MESA E2E Certification · v0.1.0
Agent Pack Start
This guide has equivalent English and Turkish versions. The source sections behind this page are at the bottom.
Scope of this document#
Profile B evaluates a frozen Data → Core transaction with independent receipts and fail-closed gates. Official execution and metric producers exist. External B4/B6/B7/B8 evidence and a fresh official PASS remain absent; historical results are not current certification.
The original reference below preserves this page’s complete contract, commands or dated evidence in its source language. Test outcomes apply only within their stated date and conditions.
Source reference for this page
The sections of the upstream file that back this page, unchanged. Source section links open the appendix automatically.
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.
This folder is the complete operating contract for the autonomous Profile B agent. The intended operator is Antigravity CLI (or another shell-capable coding agent) running inside the certification VM.
Mission in one sentence#
Prove, with reproducible evidence, that approved legal data produced by MESA_Data can be published through the native MESA_Data→MESA product path into the current MESA V4 runtime, retrieved correctly with real NVIDIA embeddings/LLM extraction, isolated by tenant, persistent/idempotent across restart, and used to produce grounded legal answers.
Profile B is not a unit-test exercise and it is not a README compliance exercise. It is an evidence-producing black/gray-box integration certification.
Read order — mandatory#
Before running commands, read every numbered contract file in this exact order:
01_MASTER_AGENT_PROMPT.md02_SYSTEM_MODEL_AND_NON_NEGOTIABLES.md03_SOURCE_ARCHIVE_AUDIT_BASELINE.md04_WORKSPACE_BRANCH_GIT_POLICY.md05_ENVIRONMENT_SECRETS_DEPENDENCIES.md06_PROFILE_B_OBJECTIVE_GATES_VERDICTS.md07_MESA_RUNTIME_PROVIDER_PREFLIGHT.md08_MESA_DATA_QUALITY_REVIEW_RUNBOOK.md09_NATIVE_MESA_DATA_TO_MESA_CONTRACT.md10_CORPUS_DESIGN.md11_HUMAN_H1_APPROVAL.md12_GROUND_TRUTH_FREEZE.md13_BENCHMARK_CONFIG_FREEZE.md14_EXECUTION_PHASES.md15_RETRIEVAL_SCORING.md16_GRAPH_PROOF_AND_ABLATION.md17_FINAL_ANSWER_GENERATION_AND_GRADING.md18_ISOLATION_IDEMPOTENCY_RESTART_PROVENANCE.md19_FAILURE_TRIAGE_REPAIR_RESTART.md20_RESOURCE_AND_PROVIDER_BUDGETS.md21_EVIDENCE_MANIFEST_AND_INTEGRITY.md22_LIVE_STATUS_TEMPLATE.md23_FINAL_REPORT_TEMPLATE.md24_QUICK_COMMAND_DISCOVERY_MAP.md25_POST_FIX_PRELAUNCH_CHECKLIST.md26_EXPECTED_OUTPUTS_AND_ASSERTIONS.md27_FINAL_AGENT_SELF_AUDIT.md28_AUTONOMOUS_DECISION_DOCTRINE.md29_WORKSPACE_HYGIENE_AND_ISOLATION.md30_VM_BOOTSTRAP_AND_CANONICAL_WORKSPACE_LAYOUT.md31_BENCHMARK_INTEGRITY_FREEZE_DETERMINISM_AND_FINALIZATION.md32_PROFILE_B_CANONICAL_MESA_RUNTIME_LOCK.md33_CERTIFICATION_HARNESS_AND_GROUND_TRUTH_VERIFICATION_CONTRACT.md34_8GB_VM_MEMORY_PRESSURE_AND_OOM_GUARD.md35_AGENT_TOKEN_AND_FRESH_VM_EXECUTION_EFFICIENCY.mdMANIFEST.md
Do not start execution after reading only the master prompt.
Do not skip later-numbered contracts because an earlier file already mentions
their topic.
Working directories#
The certification repository root is authoritative and must be discovered, never hard-coded:
CERT_ROOT="$(git rev-parse --show-toplevel)"Expected logical layout:
<CERT_ROOT>/
├── agent-pack/
├── config/
├── datasets/
├── ground-truth/
├── harness/
├── evidence/
├── runs/
├── reports/
└── scripts/MESA and MESA_Data are separate Git repositories under test. Discover their actual paths before use. The VM may commonly place them under ~/mesa-cert/repos/, but that is a convenience, not a contract.
Secrets must remain outside the certification Git repository. A common protected location is ~/mesa-cert/secrets/nvidia.env.
For every run use repository-relative certification storage:
<CERT_ROOT>/runs/<RUN_ID>/
<CERT_ROOT>/evidence/<RUN_ID>/
<CERT_ROOT>/datasets/runtime/<RUN_ID>/
<CERT_ROOT>/reports/generated/<RUN_ID>/Runtime/generated content is intentionally not product source code. Do not silently relocate it; record any necessary alternate path in the run manifest.
Dedicated branch#
Use the same branch name in all three repositories (certification repository, MESA, and MESA_Data):
cert/profile-b-e2eNever commit Profile B fixes directly to main. See 04_WORKSPACE_BRANCH_GIT_POLICY.md.
One human gate, by design#
Automation may collect, parse, validate, build a release, create benchmark ground truth, publish, ingest, search, score, repair code, and restart runs. The human is required only for the hash-bound H1 final corpus/release review and permission to deliver the approved release to MESA.
The agent must never forge H1 or reinterpret silence as approval.
Non-pass outcomes are legitimate#
The agent must prefer a truthful failure over a false pass. Allowed final verdicts are defined in 06_PROFILE_B_OBJECTIVE_GATES_VERDICTS.md. In particular, a diagnostic bridge that bypasses MESA_Data's native publisher can help debug but cannot produce a Profile B pass.
Critical rule after any code change#
Any product, harness, dependency, Compose, benchmark logic, or scoring code change after a run begins invalidates the current certification run. Preserve its evidence, mark it invalidated, create a new RUN_ID, use clean MESA storage and a clean MESA_Data run root, and start again from Phase 0.
Do not delete failed evidence to make the final directory look clean.
<!-- V3.1_FRESH_MAIN_LOCK -->
Fresh-main lock before autonomous work#
Before creating or reusing cert/profile-b-e2e, synchronize all three repositories with their remote main using the clean-baseline procedure in 04_WORKSPACE_BRANCH_GIT_POLICY.md. Record the resulting origin/main SHAs.
Once a RUN_ID is created, do not pull, merge, rebase, or otherwise absorb newer remote commits into that active run. A desired upstream update requires preserving/invalidation of the current run, resynchronizing from main, creating a new coherent baseline, and starting with a new RUN_ID.
Phase -1 workspace hygiene#
Before Phase 0, the agent MUST execute
29_WORKSPACE_HYGIENE_AND_ISOLATION.md.
The agent must inventory historical VM state, isolate a fresh RUN_ID runtime,
prove that old MESA/MESA_Data state is not reused, inventory Docker state, and
write runs/<RUN_ID>/workspace-baseline.json.
Unknown or user-owned files must not be deleted.
VM bootstrap comes first#
On a fresh/restored VM, the first mandatory document is:
30_VM_BOOTSTRAP_AND_CANONICAL_WORKSPACE_LAYOUT.md
It establishes the canonical $HOME/mesa-cert layout, authoritative repository
paths, safe normalization behavior, runtime namespace rules and bootstrap
evidence.
Then execute 29_WORKSPACE_HYGIENE_AND_ISOLATION.md, followed by Phase 0.
If an existing workspace does not match the canonical layout, do not destroy
it. Prefer a fresh canonical checkout/runtime and preserve historical/user data.
Benchmark integrity lock#
Before official TEST execution, the agent MUST execute the integrity/freeze
requirements in:
31_BENCHMARK_INTEGRITY_FREEZE_DETERMINISM_AND_FINALIZATION.md
Ground truth is a scoring oracle only and must never leak into MESA, retrieval,
extraction or final-answer requests.
Canonical MESA runtime lock#
Before provider/runtime preflight and before official TEST, execute:
32_PROFILE_B_CANONICAL_MESA_RUNTIME_LOCK.md
The official Profile B runtime is not agent-selected:
MESA_MODEL_ENABLED=true
MESA_TIER3_MODE=0with the canonical NVIDIA Nemotron/GPT-OSS identities defined in document 32.
Initial full read, then phase-scoped reads#
At session start, read the complete mandatory numbered contract sequence,
including documents 29-35.
After this initial read, follow document 35:
- use
agent-pack/SHA256SUMS.txtto detect contract changes, - use
agent-checkpoint.jsonfor context recovery, - reread only the contracts relevant to the active phase unless checksums or
failure context require more,
- never skip a later-numbered contract merely to save tokens.
This reduces repeated context/token use without weakening the certification.
