Vibeleaderboard
← Course overviewModule 06 · Choose a setup that fits · Beta
Lesson 6.2

Module contents

Choose a setup that fits

Lesson 2 of 5

Model, harness, connections, or data path?

When two setups differ, the difference lives somewhere: in the model's learned values, in what the harness can do and enforce, in which operations are connected, or in where the tenders' contents travel. Product descriptions blur these, and the words open, local and private blur them further. This lesson separates the four parts and gives four questions that turn a privacy claim into something you can check.

Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.

Four parts that can differ

Module 1 defined a by its , the learned numerical values that generation uses. Two setups with different models will tend to generate differently from the same context, though a different model does not guarantee a different answer on any one request, and the harness also shapes what the model sees and how generation is sampled. Module 3 defined the as the software that sends the model its context, recognizes tool calls, runs or refuses them and keeps the record; two setups with the same model and different harnesses can behave very differently, because the harness decides what can be attempted and what is enforced. Module 5 showed that a harness's operations come from connections, through a command line, an API or an ; two harnesses with the same features and different connections can reach different things. And module 5's last two lessons showed that the path data takes, into a provider, into an , onward to a connected service, is a property of the whole arrangement and not of any one part.

For the tender review, each part carries a different question. Model: does it find discrepancies in documents of this kind, cite pages accurately and add the price schedule correctly? Harness: can it read the files and write one report, does it ask before writing, what does it enforce? Connections: is there a route to the tender files, and what may that route reach? Data path: where do the tenders' contents go at each step, and who keeps what? Lesson 1's four requirements map onto these, and a comparison that names the part it is talking about is one that can be checked.

Available, permitted, run where, sent where

The words that cause the most confusion are the ones used for the data path, so this section takes them one at a time and turns them into four questions.

A model's weights are its parameters as stored numbers. A model has when its publisher has made those numbers available to download. That answers one question: are the weights available? It does not answer the next: may the committee use them for this purpose? That is set by the license the publisher attaches. Hugging Face, a service that hosts models, documents that a model's card names a license, either a standard one or a custom one with a link to its terms, and that the card also carries evaluation results the publisher supplied. 4 A card is the publisher's documentation, not an independent check, and a downloadable model is not thereby licensed for every use. Two questions, then, before anything runs: available, and permitted.

The third question is where inference runs. means the model's computation happens on a machine you control, whether the laptop in front of you or a server you administer somewhere else, rather than on a provider's servers. Open weights make that possible; they do not make it so, since the same weights can be run by a hosted provider. Ollama, one program for running models locally, documents in its FAQ that its local models process prompts on the user's machine, that it also offers cloud models which process requests on Ollama's servers without storing them or training on them by the FAQ's account, and that cloud features, including web search, can be switched off by a setting at the cost of losing them. 3 That page, read on 29 September 2026, is one product's account of itself on one date. The reason it is here is that it shows the third and fourth questions coming apart inside one program: the same harness can run a model locally and still send a request elsewhere for a feature. Running locally also has costs that no provider invoices: the machine, its power, and the time to set it up and keep it working.

The fourth question is where the surrounding application sends data, and it has more than one answer even for a single provider. OpenAI's documentation of its API data controls, read on the same date, says that data sent through its API is not used to train its models by default, that it may be retained for a period for abuse monitoring, that retention differs by endpoint and by configured controls, and that data an application sends on to a third-party MCP server is subject to that server's own retention and location policies. 5 Four separate facts, and one of them is about a route the provider does not control. A statement that data is not used for training says nothing about retention; a retention period says nothing about location; and a connected service has its own answers.

So is not one property but a set: training use, retention, location, and onward recipients, each answered for each place the data goes. Two constraints an organization might set land on different members of that set. A retention constraint, like the committee's policy, asks how long each recipient keeps the data; a provider's documented retention period is the fact to compare it with. A transfer constraint asks whether the data may be sent to any third party at all; no retention term answers it, because the transfer has happened before retention begins. To check a candidate against either you would ask: where does the harness send the tenders' contents (to which model, run where, through which connections), what does each recipient's documentation say, and what has the harness itself been configured to do with search, telemetry, logs and backups. A setup with an open-weights model run on the laptop answers the model's part of that with “nowhere,” and still owes answers for every other route.

Harness features are capabilities and rules, not outcomes

The other blurred comparison is between harnesses. A harness with twenty integrations, several agent roles and a planning mode looks more capable than one with a file tool and a shell tool, and a comparison by feature count would say so. Module 3's layers say what a feature actually is: a capability the model may call, a rule the harness applies, or a boundary it enforces. Each is something the model may use well, badly or not at all.

Two dated examples show the range. OpenCode, a coding harness, documents its agents as a combination of a prompt, a model and a set of permissions; its documentation gives a reviewer agent that is denied editing and shell access, and it describes subagents that start with their own context and their own permissions. 6 Read the reviewer example the way module 5 read an access line: two routes to changing files are closed, and any other tool that can write would need its own denial, so the scope of the configuration is what to inspect. Mario Zechner, whose minimal agent module 5 mentioned, describes building his with four tools and preferring command-line programs with a short instruction file over dedicated integrations. 7 Neither is a recommendation for the tender review. The first shows a harness whose features are largely rules about roles; the second shows one whose author deliberately kept the capability list short. Which reviews tenders better is not answerable from either description.

What can be said in general is narrower. Capability descriptions have to reach the model to be used, and module 5's fourth lesson showed that some harnesses load them all up front and others on demand; either way they are context, which module 4 explained is not free. Every mode or role is a further place for the arrangement to go right or wrong. Whether a feature helps is a question about the task. A mode that restricts the harness to reading and planning, for example, might suit a first pass whose only write is the final report, or might add steps the review does not need; the trial says which. One vendor's account of building an elaborate harness for long application-building runs, with separate planning, generating and evaluating components, reports that the components addressed observed failures at real cost, and recommends removing them one at a time when testing a stronger model to see which are still earning their place. 8 That is vendor practice on its own tasks with its named models; the transferable part is the method, which is the same as lesson 1's: try the task, observe, and let the observation decide. The site's harness comparison lists current harnesses and their documented features; read it as a list of what can be attempted and enforced, and not as a ranking of outcomes.

Use the Vibers directory to find practitioners whose work resembles yours. Read a recommendation alongside who tried it, on what task and with which setup; a preference can be useful without becoming a rule for everyone.

Carry this forward

  • A setup differs in four places: the model's learned values, the harness's capabilities and rules, the connections that supply operations, and the path the data takes. Name the part a comparison is about.
  • Open weights answers whether the weights are available; the license answers whether the use is permitted; local inference answers where the model runs; data handling answers what each recipient does with the data. A privacy claim has to answer all four, for the whole path.
  • Data handling is a set of questions, training use, retention, location and onward recipients, answered per recipient from that recipient's documentation on a stated date. A retention constraint and a transfer constraint land on different questions, and not-for-training is not not-retained.
  • Harness features are capabilities and rules, not outcomes. Whether a feature helps is a question about the task, answered by trying it; a permission profile closes the routes it names and no others.

Lesson check

A member proposes running an open-weights model on the treasurer's laptop for the tender review and says this settles the committee's data-handling questions, because nothing leaves the machine. What does this lesson say?

Choose an answer for each question.

This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.

One of the four parts, the model, arrives with published scores attached. The next lesson reads one of those scores as the measurement it is.

Sources · Reviewed 2026-09-29
  1. 4. Hugging Face Hub documentation: Model Cards (specifying a license; evaluation results)
  2. 3. Ollama documentation: FAQ (local and cloud models, disabling cloud features)
  3. 5. OpenAI platform documentation: Your data (retention by endpoint, abuse monitoring, third-party MCP servers)
  4. 6. OpenCode documentation: Agents (prompt, model and permissions; read-only reviewer example)
  5. 7. Mario Zechner: What I learned building an opinionated and minimal coding agent (2025-11-30)
  6. 8. Anthropic Engineering: Harness design for long-running application development (Prithvi Rajasekaran, 2026-03-24)
  7. 14. VibeLeaderboard: Harnesses (current harness comparison)
  8. 15. VibeLeaderboard: Vibers practitioner directory

Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.