Definition
Conviction without dependence means choosing the best available instrument deliberately while owning the system that gives the instrument context, authority, memory, and proof.
Use the best instrument. Own the system.
This is neither vendor neutrality nor vendor capture. A team should be able to say which model and agent environment it prefers, why it prefers them, and what evidence supports the choice. The same team should be able to change that choice without losing its workflow, policy, data, or institutional learning.
The technical shorthand is:
Model-opinionated in practice. Model-portable by design.
The Language System
Use each expression for its own job:
| Layer | Canonical language | Job |
|---|---|---|
| Doctrine | Conviction without dependence | Names the enduring relationship between an instrument and the owned system around it. |
| Maxim | Use the best instrument. Own the system. | Gives operators a compact decision rule. |
| Technical | Model-opinionated in practice. Model-portable by design. | States that current choices are deliberate while interfaces, policy, and evidence remain portable. |
| Commercial | Built primarily with OpenAI Codex. Designed to outlast any model. | Names the current .agency operating conviction without turning a vendor into the offer. |
| Ownership | Model capability is rented. Organizational intelligence and control are owned. | Identifies where durable value must remain. |
The language is intentionally asymmetric. The Canon names a durable principle. A commercial or technical surface may name the current platform. A current platform choice can change; the doctrine should not.
What the System Owns
A model can reason, generate, classify, and call tools. Those capabilities are important, but they do not constitute the operating system around delegated work. Durable ownership sits in the layers a business can inspect, version, move, and improve:
| Owned layer | What remains portable |
|---|---|
| Substrate data and review | Source records, schemas, provenance, identity, workflow state, human review, and organizational context. |
| MCP boundaries | Tool and resource contracts, authentication scopes, errors, and allowed actions. |
| Harness | Context assembly, tool exposure, iteration loops, budgets, and recovery behavior. |
| Skills and prompts | Reusable behavior, domain instructions, examples, and task framing. |
| Policy | Approval rules, escalation, stop conditions, entitlements, and human authority. |
| Evals | Golden tasks, failure cases, comparison criteria, and promotion thresholds. |
| Receipts | Evidence, decisions, traces, outcomes, rollback notes, and audit history. |
| Routing | The ability to choose hosted, open-weight, or custom models by task and risk. |
The model participates in this system. It does not own the system.
Current Expression
CREATE SOMETHING currently builds primarily with OpenAI Codex. This is a real operating conviction: Codex is the primary environment for setup, demonstration, repository work, and agent-operable delivery. OpenAI can also provide reasoning and agent hosting when a workflow benefits from it.
Substrate is the owned database and operator layer. It keeps source records, Atlas bindings, workflow actions, approvals, runs, receipts, and API/MCP access inside the system CREATE SOMETHING can inspect and evolve. OpenAI, Dify, and Cloudflare are the active external stack: OpenAI is the primary reasoning and agent environment, Dify is the visible agent application surface, and Cloudflare is the runtime. Each external platform has a clear job; none owns the workflow or its institutional memory.
That statement is not an official OpenAI partnership, certification, affiliation, reseller, or endorsement claim. It describes how CREATE SOMETHING works today.
The surrounding artifacts remain portable across Claude, Pi, compatible agent harnesses, open-weight executors, and custom models. MCP contracts, policy artifacts, evals, receipts, and source data should survive a change in model or runtime. Portability is proven by the ability to route, compare, fall back, and recover—not by claiming indifference to model quality.
Rules
1. Name the instrument honestly
Do not hide a strong technical preference behind generic “best model for the job” language. State the current default and why it earns that position.
2. Keep authority outside the model
The model may propose or execute. Policy decides what can run, what waits for a named person, and what must stop. A vendor’s system prompt is not a substitute for an owned approval and escalation policy.
3. Prefer portable contracts
Keep tools, resources, identity boundaries, and evidence accessible through documented interfaces. MCP, HTTP, OAuth, JSON, versioned schemas, and exported artifacts are materials that can outlast one model release.
4. Evaluate before routing
Model routing is an evidence decision. Use the same golden tasks, forbidden actions, latency measures, cost measures, trace requirements, and recovery checks when comparing frontier, open-weight, and custom models.
5. Preserve a fallback
Every production model choice needs a pinned version or named release, a known fallback, and a rollback path. Portability without a rehearsed recovery path is only an architectural claim.
6. Keep the proof
When the model or vendor changes, preserve the inputs, policy version, tool calls, decision, outcome, and receipt. The evidence belongs to the organization using the system.
Anti-Patterns
- Model reseller — Selling access to a generally available model as though the access itself were proprietary value.
- Badge-first marketing — Using vendor marks or lab names as a substitute for workflow boundaries, policy, evals, and proof.
- False agnosticism — Claiming every model is interchangeable while hiding real quality, cost, tool-use, safety, or latency differences.
- Shadow coupling — Calling a system portable while prompts, tool schemas, identity, traces, or state depend on undocumented vendor behavior.
- Outsourced judgment — Allowing a model provider’s defaults to become the organization’s approval, escalation, or risk policy.
- Implied endorsement — Describing ordinary product use in language that suggests an official partnership, certification, affiliation, or approval.
- Open weights without operations — Treating downloadable weights as independence without owning serving, evaluation, security, observability, fallback, and upgrade responsibility.
The Test
When evaluating a model or agent platform, ask:
Could we replace this instrument without losing the workflow, policy, evidence, or institutional learning?
If the answer is no, the system is captured. If the answer is yes but no model has been selected deliberately, the system lacks conviction. Good architecture holds both conditions at once: a strong present choice and a credible path to change it.
Relation to Other Concepts
- Gelassenheit supplies the stance: use technology fully while remaining free from capture by it.
- Complementarity locates human judgment and machine execution inside one accountable whole.
- Crystallization turns judgment into portable, inspectable policy rather than leaving it inside a person or a model session.
- Timeless Materials favors standards and interfaces with enough durability to survive vendor and framework churn.
- The Hermeneutic Circle requires every platform choice to serve the whole system and be revised when practical evidence changes the whole.
Conviction without dependence is the operational form of these principles for the model era: engage fully, choose deliberately, keep judgment visible, and remain able to change the instrument without losing the work.