chore(release): v1.16.4 — KB-refresh MEDIUM-tier (33 filer): prompt-caching/Realtime/Guardrails/PTU/Foundry-deprecations + 18 TRYGT (medium 33→0)

CHANGELOG.md

@@ -5,6 +5,21 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [1.16.4] - 2026-06-24
+
+KB-refresh-patch. MEDIUM-tier (33 reference docs) verifisert mot Microsoft Learn (post-1.16.3 baseline). Ingen kommando-, agent- eller skill-endringer — kun reference docs. 18 filer var TRYGT (innhold gjeldende, kun header-bump); 15 filer hadde reell drift. Hver `VERIFY`-påstand fra de 7 parallelle drift-rapportene ble re-bekreftet av main-agent mot sitert Microsoft Learn-kilde FØR skriving; disclaimed/illustrative priser urørt; GPT-4.1-pricing-påstand forble uverifiserbar (kun community-kilder) og ble IKKE endret; alle endringer review-gated på `git diff` + plugin-validering (239/0). Medium-count 33 → 0 (needing-update 85 → 52).
+
+### Changed
+
+- **Prompt-caching:** `token-optimization` — retention-modell omskrevet: in-memory (5–10 min, ≤1t, default t.o.m. gpt-5.4) vs **extended retention 24t** (`prompt_cache_retention`, default for nyere modeller); in-memory støttes av «GPT-4o eller nyere».
+- **Realtime:** `real-time-reasoning` — la til `gpt-realtime-1.5` (2026-02-23); Realtime API flyttet til **GA-endepunkt** `/openai/v1` (ikke lenger date-baserte api-version); preview→GA-statusnyanse.
+- **Guardrails-rebranding:** `regulatory-and-compliance` + `azure-ai-foundry` — «content filters» → **Guardrails (and controls)**, default `Microsoft.DefaultV2`, fire intervensjonspunkter; data-privacy-kildetittel → «Models sold by Azure».
+- **GPU/PTU:** `gpu-compute-sizing` — `ND96asr_v4` = 8×A100 **40GB = 320 GB** (var 640 GB); `load-testing` — PTU min deployment Global/Data Zone **15** (increment 5), 50 kun Regional.
+- **Foundry-deprecations:** `azure-ai-foundry` — **Deep Research-tool** og **Connected agents (classic)** deprecated (retire 2027-03-31), **Prompt Flow**-retirement 2027-04-20, **A2A-tool nå i Java SDK**.
+- **Agent-evaluering/-memory:** `agent-evaluation` — nye agent-evaluatorer (Task Completion/Customer Satisfaction/Tool Selection m.fl.) + `FoundryEvals` + `gpt-4.1-mini` judge; `agent-memory` — **procedural memory** (3. type), TTL/CRUD, **Norway East**-region bekreftet (VNet ikke støttet).
+- **Øvrig faktafiks:** `multi-turn` RPM/TPM-forhold varierer per modell; `structured-output` usupporterte keywords utvidet + `gpt-4o-mini-audio-preview`; `temperature-sampling` Copilot Studio prompt builder-temp + fjernet uoffisielt M365-tall; `rag-hallucination` groundedness `mitigating`/`correctionText`; `model-versioning` MLflow stages legacy i UC; `speaker-recognition` **Limited Access**-status; `chain-of-thought` reasoning_effort per-modell-nyanse.
+- **Header-bump (TRYGT, 18 filer):** domain-specific, few-shot, function-calling, grounding, role-playing, system-message-design, async-processing, batch-api-usage, concurrent-request, performance-benchmarking, response-chunking, throughput-optimization, token-per-second, late-chunking, rag-evaluation, self-reflective-rag, streaming-rag, feedback-loops.
+
 ## [1.16.3] - 2026-06-24
 
 KB-refresh-patch. HIGH-tier (11 reference docs) + 1 dateless kjernefil (`ms-ai-advisor/references/architecture/security.md`), flagget av endrings-detektoren mot Microsoft Learn-sitemaps (post-1.16.2 baseline), er verifisert og oppdatert. Ingen kommando-, agent- eller skill-endringer — kun reference docs. Hver `VERIFY`-påstand fra de parallelle drift-rapportene ble bekreftet mot sitert Microsoft Learn-kilde FØR skriving (main-agent som verifiseringsgate); disclaimed/illustrative priser urørt; alle endringer review-gated på `git diff` + plugin-validering (239/0). High-count 11 → 0, Unverified 1 → 0 (needing-update 97 → 85).