docs/research-jev-and-laya-decision-models #41

Merged

clawdie merged 5 commits from docs/research-jev-and-laya-decision-models into main

2026-10-08 12:29:58 +02:00

clawdie commented

2026-10-08 12:29:41 +02:00

Owner

No description provided.

clawdie added 5 commits

2026-10-08 12:29:42 +02:00

docs(research): add Cole Medin and Under the Hood AI decision models guide 4a7f4e3dce

- Part 1: AI coding workflows (lifecycle hook guardrails, real-time control, web navigation, task triage).
- Part 2: AI second brain (inbox triage, prompt injection defense, weighted priority scoring, deduplication).
- Part 3: Jev vs Laya (ModernBERT open-source local decision model alternative, 40ms local latency, sovereign airgap fallback under 20 options).
- Complete structured timestamped transcripts for all three videos.

docs(research): expand decision models guide with James Briggs GLiNER analysis d12129b762

- Define Jev under the hood as a non-autoregressive bidirectional transformer
  encoder (ModernBERT/GLiNER lineage) with multi-task classification heads.
- Explain probability calibration as Jev's key neurosymbolic differentiator
  over uncalibrated LLM logprobs.
- Document open-source alternatives (GLiNER, GLiClass, GlideGuard < 300M params,
  sub-50ms CPU latency).
- Add Part 4 executive summary and full structured transcript of James Briggs
  analysis.

docs(research): align Jev interface definition and attribute claims cc884b2b53

- Define Jev contractually by its typed decision interface (choice, score, boolean).
- Clarify that internal architecture is undisclosed by TypeSafe; attribute
  non-autoregressive encoder hypotheses and calibration claims to external
  reviewers (James Briggs, Archer Hume, Under the Hood AI).
- Keep Laya labeled as an optional open-source alternative to explore, not
  a Jev provider or established fallback tier in Colibri.
- Maintain invariant that passing Jev never authorizes execution; deterministic
  policy is the mandatory authorizer for all tool execution.
- docs only

docs: clarify Jev route authority and boundaries 309868ff5d

docs(research): add Part 5 Laya fine-tuning and temperature calibration analysis

go-ci / linux (build, vet, test) (pull_request) Successful in 2m3s

Details

verify-evidence / verify evidence reports (pull_request) Successful in 3s

Details

963eacb1bc

- Document empirical head-to-head evaluation by The AI Automators (Laya vs. Jev).
- Detail 4-bucket dataset split methodology (train, val, calibrate, test).
- Detail post-hoc temperature scaling (T=1.63) resolving open-source calibration
  deficits, achieving 36/1,000 errors at 90% gate vs. Jev's 39/1,000 errors.
- Document basic state performance (fine-tuned Laya at 82.2% vs. Jev at 67.3%)
  and rich state performance (86.1% vs. 87.4% on 7k tokens).
- Add full structured transcript for Part 5.
- docs only

clawdie merged commit ae70ffab48 into main

2026-10-08 12:29:58 +02:00

clawdie deleted branch docs/research-jev-and-laya-decision-models

2026-10-08 12:29:58 +02:00

clawdie referenced this pull request from a commit

2026-10-08 12:30:00 +02:00

Merge pull request 'docs/research-jev-and-laya-decision-models' (#41) from docs/research-jev-and-laya-decision-models into main