What AI evidence actually exists: inference layer and model layer artifacts
Free. Watch without an account. Sign in to save progress and take the exam.
Up next
Why AI evidence disappears before you ever request it in discoveryTranscriptThe player shows YouTube’s automatic captions, which are machine captions we do not review. The full transcript is below.
00:00:00Maybe, you know, Mari, you want to walk through sort of the inference layer artifacts and then Karen, sort of the model layer artifacts.
00:00:07And these are just basically what exists, right? Just at a basic level. And it varies based on your model and as an open source and who hosted it and all these things.
00:00:17But, Mari?
00:00:19Sure. So at a high level, I kind of think about the inference layer artifacts as the things that sort of the individual is interacting with or reflects what the individual is doing.
00:00:30So prompts and outputs, sort of the questions or statements that you put into a tool and what it feeds back to you.
00:00:37The system prompts and configuration is basically like a layer above that what sorts of instructions exist.
00:00:43Tool call and agent logs basically showing what agentic workflows were allowed to do or not do.
00:00:49And then retrieval records and usage telemetry basically show for the reference records, it's like the documents that the tool was allowed to go access to the extent that you had those.
00:01:01And then telemetry, I almost view as like, I think the easiest corollary is like the fax machine admin list that you used to get.
00:01:11So I used to get like this printout at the end of the day that had like what time faxes were sent out and which phone numbers and how long it took and like whether it worked or didn't work.
00:01:22To the extent that anyone on this call still remembers fax machines. So usage telemetry, I sort of put in that bucket, you know, who did what at what time, how long as sort of like the admin data points.
00:01:36Karen, model layer artifacts.
00:01:38Yeah, I do want to distinguish between inference and model layer. Is that helpful?
00:01:42Yeah.
00:01:44So inference is sort of what the user, I mean, just for like, again, this is one of those words that will need definition in any context you use it.
00:01:52But generally, it's meant to indicate that that user layer on the model, right?
00:01:58So it's what we all do with these models that we've been handed.
00:02:03The model layer artifacts we'll talk about is what's going on with that underlying engine, as opposed to what somebody did with that engine and where they drove the car and when and how long they stayed and that kind of thing.
00:02:14So the model layer artifacts, and this is tricky because the models themselves are very mutable. They update the outputs they produce are themselves changeable as we discussed.
00:02:25So what do we have that sort of artifactual that could be discovered in retrospect?
00:02:33If not the whole training and fine tuning data sets and provenance for those, at least, which would be, to Dan's point, enormous.
00:02:43Some description of them, some archival kind of reflection of them should be discoverable in most instances.
00:02:57And it will have been produced in the context of other work that they have to do, whether that's reporting work or licensing or some other, some such, or in the course of prior litigation.
00:03:09Model cards and evaluation reports are the materials that the labs themselves produce to describe the models, their capabilities, their limitations, inputs and, you know, expected outputs.
00:03:23Sometimes they include, you know, name states and places. Basically, it's sort of if you were to, you know, if you were to describe a model on a, you know, on a two dimensional piece of paper.
00:03:39This is what this is what it's a description. It's a model description at the time, generally speaking, of testing and deployment, not always, but that's generally speaking.
00:03:49Version history, like anything else, you know, the 1.0, 2.0 should be, should be discoverable. Validation and testing records, I would imagine that sometimes that's available.
00:04:03I think the disappointment most of us would find is that those are much less comprehensive than we might expect.
00:04:12What have these models been tested for and why and by whom is a big conversation right now, particularly in around like when is independent testing required and what is independence really mean in this context.
00:04:26Also interesting here, just to put a footnote in it is there's sort of increasing evidence that these models are very good at detecting when they're being tested.
00:04:35And so, you know, and adapting and adjusting their approaches and results in response to that insight. And so, big question about what really qualifies as as validation at this point.
00:04:53But but efforts that this should be discoverable, at least around the they want to make you happy. That's why I tell people about the AI models generally want to make you happy.
00:05:05So, like, I don't want to over generalize, but like I've seen a lot of this validation testing and records and like.
00:05:15If you tweak it a little bit, are you coming clean or like it remembers the history, there's a whole lot of pieces that come into this, but they're important records.
00:05:26There should be records. I just, it's just never going to be as clean as, you know, we tested it at past so we released it. Right. I mean that sort of the way you would.
00:05:36Don't ever. Yeah, I mean, this is not a car factory right in that way.
00:05:42And then the governance materials, which I think are also really spotty and all over the places, people are trying to figure out sort of what they need to have by way of policy, what kinds of training need to underpin that policy.
00:05:54What kinds of documentation is one of the things that that most organizations we work with either forget about or don't think about being a core piece of their governance process.
00:06:08That was changing a little bit, but but I'm just going to guess that's going to be all over the place if you're looking at anything 2026 or earlier.
00:06:18Because it's very it's very uneven terrain. Yeah. Well, like I tell people, depending on how sophisticated your AI is, some of the AI I'm working with.
00:06:29The governance actually can serve as feeds into a heart into a heart. I don't want to be too technical, but can you use to help inform how a harness will operate to make sure that your AI systems generally are following in certain things.
00:06:44So having good governance actually has a lot of value. I think you're going to see a huge shift as AI gets more widely deployed into having more robust AI governance because I think that will be a.
00:06:57I call it a two for right. You can use the governance materials to show regulators that you're doing your razzle dazzle magic and you can also use it to help guide and navigate your actual deployment of your AI model.
00:07:09So it sort of offers a double win and it's but I could be wrong. So who knows why for what it's worth. I think it's important.
Automatically generated; it may contain errors.
Study materials
Notes, flashcards and practice questions for this lesson, reviewed before publication.
Study notes
Distinguishing inference layer and model layer artifacts in AI discovery
Daniel B. Garrie frames the discussion by asking Mari Grace to walk through inference layer artifacts and Karen Silverman to walk through model layer artifacts. He notes that what actually exists as evidence varies depending on the model, whether it is open source, and who is hosting it.
Karen Silverman draws the core distinction: the inference layer is the user layer on top of the model — what people do with a model they have been handed. The model layer is the underlying engine itself, separate from how someone used it, where they "drove the car," and for how long.
Inference layer artifacts described by Mari Grace
Mari Grace describes inference layer artifacts as reflecting what an individual is doing when interacting with a tool:
- Prompts and outputs – the questions or statements a user puts into a tool and what the tool feeds back.
- System prompts and configuration – the layer of instructions sitting above the prompts and outputs, governing what the tool is set up to do.
- Tool call and agent logs – records showing what agentic workflows were allowed to do or not do.
- Retrieval records – reference records showing which documents a tool was permitted to access, to the extent such access existed.
- Usage telemetry – administrative data such as who did what, at what time, and for how long.
Mari Grace compares usage telemetry to an old fax machine admin printout that listed what time faxes were sent, which phone numbers were used, how long transmission took, and whether it succeeded or failed. Usage telemetry serves a similar administrative, record-keeping function for AI tool use.
Model layer artifacts described by Karen Silverman
Karen Silverman explains that model layer artifacts are harder to pin down because models themselves are mutable — both the underlying engine and its outputs can change over time. She identifies what might still be discoverable:
- Training and fine-tuning data sets and provenance – the full data sets would be enormous, but some description or archival reflection of them should be discoverable in most instances, often because it was produced for other purposes such as reporting, licensing, or prior litigation.
- Model cards and evaluation reports – materials labs produce describing a model's capabilities, limitations, inputs, and expected outputs, sometimes including named states and places. Karen Silverman likens this to a description of the model on a two-dimensional piece of paper, generally created at the time of testing and deployment.
- Version history – like any versioned product, the 1.0, 2.0, and so on should be discoverable.
- Validation and testing records – sometimes available, but often far less comprehensive than expected. Karen Silverman raises an open question about when independent testing is required and what independence really means in this context. She also flags growing evidence that models are good at detecting when they are being tested and adjusting their behavior in response, which complicates what counts as genuine validation.
How testing records fall short of expectations
Daniel B. Garrie adds that AI models generally "want to make you happy," which affects the reliability of validation testing and records. He notes there are many open questions, such as whether a system is "coming clean" after being tweaked or whether it retains memory of its own history. He concludes that testing and validation records will never be as clean as a simple claim that a product was tested and then released — this is not like a car factory process, as Karen Silverman agrees.
Governance materials and their value in AI discovery
Karen Silverman describes governance materials as also "spotty and all over the place." Organizations are still working out what policies they need, what training should underpin those policies, and what documentation should exist — something many organizations either forget about or do not treat as a core part of governance. She expects this to remain uneven terrain for materials dated 2026 or earlier, though the situation is slowly changing.
Daniel B. Garrie notes that, depending on how sophisticated an AI system is, governance materials can feed into a "harness" that helps ensure AI systems follow certain behaviors, giving good governance real operational value. He predicts a shift toward more robust AI governance as AI is more widely deployed, describing this as a "two for": governance materials can be used both to show regulators that an organization is managing its AI responsibly and to guide and navigate the actual deployment of the AI model itself.
Flashcards
13 cards. Test your recall, then flip to check.
Select the card, or press Space, to flip it. Left and right arrow keys move between cards.
1 of 13
Practice quiz
Practice — not graded, not the final exam