Skip to content

About.md

Latest commit

 

History

History
71 lines (49 loc) · 3.2 KB

About.md

File metadata and controls

71 lines (49 loc) · 3.2 KB

About — GCP V50 Flagship

What it is. A single protocol an advanced LLM can follow to produce auditable, benchmarked, policy‑gated invention packages with provenance and a verifiable Exit Wizard export. V50 builds on the V49 Flagship edition with autonomous spark generation, refined governance and compliance mapping, and an expanded library of runners and cartridges.

Core Principles (LLM‑native)

  1. Policy‑as‑Code gates (Rego/OPA‑style patterns).
  2. Observability from birth (traces/metrics/logs).
  3. Evidence over vibes: Evidence Index & Claim Graph.
  4. Adversarial & metamorphic testing.
  5. Formal constraints (SMT/proof assistants).
  6. Novelty vs SOTA with thresholds.
  7. Supply‑chain & provenance (SBOM/AI‑BOM, signatures).
  8. Rehydration proof (rebuild & re‑run).
  9. Modular orchestration (Runners & Cartridges).
  10. Export with guarantees (Exit Wizard).

What GCP Is / Is Not

  • Is: a deterministic, auditable protocol any major LLM can run to emit release‑grade artifacts.
  • Is: gate‑driven; can stop early at No‑Go points.
  • Is not: a repo crawler or production deployer.
  • Is not: a substitute for domain expertise, legal review or field validation.

Modes & Startup

GCP V50 supports two modes of operation:

  • Architect‑led (spark_provided): you supply a well‑defined problem spark; the protocol starts at I‑0 initialization and proceeds through the phases.
  • Collaborative brainstorm (autonomous): you supply only a high‑level domain; the I‑1 Autonomous Spark Generation phase proposes a spark based on the domain, performs a worth‑it analysis and presents a brief for acceptance.

Operational modes (auto, full, manual) control gate progression. Universal commands (proceed, branch, return, end & export, etc.) remain unchanged from V49.

Interfaces (stable in V50)

  • Runner API: deterministic modules with inputs, steps, artifacts, gates and telemetry as defined in the Master Runners Codex.
  • Cartridge API: domain knowledge modules with embeddings, indexes, lexicons, tools and governance notes.

Roles (internal council)

Planner • Researcher • Engineer • Adversary • Auditor • Operator • Scribe.

Risk‑Tiered Rigor

Tier Reps/Seeds Notes
R1 ≥15 / ≥5 Integrity‑lite
R2 ≥30 / ≥10 Full integrity; SBOM/signing
R3 ≥50 / ≥20 Includes FMEA/STPA/GSN/DPIA/HIL; CE default

Reproducibility & Evidence TTL: C7.Repro templates included. TTL: R1 = 180 days, R2 = 90 days, R3 = 30–60 days.

Artifact Model

Artifacts are stored in run directories or emitted inline. Key folders include:

agents/  memory/  tools/  novelty/  redteam/  metamorphic/  observability/
policies/  SLOs/  SBOM/  provenance/  Evidence_Log/  Rehydration_Test/
Audit_Package/  XW/  INDEX.md  MANIFEST.json  Gate_Signals.json

When file I/O is unavailable, emit artifacts inline using:

BEGIN ARTIFACT:<virtual_path>
<content>
END ARTIFACT

Rehydration tests rebuild the environment and re‑run checks; the Exit Wizard packages a signed, transparent export.