As far as I can infer research isn’t uniformly landing on one end or the other of the spectrum, but it’s easy to find less useful explanations on one side of it.

And so there’s a distinction between describing this “this is how people report they build awareness and update models of the world” and noting contextual challenges to this process, vs. “the bad decision implies bad awareness and since we designed for awareness, this has to be individual issues”

Underwood distinguishes five levels of con-cepts. However, the distinction between Level 2 and Level 3 is where I would like to focus: A Level-2 concept is one which summarizes the operations used to define a phenomenon and therefore merely identifies the phenomenon. I call it phenomenon identification or phenomenon naming. The definition of the phenomenon

implies not one thing about a causal process or condition over and above the operations per se. (P. 198) Level-3 concepts name or identify a phenomenon just as do Level-2 concepts; but, the name is applied to a hypothetical process, state, or capacity as a cause for the observations indicating the phenomenon. (P. 200) Is SA a Level 2 concept or a Level 3 concept? And why should anybody (except a fuzzy academic such as myself) care? Underwood's answer is that Level 2 and Level 3 concepts are based on the same formal operations but are "'thought about' differently by psychologists" (p. 202). In particular, he warns that "occasionally, having defined the term at Level 2, the writer may slip and talk as if it (the defined phe-nomenon) is now causing itself" (p. 202). Figure 1, which is reprinted from Underwood (p. 203), illustrates the difference between these two levels of concepts. In describing this figure, Underwood wrote, For the Level-3 concept I have drawn bidirectional arrows between X and Rd. For in using these concepts the investigator infers a state or process only if a reliable difference in response occurs and then says that this difference is caused by the state or process (X). Differences in X are in turn caused by Sm. If this sounds to you like scientific double-talk, then at this point I must agree. And it should be mentioned that Level-3 definitions do not always make circularity of the inference so obvious as I have made it here, but it is inevitably present. (P. 203) Consider some definitions for SA in light of Underwood's cautions. Figure 1. A comparison of Level 2 and Level 3 concepts. Sm indicates stimulus manipulation (i.e., independent variable). Rd indicates response differences (i.e., dependent variable). A Level-2 concept is defined by referring directly to the relation between Sm and Rd. A Level 3 concept identifies a state (X) as causing Rd and this state is, in tum, related to Sm.

Folk models typically describe measurements that reflect an important aspect of the human condition but that refer to intervening variables - representing intermediate mental states - rather than to observable performance. As shown by Figure 3.2, performance can be characterised in many different ways, such as degree of efficiency, speed and precision of actions, performance deviations or 'errors, etc. Performance depends on individual attributes such as experience, motivation, capacity, skills, etc. Performance is also affected by working conditions, which in turn can be described by means of a number of general and specific factors. Folk models propose that one or more of the intervening variables are strongly correlated with the quality of performance, and that measurements of this intervening variable provide essential information about how well the individual is able to accomplish a given task. Some of the more common intervening variables are shown in Figure 3.2. In order for a folk model to be of practical use, it is necessary to assume that the measurement - or rather the hypothetical state that the measurement refers to - is affected by the working conditions in the same way that an individual being in the situation would be. (It also implies that the hypothetical state is the prime determiner of operator performance.) It must also be assumed that the measurement can be used to predict short-term changes in performance, since that is the main reason for not just observing what happens. Figure 3.2: Loosely defined measurements. - Intervening variables: workload, situation awareness, attention, fatigue, WM management - Working conditions: general / specific factors - individual elements: Experience, motivation, capacity, skills,... - Direct measures of performance quality: efficiency, speed & precision, performance deviations

I don’t know that it’s fully invalid as a construct or been strengthened enough over time, but I liked this little distinction by Flach (1995) to show the risk of turning it from a description of a phenomenon to a causal process, and then Hollnagel’s later complaints that sort of crystallizes it.

I have spent the last few days reading over fifteen papers on situation awareness and the validity (or lack thereof) of the model and its epistemological stances. I am now distressingly aware of the situation with situation awareness, thus giving me a final reason for avoiding the construct.

Black hoodie showing 'anti complexity complexity club' in a red explosion based on the logo of the Resilience in Software Foundation.

hell yeah @resilienceinsoftware.org swag is in!

the law of requisite variety states that only variety in the regulator can destroy variety in the system being regulated.

so if you need to deal with complexity you know you gotta join the club & begrudgingly increase complexity to keep things simple

Imaginaries of omniscience: Automating intelligence in the US Department of Defense - Lucy Suchman, 2023 The current reanimation of artificial intelligence includes a resurgence of investment in automating military intelligence on the part of the US Department of D...

You might also like Suchman’s Imaginaries of Omniscience: automating intelligence in the US department of defence (journals.sagepub.com/doi/10.1177/...) which criticizes the concepts behind this type of automation feeding into itself and lacking orientation (my summary at ferd.ca/notes/paper-...)

That being said, modelling is a definite challenge and part of the reason people see a value in prototyping and iterating fast, not arguing with that at all, mostly wondering about the proportions.

Like if we suppose faster production of code increases interactivity at a pace the testing methods aren’t keeping up with (hence falling back to formalisms in places), it’s unclear better modelling would change the underlying properties except by reducing the need for new code by narrower designs.

not sure I’d fully agree that MOST problems would be domain modelling problems in quality issues involving AI. If many bugs are found through users walking code paths where the intent was right but outcome not, then maybe insufficient walking of designed paths are more relevant than bad modelling?

Exactly.

More like how a plane that blows up tends to not break down anymore after that one time

What if I told you you could cause it on purpose today

would you rather have many incidents of various small to moderate size for the foreseeable future or just one very big incident and then be done with them for good?

I recall Jim saying, on a day when I felt particularly helpless: "The doors of the people who should be talking to you will likely be closed, stop knocking on them. Look for the open doors' This took me out of linear thinking, into the emergent maze of change and complexity. I've stopped knocking on closed doors. I've stopped engaging in activities or conversations that will do nothing but maintain the status quo. I've learned that systemic change requires disruption - that I too need to shift how I think and act. I've come to see that kindness can be more disruptive than aggression, that compas- sion can be more impactful than taking positions. I've learned to plant seeds, and not to worry about which ones live and which ones die. That is the nature of complex change.

Here's a quote I always loved from Gill Kernick's "Catastrophe and Systemic Change" that I feel you might like given these recent experiences. Even if it's hard to apply, it's been stuck in my mind since then.

It sounds like this approach, or something close to it, is working well for you :)

They’re from my hometown, saw them a few times live in tiny venues! Good shit, feels even nicer live with the loudness and vibrations and everything.

A comparison I heard was about specific cockpit automation (which led to a crash) was to compare it to “management incursion at a distance”, because it encoded rules and policies to override pilots at landing time in certain conditions without knowledge of local context.

I still think about it.

As usual @grimalkina.bsky.social is worth listening to.

These principles are also worth considering and applying in all sorts of contexts. Here’s a sample from safety research I happened to read just yesterday (Dekker - Reconstructing human contributions to accidents, 2002) that aligns with it!

Many of the “risks” of building software without specialized expertise are also negated by keeping it personalized and local; the potential for harms and surprises is reduced, and remains easier to work around while gaining from the upsides. This is a good fit!

(Excel may be the closest past case?)

When I joined the program I'm currently in, I told myself I would not read more papers and technical books outside of it because I'd need to balance about my energy levels—not spending it all on this.

I was right, but also there's lots of other cool nerd shit I want to read through now and welp.

Incident Report: Exercises, Cleanups, and Evacuations Every year, Honeycomb runs disaster recovery scenarios in multiple environments, including in production. Although each of our instances runs in a single region, on at least three Availability Zones ...

Here is the fuller writeup I promised, a little bit overdue (I said Jan but it actually went out in Feb). Credit to @ferd.ca for the writeup, as well as to all of our incident responders who worked this 12+ day incident. www.honeycomb.io/blog/inciden...

I really need to do this year’s garden planning. It’s gonna be time for seedlings in a few weeks and it’s gonna be nice to once again get going on one of them hobbies where you can’t really obsessively dictate the pace nor feel pressured in going faster. Just watch the plants grow and see what goes.

Incident Report: Exercises, Cleanups, and Evacuations Every year, Honeycomb runs disaster recovery scenarios in multiple environments, including in production. Although each of our instances runs in a single region, on at least three Availability Zones ...

Back in December, we had a large outage at work. The internal investigation took a while and the internal report was roughly 40 pages long. For the public, we managed to try and condense it to a much shorter format that we think can still offer useful insights to other organizations:

You gotta taste the bits where the cheese is not there

I believe I had a relatively intuitive sense of how much Swiss cheese I could consume in one sitting before I can take no more.

I also believe I now have a fairly empirical sense of how much literature about the Swiss Cheese Model I can consume in one sitting before I can take no more.

I also think this discussion about how SRE work is being devalued by these products is at least parallel to the discussion about how people are willing to write clear documentation for *AI* consumption, but never put any value on it when it was for *actual people*.

AI SREs are framed as nameless job automation; Coding Assistants as named partners.

The framing used to create these products reveals a lot about how the builders and buyers perceive these roles. I also write about the challenges and risks of picking self-limiting analogies in building systems.

this thread is touching on something i find very important: efficiency is sometimes opposed to (or in tension with) other goals

efficiency can preclude generalist systems that create flexibility, decentralization and surplus that add resiliency, or non-expert participation that gets people involved

From a discussion in RISF based on the old IBM adage, an updated version for the modern era:

Table 5.2: Example of transfer surprises. 3 columns are given: transfer tasks, expectations based on common elements theory, actual outcomes. Example row: Rodeo riding vs. tournament jousting; expected positive transfers due to similarity in riding and falling without injury; found a large negative transfer because rodeo riders look to the back which doesn't work in jousting.

This sent me skimming through Accelerated Expertise by Robert H. Hoffman et al. again. I was wondering if there's a backwards way of going "if expertise transfer works, domains have to be similar", but they make the point that Transfer is incredibly difficult to predict and measure :/

A manual remake of "A successful Git branching model" by Vincent Driessen (nvie.com). Flowchart illustrating a Git branching strategy with time flowing downward (bold arrow labelled "Time" on the left side). Five vertical lanes across the top: "feature branches", "develop" (bold), "release branches", "hotfixes", and "main" (bold). Coloured circles represent commits on each branch: pink for feature branches, yellow for develop, green for release branches, red for hotfixes, and cyan for main. Arrows between commits show the direction of merges. Text: - "Tag 0.1" (on main, at the top) - "Major feature for next release" (callout pointing to a feature branch merging into develop) - "Feature for future release" (callout pointing to a feature branch on the far left) - "Severe bug fixed for production: hotfix 0.2" (callout pointing to a hotfix branch forking from main) - "Incorporate bugfix in develop" (callout pointing to the hotfix merging into develop) - "Tag 0.2" (on main, after the hotfix merge) - "Start of release branch for 1.0" (callout pointing to a release branch forking from develop) - "From this point on, 'next release' means the release *after* 1.0" (callout with dashed line pointing to develop, where "after" is italicised) - "Only bugfixes!" (callout pointing to commits on the release branch) - "Bugfixes from **rel. branch** may be continuously merged back into **develop**" (callout with arrows showing merges from the release branch back into develop; "rel. branch" and "develop" are bold) - "Tag 1.0" (on main, after the release branch merges in) The diagram shows the complete lifecycle: feature branches fork from and merge back into develop; a release branch forks from develop when ready, receives only bugfixes, and merges into both main (creating a tagged release) and back into develop; hotfix branches fork from main for urgent production fixes and merge into both main (creating a tagged patch release) and develop.

same as previous but in hand-drawn style

here, I tried it and did it in ~12 minutes, fully vectorized and editable with an off-the-shelf app like draw.io. Now it's fully maintainable and styles modifiable, without working directly in SVG, and it's all costing nothing extra.