MDLC Thesis
Current point of view. Revised as evidence accumulates; every change is dated.
Draft v0.2. Claims 1 and 3 rest on hypotheses only; claims 2, 4 and 5 now have external evidence, mostly vendor documentation. Each claim links to the hypothesis or note it rests on.
Thesis statement
As agents take over building and validation, the model development lifecycle reorganizes around evidence and approval: humans set direction and accept evidence at gates; policy, knowledge and skills become versioned infrastructure; and the systems of record that hold evidence and approvals become the control points. That control plane is shared across classical ML, LLM and agentic systems, while how each is built and evaluated stays specialized.
Supporting claims
| # | Claim | Type | Rests on | Confidence |
|---|---|---|---|---|
| 1 | Evaluation evidence becomes the primary artifact. | hypothesis | H1 | low |
| 2 | Policy, knowledge and skills are versioned like code. | hypothesis | H2 · vendor profiles | low–medium |
| 3 | Workflow structure matters more than model choice for reliability. | hypothesis | H3, benchmarks | low |
| 4 | MDLC tooling converges in the control plane (governance, approval, evidence, registry), so far mainly inside each vendor’s ecosystem, with cross-vendor convergence limited to a few aggregation layers and open interfaces; build and evaluation tooling stays specialized by system type. | hypothesis | H4 · F5 · control-layer note · vendor matrix · audit-logs note | medium |
| 5 | Formal supervisory guidance treats generative and agentic AI separately in the near term, while organizations extend internal governance to cover them. | hypothesis (revised) | H5 · F1, F2, F4 | medium |
Open questions
What the thesis still cannot answer. Full wording, reasoning and evidence needed: all open questions.
Revision history
- v0.2 — monthly review after five new notes and six vendor profiles. Claim 4 raised to medium and reworded: control-plane convergence is so far inside each vendor’s ecosystem (three large clouds document an agent registry, identity and policy layer; the frontier labs document less), with cross-vendor convergence limited to a few aggregation layers and open interfaces; evaluation tooling is specialised (OpenAI is retiring its Evals platform; open-source tools fill the gap). Claim 2 raised to low–medium: vendor registries now carry versioned agents, policies and skills, which is not the same as keeping them in the code repository. Claims 1, 3 and 5 unchanged. All evidence is vendor documentation; no independent test or customer evidence. Q2 and Q6 moved to researching.
- v0.1 — added claims 4–5 from the convergence note; first claims backed by external sources (SR 26-2, MLflow 3, EU AI Act).
- v0 — initial framing from hypotheses only.