Skip to content
OpenJevX wiki
Every statement on this page is cited and was checked against jev-cloud origin/main (10a744c), jevx HEAD (4a4722d), openjevx v0.5.9 (ee2a1f4) on 2026-10-04.

Glossary

Terms are grouped: the decision API first, then the model, then training, then serving and jevx.

Decisions

Section titled “Decisions”
Term Meaning
System One A decision model that answers in well under a second with a number or a key, never prose. 1
/v1/systemone The OpenJevX decision endpoint, served by the same process as /health, behind an API-key check. 2 3
noul A yes/no question; its two keys are false and true. 4
choice A question that picks one option key; the answer carries the winning key as choice. 5 6 7
score A question whose options are levels, lowest first, such as low|medium|high. 8
state The facts the decision depends on, as a JSON object or plain text. 9
question: type, instructions, criteria A training row’s questions are {id:{type: noul|choice|score, instructions, criteria}}. 10
probabilities One probability per option key, for example {"api":0.0073,"billing":0.0066,"docs":0.0073,"web":0.9788} for a choice. 11
confidence / answer_confidence For noul, max(p, 1-p); in the routing recipe the pick web has probability 0.9788 and confidence 0.9788. 12 11
act_probability A field under action in each answer, for example "action":{"act_probability":1}. 11

The model

Section titled “The model”
Term Meaning
Laya The base model, convaiinnovations/laya, built by Nandakishor Mukkunnoth, ConvAI Innovations. 13
ModernBERT The encoder, answerdotai/ModernBERT-large (395M), plus Laya’s decision head (421M total). 14
RLCD Reinforcement Learning for Calibrated Decisions, the training method. 15
model folder One folder holding openjevx.w8.onnx, config.json and tokenizer.json; the server loads it and uses that model’s own temperatures. 16
openjevx.w8.onnx The graph, 8-bit weight-only, at most 750 MB (model.max_w8_mb). 17 18
config.json (model) The model’s version, calibration temperatures (choice, score, noul), max_len, special ids and sha256 of the graph. 19 20
tokenizer.json Optional in a model folder; the built-in ModernBERT tokenizer is used otherwise. 21
QDQ weight-only / MatMulNBits The graph is QDQ weight-only: int8 DequantizeLinear feed fp32 MatMul, and ORT fuses them into MatMulNBits (8-bit, accuracy level 4 = int8 activations). 22
calibration temperature temperature in config.json is that model’s own confidence calibration, so each model carries its own. 23 24
max_len The token length the model was trained at; model 0.5.2’s config.json has 512, so longer inputs are cut at 512 tokens. 25 26
head_max The token budget shared by the question head and its options; 256 in the model 0.5.2 folder. 27 28 25
marker The input position of each option; each option’s token ids start with p.MASK, and markers[i] records where option i begins. 29 30
[MASK] in text Any literal [MASK] in the state, instructions or options is replaced by a space. 31 32 33

Training and the gate

Section titled “Training and the gate”
Term Meaning
gold The target in a training row: full probability distributions, not hard labels. 34 35
gate The test a model must pass to ship: it is served locally and scored, and ft.py gate exits 1 if it misses the config thresholds. 36 37
basics gate gate/basics_gate.jsonl and gate/conditions_basics_gate.jsonl, the basics_files the thresholds apply to. 38
jevx 13 fundamentals 13 checks run through jevx; a model needs at least 12 of 13. 39
right & confident Right and sure enough to act on: yes ≥ 0.8, no ≤ 0.2, a choice or score ≥ 0.6. 40
confidently wrong Sure and wrong; the dangerous one. 41
usable Right AND past jevx’s default thresholds (yes ≥ 0.8, no ≤ 0.2, choice/score ≥ 0.6). 42
leakage check Removes any training question that also appears in a test or gate file, so the gate measures unseen cases. 43
shard Built by ft.py package in <data>/work/shards/<version>[-smoke]/; ft.py train runs the shard on the GPU provider. 44
smoke run The run ft.py all makes before the full run: a ~10-minute smoke job on the GPU provider. 45 46
--repeat How many times your train file is repeated in the training mix; default 3. 47
trainable checkpoint The training checkpoint that is kept so the model can be fine-tuned again. 48
fine-tune kit The v0.5.2 trainable checkpoint (777 MB): model.safetensors, encoder/, tokenizer/, rl_agent_config.json, and no training data. 49
DESTROY-AFTER label The self-destroy function destroys only openjevx-*-DESTROY-AFTER boxes when given the kill token. 50
private / public R2 bucket openjevx-train holds shards, runs, checkpoints and gate reports and stays private; only released models go to the public openjevx bucket. 51
jev-train Trains Jev on your own decisions: it validates your CSV, uploads it to S3 in your AWS account, starts a SageMaker training job and watches it. 52

Serving and jevx

Section titled “Serving and jevx”
Term Meaning
models/current/ The S3 prefix the server loads (model = s3://<bucket>/models/current/). 53
promote / rollback Promote copies the 3 files of models/<version>/ into models/current/; rollback promotes the previous version again. 54
ETag reload On a new ETag the server downloads into a new cache folder, verifies it, swaps it in between requests, and keeps the previous folder for rollback. 55
model_sha256 pin The sha256 of the .tar.gz or the folder’s openjevx.w8.onnx; a mismatch fails loudly. 56 57
fallback model The folder served while the bucket holds no model yet; by default the model/ lookup next to the executable. 58
Server-Timing A header on every decision response with encode, wait, run and total in ms. 59
usage.server_ms The Server-Timing total, repeated as usage.server_ms in the JSON body. 59
openjevx.api-key / openjevx.password Files the server generates (mode 0600) when the API key or dashboard password is needed but unset: beside openjevx.json, else in $OPENJEVX_DATA, else the working folder; reused on later starts. 60
jevx profile A named System One endpoint in jevx, added with jevx profile add NAME URL --model MODEL. 61
jevx cache Answers are cached on your machine, a hash-keyed file each; --fresh asks again. 62

See also Fine-tune and Troubleshooting.

Sources

Section titled “Sources”

Footnotes

Section titled “Footnotes”

  1. jevx @ 4a4722d (4a4722d) · README.md L14–16 ↩

  2. openjevx @ v0.5.9 (ee2a1f4) · cmd/openjevx/main.go L190 ↩

  3. openjevx @ v0.5.9 (ee2a1f4) · cmd/openjevx/main.go L206–208 ↩

  4. openjevx @ v0.5.9 (ee2a1f4) · internal/decide/decide.go L98–100 ↩

  5. openjevx @ v0.5.9 (ee2a1f4) · recipes/README.md L37 ↩

  6. openjevx @ v0.5.9 (ee2a1f4) · recipes/README.md L40 ↩

  7. openjevx @ v0.5.9 (ee2a1f4) · internal/decide/decide.go L236–237 ↩

  8. openjevx @ v0.5.9 (ee2a1f4) · finetuning/dataprep/import_csv.py L11 ↩

  9. openjevx @ v0.5.9 (ee2a1f4) · finetuning/dataprep/import_csv.py L7 ↩

  10. openjevx @ v0.5.9 (ee2a1f4) · README.md L250 ↩

  11. openjevx @ v0.5.9 (ee2a1f4) · recipes/02-route-a-ticket-queue.md L28 ↩ ↩2 ↩3

  12. openjevx @ v0.5.9 (ee2a1f4) · internal/decide/decide.go L244–246 ↩

  13. openjevx @ v0.5.9 (ee2a1f4) · docs/adr/0001-base-model.md L12–16 ↩

  14. openjevx @ v0.5.9 (ee2a1f4) · docs/adr/0001-base-model.md L14 ↩

  15. openjevx @ v0.5.9 (ee2a1f4) · docs/adr/0001-base-model.md L15 ↩

  16. openjevx @ v0.5.9 (ee2a1f4) · docs/RELEASE_PROCESS.md L60–68 ↩

  17. openjevx @ v0.5.9 (ee2a1f4) · AGENTS.md L15 ↩

  18. openjevx @ v0.5.9 (ee2a1f4) · docs/RELEASE_PROCESS.md L62 ↩

  19. openjevx @ v0.5.9 (ee2a1f4) · docs/RELEASE_PROCESS.md L63 ↩

  20. openjevx @ v0.5.9 (ee2a1f4) · README.md L69–76 ↩

  21. openjevx @ v0.5.9 (ee2a1f4) · README.md L76 ↩

  22. openjevx @ v0.5.9 (ee2a1f4) · llmresults/14-x86-cpu-latency.md L15–17 ↩

  23. openjevx @ v0.5.9 (ee2a1f4) · README.md L70–71 ↩

  24. openjevx @ v0.5.9 (ee2a1f4) · README.md L79 ↩

  25. openjevx @ v0.5.9 (ee2a1f4) · README.md L70–72 ↩ ↩2

  26. openjevx @ v0.5.9 (ee2a1f4) · llmresults/13-v0.5.2-gate-misses.md L120–121 ↩

  27. openjevx @ v0.5.9 (ee2a1f4) · internal/decide/decide.go L153 ↩

  28. openjevx @ v0.5.9 (ee2a1f4) · internal/decide/decide.go L167–168 ↩

  29. openjevx @ v0.5.9 (ee2a1f4) · internal/decide/decide.go L147 ↩

  30. openjevx @ v0.5.9 (ee2a1f4) · internal/decide/decide.go L172–176 ↩

  31. openjevx @ v0.5.9 (ee2a1f4) · internal/decide/decide.go L127 ↩

  32. openjevx @ v0.5.9 (ee2a1f4) · internal/decide/decide.go L138 ↩

  33. openjevx @ v0.5.9 (ee2a1f4) · internal/decide/decide.go L142 ↩

  34. openjevx @ v0.5.9 (ee2a1f4) · docs/adr/0002-training-run.md L14 ↩

  35. openjevx @ v0.5.9 (ee2a1f4) · docs/adr/0001-base-model.md L22 ↩

  36. openjevx @ v0.5.9 (ee2a1f4) · docs/RELEASE_PROCESS.md L71 ↩

  37. openjevx @ v0.5.9 (ee2a1f4) · finetuning/ft.py L14–15 ↩

  38. openjevx @ v0.5.9 (ee2a1f4) · finetuning/config.example.json L136–142 ↩

  39. openjevx @ v0.5.9 (ee2a1f4) · docs/RELEASE_PROCESS.md L77–78 ↩

  40. openjevx @ v0.5.9 (ee2a1f4) · docs/book/train-your-own-jev.md L217 ↩

  41. openjevx @ v0.5.9 (ee2a1f4) · docs/book/train-your-own-jev.md L219 ↩

  42. openjevx @ v0.5.9 (ee2a1f4) · llmresults/10-v041-baseline.md L4–5 ↩

  43. openjevx @ v0.5.9 (ee2a1f4) · docs/book/train-your-own-jev.md L155–156 ↩

  44. openjevx @ v0.5.9 (ee2a1f4) · finetuning/ft.py L11–13 ↩

  45. openjevx @ v0.5.9 (ee2a1f4) · finetuning/ft.py L16 ↩

  46. openjevx @ v0.5.9 (ee2a1f4) · finetuning/Taskfile.yml L19–21 ↩

  47. openjevx @ v0.5.9 (ee2a1f4) · finetuning/dataprep/import_csv.py L249 ↩

  48. openjevx @ v0.5.9 (ee2a1f4) · AGENTS.md L18–19 ↩

  49. openjevx @ v0.5.9 (ee2a1f4) · README.md L248 ↩

  50. openjevx @ v0.5.9 (ee2a1f4) · docs/adr/0009-one-job-gpu-run-via-r2.md L36–37 ↩

  51. openjevx @ v0.5.9 (ee2a1f4) · docs/adr/0009-one-job-gpu-run-via-r2.md L67–68 ↩

  52. jev-cloud @ origin/main (10a744c) · client/README.md L3–9 ↩

  53. jev-cloud @ origin/main (10a744c) · docs/model-s3-contract.md L8 ↩

  54. jev-cloud @ origin/main (10a744c) · docs/model-s3-contract.md L20–21 ↩

  55. openjevx @ v0.5.9 (ee2a1f4) · README.md L113–114 ↩

  56. openjevx @ v0.5.9 (ee2a1f4) · README.md L97 ↩

  57. openjevx @ v0.5.9 (ee2a1f4) · README.md L108 ↩

  58. openjevx @ v0.5.9 (ee2a1f4) · README.md L100 ↩

  59. openjevx @ v0.5.9 (ee2a1f4) · README.md L59–61 ↩ ↩2

  60. openjevx @ v0.5.9 (ee2a1f4) · cmd/openjevx/auth.go L25–28 ↩

  61. jevx @ 4a4722d (4a4722d) · README.md L63–66 ↩

  62. jevx @ 4a4722d (4a4722d) · README.md L29–30 ↩