The newest news in raw, unedited form is in the Journal of the PROCESS, of BUILDING this little repository or agentic mdBook monorepo of skill *.md files ... check the Journal first, if you are revisiting -- otherwise, look at the repository or maybe fork/clone it.

Five Strategic Causal Engineering Visions

Causal Engineering has one objective: maximize the discovery of surprises — new actionable insights and useful knowledge. We focus on causal drivers because that's what makes the whole process efficient. Take business prospecting for a scalable company. Instead of spamming people, getting permanently blocked, and irritating uninterested prospects with robo-pitches, we design campaigns — freemium models, for instance — to maximize closable leads per $1,000 spent. The point of the campaign is to understand what actually causes someone to use our product, not to assume prospecting is just a numbers game.

1. The Orthogonal Opportunist: Exploiting Natural Experiments in Administrative Data

Treat every policy change, geographic boundary, and bureaucratic threshold as a free randomization you didn't have to pay for, then deliberately construct a portfolio of these natural experiments whose identifying assumptions fail in different ways—so that no single unobserved confounder can kill the whole portfolio. This is Taguchi thinking applied to causal identification itself: you are not just estimating effects, you are arraying designs orthogonally so that the weaknesses of one design (e.g., a regression discontinuity's local external validity) are covered by the strengths of another (e.g., a difference-in-differences' broader scope). The strategic bet is that the world is littered with quasi-random variation that no one has bothered to harvest because it's unglamorous, and that the marginal cost of adding the twelfth natural experiment to your array is nearly zero once the data pipeline exists. The risk is real: reviewers may see a "grab bag" of designs rather than a coherent identification strategy, and if your portfolio is correlated in its failure modes (e.g., all designs rely on the same parallel-trends assumption), you've built false confidence at scale. Mitigate by pre-registering the heterogeneity of assumptions across designs—making the fragility itself the object of study—and by publishing the "assumption coverage matrix" alongside the estimates.

2. The Negative-Space Cartographer: Mapping What Isn't There

Systematically catalog the absences in free data—the products never launched, the trials never run, the regions never treated, the queries never asked—and treat these null spaces as your most information-dense research frontier, because selection into non-existence is itself a causal process that reveals the decision rules of the actors you're studying. This is the prospector's inversion: instead of drilling where everyone else drills (published effects, successful launches, treated units), you drill where the absence of a drill-hole is informative, using the pattern of missingness as a signal about costs, beliefs, and constraints. The efficiency argument is that absence-data is almost always free (it's the complement of data someone else already collected), and the surprise-per-dollar is enormous because you are testing hypotheses no one has articulated—"why did this firm stop at formulation 47?" is a question with a cheap answer and a potentially huge payoff. The risk is that absence is unfalsifiable by default: you cannot distinguish "no one tried because it wouldn't work" from "no one tried because of a historical accident," and you may spend your budget building elegant theories of nothing. Defend against this by triangulating absences against revealed constraints (patent cliffs, regulatory calendars, procurement records) so that each null-space claim is anchored to at least one observable decision cost.

3. The Fractional-Factorial Field Lab: Running 2^(k-p) Designs on Cheap Interventions

Embrace the Taguchi insight that you can estimate main effects of many factors with a fraction of the runs, and apply it not to manufacturing but to behavioral or operational interventions in the field—running 8 or 16 armed mini-experiments where a naive team would run 64, accepting aliasing between high-order interactions in exchange for testing eight hypotheses per dollar instead of one. The vision is a "screening factory": cheap, fast, low-power experiments whose purpose is not to confirm but to rank which factors deserve the expensive follow-up, with the explicit understanding that Resolution III designs will confound main effects with two-way interactions and that this is an acceptable price for breadth. The strategic payoff is that most intervention effects are main-effects-dominated, so the aliasing cost is small relative to the discovery yield, and the free infrastructure (existing survey panels, app telemetry, call-center logs) absorbs most of the cost. The risk is the classic one: if the true structure is interaction-heavy (e.g., the intervention only works in combination), your fractional design will systematically mislead you into killing a factor that matters, and you'll never know it. Insure against this by reserving 20% of budget for foldover runs on the top-ranked factors, converting your screening array into a confirmatory one exactly where the stakes justify it.

4. The Adversarial Replication Exchange: Paying for Disconfirmation, Not Confirmation

Restructure your research budget so that a fixed fraction (say 30–40%) is spent only on attempts to falsify your own prior findings using free or cheap external data, and make the surprise—the failure to replicate—the primary deliverable rather than a embarrassment to be buried. The philosophical bet is Popperian and economic at once: the marginal dollar spent trying to kill a result buys more information than the marginal dollar spent extending it, because confirmation has diminishing returns while disconfirmation has option value—it tells you which of your beliefs are load-bearing before you build on them. This is efficient because replication attempts on already-collected public data (or on data you can obtain through FOIA, open science repositories, or scraping) cost a fraction of original data collection, and the "orthogonal array" here is over threats to validity: you deliberately array your replications so that each targets a different failure mode (measurement, selection, specification, publication). The risk is institutional and reputational: funders and collaborators may read a portfolio of failed replications as a portfolio of your failures, and the strategy can be gamed into nihilism where nothing is ever believed. Guard against this by pre-committing to a decision rule—"we will treat a finding as load-bearing only if it survives at least two of four adversarial probes"—so that disconfirmation buys calibrated confidence, not reflexive skepticism.

5. The Surprise-Weighted Portfolio: Budgeting for Maximum Entropy Reduction

Treat your research portfolio as an information portfolio and explicitly optimize for expected entropy reduction per dollar rather than for expected effect size or expected publication, which means deliberately over-weighting hypotheses that are (a) cheap to test, (b) currently near 50/50 in your prior, and (c) mutually orthogonal in what they would teach you if resolved. This is the deepest Taguchi analogy: just as orthogonal arrays maximize the information extracted per run by making columns independent, a surprise-weighted portfolio maximizes learning per dollar by making your hypotheses independent in information space—so that resolving one doesn't make the next redundant. The efficiency comes from the free-data augmentation: many of the highest-uncertainty, cheapest-to-test hypotheses live in data you already have but haven't interrogated (the unused columns of a survey, the untapped logs, the second-order interactions you never checked), so the marginal cost approaches the cost of attention rather than the cost of collection. The risk is that entropy-reduction is a process objective, not a substantive one—you can maximize surprise per dollar and end up with a portfolio of true-but-trivial findings, having optimized for the thrill of learning rather than for learning what matters. Defend against this by imposing a substantive constraint on the portfolio (e.g., "every hypothesis must connect to a decision someone will actually make") and by periodically re-weighting toward consequential uncertainty, so that the entropy you reduce is the entropy that stands between you and action.