MESA E2E Sertifikasyonu · v0.1.0

Sistem Modeli ve Esnetilemezler

v0.1.0 kaynak: agent-pack/02_SYSTEM_MODEL_AND_NON_NEGOTIABLES.md

Bu rehberin Türkçe ve İngilizce sürümleri aynı konuyu açıklar. Bu sayfayla ilgili kaynak bölümleri en altta yer alır.

Bu belgenin kapsamı#

Profile B, veri ve ayarları sabitlenmiş Data → Core akışını bağımsız işlem alındılarıyla değerlendirir; doğrulanamayan koşullarda ilerlemez. Resmî çalıştırma ve ölçüm bileşenleri hazır, ancak B4/B6/B7/B8 için dış kanıtlar ve yeni bir resmî PASS sonucu henüz yoktur. Eski sonuçlar güncel sertifikasyon sayılmaz.

Aşağıda özgün belgeyi, komutları ve varsa test kayıtlarını kendi dilinde okuyabilirsiniz. Sonuçlar yalnızca belgede belirtilen tarih ve koşullar için geçerlidir.

Kaynak belgedeki ilgili bölümler

Bu bölüm, kaynak belgenin bu sayfayla ilgili kısımlarını değiştirmeden gösterir. Kaynak içindeki bir bölüme giden bağlantıyı açtığınızda metin otomatik olarak görünür.

Profile B Autonomous Agent Pack v3.1
Scope: MESA + MESA_Data native legal E2E certification on a dedicated certification branch.
Authority rule: the live checked-out repositories and runtime behavior outrank this document when code has changed. Never guess an API, CLI, schema, path, or capability: discover and record it first.

Product boundary#

Profile B evaluates the integrated chain:

text
Official Turkish legal source
  → MESA_Data raw immutable artifact
  → decode / parse / canonicalize
  → quality + review
  → immutable verified release
  → deterministic SourceChunk plan
  → native MESA_Data publisher
  → MESA V4 authenticated/session-scoped ingestion
  → async mutation COMMITTED
  → canonical SQL + LanceDB + FTS/assertions + Kùzu projection
  → public MESA V4 search
  → retrieval provenance
  → context-only GPT-OSS answer
  → deterministic scoring

A pass must prove the chain. Testing components independently is useful for diagnosis but cannot substitute for the chain.

What Profile B is for#

Profile B answers: “Can the current product ingest and retrieve a meaningful, approved real legal corpus with the intended external models and preserve enough provenance/isolation to support grounded answers?”

It does not attempt to certify long-duration memory-leak behavior (Profile A) or exhaustive temporal/contradiction/crash/adversarial semantics (Profile C), although B includes bounded restart, idempotency and graph checks because they are necessary to trust the benchmark.

Invariants#

The following are non-negotiable:

  • Native publisher: Full PASS requires MESA_Data's product publisher to interoperate with MESA. Diagnostic harness bridges do not satisfy this.
  • Real providers: Do not replace NVIDIA models with mocks/local substitutes for hard provider gates.
  • Asymmetric embedding: documents use Nemotron input_type=passage; queries use input_type=query.
  • Embedding identity: expected model nvidia/nemotron-3-embed-1b, dimension 2048, configured version nemotron-qpass-v1 unless the current product contract deliberately changes and is requalified.
  • Extraction: expected model openai/gpt-oss-20b, Turkish extraction, real FactExtractionService/current production equivalent, with sufficient bounded output budget (archive baseline: 4096).
  • Immutable evidence: failed runs remain available.
  • One coherent final run: do not cherry-pick passing phases from different code SHAs/runs.
  • No TEST tuning: TEST is evaluation, not development.
  • No source falsification: never edit official legal text, canonical expected evidence, or qrels merely to fit MESA output.
  • No secret exposure: redact credentials and bearer headers.
  • Authorization: tenant/dataset/session boundaries are hard gates, not diagnostics.
  • Graph proof: code presence or Kùzu writes alone do not prove graph retrieval usefulness.
  • Grounding: final answer correctness without evidence is a fail for grounded-answer scoring.

Evidence hierarchy#

Strongest to weakest:

  1. current-run live HTTP/runtime evidence with exact code/config hashes;
  2. deterministic state inspection and immutable artifacts;
  3. integration tests against real local services;
  4. unit/regression tests;
  5. static source inspection;
  6. documentation/README assertions.

Higher evidence should be used where a hard gate permits it.

Truth over automation#

The agent may stop with a failure. It must not “help” by silently substituting:

  • a different model;
  • a different endpoint;
  • a fabricated official source;
  • a smaller/easier query set after results are seen;
  • relaxed thresholds;
  • a mock server;
  • a manual direct MESA client bridge in place of the native publisher;
  • a hidden second attempt whose first failure is deleted.