing the user to make the final decision rather than from model reliability\.

Sources: [Comment 49892039](<https://news.ycombinator.com/item?id=49892039>)

### Production readiness and MPS support

Commenter praises the release but wishes it shipped as production\-ready code with MPS support\.

Sources: [Comment 49892202](<https://news.ycombinator.com/item?id=49892202>)

### Tail latency and practical value

Commenter questions Jeeves' purpose if p90 latency is 17 seconds, saying one might as well use an LLM, and contrasts this with Jev's cheap and fast profile\.

Sources: [Comment 49892563](<https://news.ycombinator.com/item?id=49892563>)

### Cost comparison for spam detection

Commenter cautions against calling Jev dirt cheap, reporting that Luna was 20% cheaper due to prompt caching in their spam\-detection cases, though not as fast\.

Sources: [Comment 49892718](<https://news.ycombinator.com/item?id=49892718>)

### Embeddings plus logistic classifier for email

Commenter suggests embedding sender, recipients, subject, and body and training a logistic classifier, reporting 95% accuracy with 50\-100 examples, CPU training under one minute, inference under 20ms, and 93\-94% on Banking77\.

Sources: [Comment 49893777](<https://news.ycombinator.com/item?id=49893777>)

## Sources & coverage

AI-generated summary · 2026\-09\-29T13:40:18\.297702\+00:00

Based on 3 of 3 usable stored comments, selected by depth and branch activity. This is a sample of the discussion. Article text may also be shortened.

Generated using deepseek\-ai/DeepSeek\-V4\.1\-Flash. Check the linked sources for full context.