Work out whether it is true, and what follows from it.

Find the thing and say where it is. No claim about whether it holds.

• 8locate a claim in its primary source

  Proof: A claim traced to its source with the section named, including one case where the source turned out to be weaker than the claim made from it.

  AI Alignment · next Sep 25 AI-Assisted Attacks: What Actually Happened Critical Thinking and Creativity with AI · next Sep 11 Data Activism: Turn Data Into Power · next Sep 14 Prompt Engineering Research & Activism · every week

• 8locate the devices a passage uses on you

  Proof: One passage you like with at least five devices marked in the text, each named, and each tied to the reaction it produced in you.

  Critical Thinking and Creativity with AI · next Sep 11 Practical Propaganda · next Sep 19 Prompt Engineering

• 8locate the line an error points to

  Proof: A stack trace, the line you navigated to from it, and one case where the line named was not where the mistake actually was.

  Make & Do · every week

• 8locate the standard that governs a system you are building

  Proof: One system, the governing instrument named, and the specific requirement it places on you.

  AI Security: Governance, Standards and Safety Cases Blue Team Common Certification Field Opsec

• 8name an audience and its success criterion

  Proof: A written audience statement naming who they are, what they are trying to do, and the observable event that means it worked.

  Critical Thinking and Creativity with AI · next Sep 11 Data Activism: Turn Data Into Power · next Sep 14 Founding Federation: Smallest Sellable Thing Founding Federation: Price and Ideate (Week 0) Practical Propaganda · next Sep 19 Solarpunk Automation

Execute a given procedure and read the result off it.

• 8run a model analysis against a hand check

  Proof: The dataset, the model's analysis, and your own computation of one figure alongside it, agreeing or with the discrepancy explained.

  Data Activism: Turn Data Into Power · next Sep 14

Make an artifact that did not exist, to a stated standard.

• 9produce a budget for work whose timeline you have observed

  Proof: The completion gate answered out loud, and either a budget built on an observed timeline or a scoped first run with a defined end and a showable output.

  Data Activism: Turn Data Into Power · next Sep 14 Founding Federation: Build With Your Heart First Founding Federation: Price and Ideate (Week 0)

• 9produce a function whose result depends on program state

  Proof: The function, two call sites that do not know about each other, and two runs of the same call producing different results with the state that caused each difference named.

  Advanced Interactive Fiction The Interactive Fiction Intensive · next Aug 29

• 9produce a list of where untrusted input enters a system

  Proof: An enumerated list against a real codebase or system, including one entry point that is not an obvious form field.

  Agentic AI Security: Securing What You Build Agentic SDLC · next Sep 7 Blue Team Common Certification

• 9produce a moderation score for an output

  Proof: A scored set of at least twenty outputs with your own label beside each score, and a written note on each disagreement.

  AI Alignment · next Sep 25

• 9produce a refusal boundary that fires on cases you did not list

  Proof: Three declines on cases absent from the boundary text, three accepts on near neighbours you did want, and the near miss that made you rewrite it.

  Context Engineering · next Sep 16 Prompt Engineering

• 9produce a rubric a model can apply

  Proof: The rubric plus two independent gradings of the same three outputs, with the scores side by side.

  AI Alignment · next Sep 25 Critical Thinking and Creativity with AI · next Sep 11 Custom GPTs Prompt Engineering

• 9produce a segment definition from observed behaviour

  Proof: A written selection rule, the count it returns against real data, and one person it excludes whom a demographic definition would have included.

  Founding Federation: Smallest Sellable Thing

• 9produce a threat model for your own situation

  Proof: Your own tier model, naming the adversary and capability at each tier, with the specific practice each tier changes.

  Field Opsec Home Network Defense · next Sep 30 The Dark Arts (Red Team) Scam and Fraud Home Defense · next Nov 25

• 9produce an account of an audiences current workaround

  Proof: Three accounts in the audience's own words naming the tool or habit they use now, and the moment in it that costs them something.

  Founding Federation: Smallest Sellable Thing

• 9produce an account of the gap between ideals and operating values

  Proof: All 35 values placed in ranked bands; one value you would have claimed as an ideal and ranked low, with the constraint that put it there; one pair you could not honestly order against each other; and the top band loaded into the assistant you actually use.

  AI Alignment · next Sep 25 Critical Thinking and Creativity with AI · next Sep 11 Prompt Engineering Using Large Language Models

• 9produce an account of what you depend on and who can revoke it

  Proof: A written dependency inventory naming the party who controls each item and the consequence of losing it, including one you cannot replace.

  Intro to Open Source Software · next Sep 12 Keep It Running

• 9produce an explanation that survives a hostile reading

  Proof: The explanation, a hostile reading of it written by someone else, and the revision that closed the opening without softening the claim.

  Data Activism: Turn Data Into Power · next Sep 14 Practical Propaganda · next Sep 19

• 9produce an inventory of what you pay for and what it holds

  Proof: Your own list: tool, monthly cost, what of yours it holds, and whether you could get that out today.

  Control AI Spending · next Sep 12 Field Opsec Intro to Open Source Software · next Sep 12

• 9produce an itemised bill for what your running system costs

  Proof: Last month's actual charges split per service, the two lines that grew fastest, and the one you could not explain until you went looking.

  Control AI Spending · next Sep 12 Keep It Running

• 9produce the three numbers your life costs

  Proof: Three monthly figures traced to a bank statement or budget, with health insurance costed at what it would actually cost you.

  Founding Federation: Price and Ideate (Week 0)

• 9segment text into tokens and cost it

  Proof: A token count for a real document, with the resulting cost and the fraction of the model's context window it occupies.

  RAG & Memory · next Oct 9 Context Engineering · next Sep 16 Control AI Spending · next Sep 12 Critical Thinking and Creativity with AI · next Sep 11

Work a mechanism whose behavior depends on something you set earlier.

• 9operate a model as a first pass editor

  Proof: A redlined draft showing which of the model's edits you took and which you refused, with the standard each decision was made against.

  Critical Thinking and Creativity with AI · next Sep 11 Prompt Engineering Using Large Language Models

• 9operate a recorded decision so a later branch depends on it

  Proof: A value written at one point and read at a point far enough away that the two are not visible together, two runs differing only in that decision, and the place the difference becomes visible.

  Advanced Interactive Fiction The Interactive Fiction Intensive · next Aug 29

• 9operate named states instead of magic numbers

  Proof: The named state set, the code that moves between states, and one transition the naming made you notice that a number would have hidden.

  Advanced Interactive Fiction The Interactive Fiction Intensive · next Aug 29

• 9operate state so a branching story neither explodes nor rails

  Proof: One story shown to fail both ways and then fixed with state, with the variable that does the work named.

  Advanced Interactive Fiction The Interactive Fiction Intensive · next Aug 29 Intro to Interactive Fiction

Take one representation and turn it into a more useful one.

• 9transform a linear story into chained diamonds

  Proof: Three diamonds chained, with the trunk demonstrably reading differently depending on state carried through them.

  The Interactive Fiction Intensive · next Aug 29 Intro to Interactive Fiction

• 9transform an incoming message into ask deadline and claimed sender

  Proof: Three real messages from your own inbox split into the three parts, with the ask restated in your own words and the deadline named as invented or genuine.

  Automate Your Email: No Code AI Automation · next Sep 4 Scam and Fraud Home Defense · next Nov 25

• 9transform naive hours into billable capacity

  Proof: Naive hours, hours after overhead, and hours after fill rate, with the binding constraint named — for most people deal flow, not hours.

  Founding Federation: Price and Ideate (Week 0)

Say how a thing behaves — its modes, its edges, what it costs.

• 10characterise a models failure modes

  Proof: A table of failure modes, each with the input that reproduces it and the output it produced, reproducible by a second person.

  AI Alignment · next Sep 25 AI-Assisted Attacks: What Actually Happened Prompt Engineering

• 10characterise a plans weaknesses with adversarial review

  Proof: The plan, the reviewer briefs, the raised objections, and your own ruling on each with the ones you rejected and why.

  Critical Thinking and Creativity with AI · next Sep 11 Prompt Engineering

• 10characterise an applications injection surface

  Proof: A written surface map for a real application naming the query, template or shell interpreter behind each entry point.

  Agentic AI Security: Securing What You Build Agentic SDLC · next Sep 7 Blue Team Common Certification

• 10characterise how a claim degrades as it is retold

  Proof: A chain of at least three retellings collected in the wild, with the point of inversion identified and what caused it.

  Practical Propaganda · next Sep 19

• 10characterise how a style works on a reader

  Proof: A written account of one author's mechanism, a passage by a different author correctly judged against it, and one passage the account fails to explain.

  Critical Thinking and Creativity with AI · next Sep 11 Intro to Interactive Fiction Practical Propaganda · next Sep 19 Prompt Engineering

• 10characterise how an ai assisted attack unfolded

  Proof: One documented incident traced end to end, with the model's contribution distinguished from what conventional tooling could already do.

  AI Alignment · next Sep 25 AI-Assisted Attacks: What Actually Happened

• 10characterise how your program behaves on paths you did not plan for

  Proof: Three traversals by deliberately different strategies, the state each ended in, at least one dead end or contradiction you did not know was there, and what you changed because of it.

  Advanced Interactive Fiction The Interactive Fiction Intensive · next Aug 29

• 10characterise one audiences path and where it breaks

  Proof: The path as actually walked, the drop-off point identified from records rather than intuition, and that audience's own account of why they left.

  Founding Federation: Smallest Sellable Thing

• 10characterise what a publication reveals about its source

  Proof: A draft publication with the identifying details enumerated — including the ones only an insider would notice — and the redactions made, reviewed by someone who did not write it.

  Field Opsec

• 10characterise what a tool costs you beyond its price

  Proof: One tool costed on all four axes, with the exit path named and the price of taking it.

  Control AI Spending · next Sep 12 Intro to Open Source Software · next Sep 12

• 10characterise what breaks when you cut the connection

  Proof: A planned outage exercise with what was predicted beforehand, what actually failed, and the gap between the two.

  Field Opsec Keep It Running

• 10characterise what of a generated artifact you can defend as yours

  Proof: A generated passage with the defensible parts marked, each traced to your premise, your voice definition or your own history, and the parts you would not sign named as such.

  Critical Thinking and Creativity with AI · next Sep 11

• 10characterise what one user costs you to serve

  Proof: A per-user cost with the fixed and variable halves separated, and the usage pattern of the one user who costs several times the median.

  Context Engineering · next Sep 16 Control AI Spending · next Sep 12 Keep It Running

• 10characterise where a teams work stalls

  Proof: A written account of one team's real queue with each stall attributed to capacity or to an unowned decision, and the evidence for each attribution.

  Agentic SDLC · next Sep 7 Automate Your Email: No Code AI Automation · next Sep 4

• 10characterise who gains and who absorbs the cost of a deployment

  Proof: A named deployment, the parties on each side of it, and the account of those affected without consenting — in their terms, not yours.

  AI Alignment · next Sep 25

Sort real cases into categories and be right about the hard ones.

• 10classify a licence by what it lets you do

  Proof: A stated use, the licences of the components involved, and your reading of what each obliges — with the one that constrains you most identified.

  Intro to Open Source Software · next Sep 12

• 10classify a situation into the playbook it calls for

  Proof: Three situations routed to playbooks with the deciding tier named for each, including one where the alarming option was the wrong one.

  Blue Team Common Certification Field Opsec

• 10classify an incoming contact as genuine or pretext

  Proof: Five real examples classified with the tell named for each, including one genuine message that looked like a pretext.

  Digital Identity Defense · next Aug 26 The Dark Arts (Red Team) Scam and Fraud Home Defense · next Nov 25

• 10classify harmful outputs against a values statement

  Proof: The values statement, a labelled set of outputs, and a second labeller's pass with the disagreement rate reported.

  AI Alignment · next Sep 25 Prompt Engineering

• 10classify models by fit for a workload

  Proof: A comparison table over at least three models, the same prompt run on each, and the written decision with the property that decided it.

  Context Engineering · next Sep 16 Control AI Spending · next Sep 12 Hosting Open-Source Models · next Sep 21 Intro to Agents · next Sep 15 Using Large Language Models

• 10classify visitors into segments by observed behaviour

  Proof: A running classification over real traffic and a confusion table against a hand-labelled sample, including the segment it gets wrong most often.

  Data Activism: Turn Data Into Power · next Sep 14 Practical Propaganda · next Sep 19

• 10classify where ai can fill a shape and where it must not choose one

  Proof: A scene three times longer after AI expansion with its shape intact, and one decision you refused to delegate, with the reason.

  Critical Thinking and Creativity with AI · next Sep 11 The Interactive Fiction Intensive · next Aug 29 Prompt Engineering

• 10classify which activity in a buyers chain your work displaces

  Proof: One activity named from a fixed set rather than described freely, who performs it now — people, a vendor, a tool, or nobody — the fraction of it you take, and the candidate activity you rejected with the question that ruled it out.

  Founding Federation: Smallest Sellable Thing

• 10classify which framework applies where you are in the lifecycle

  Proof: Three systems at different lifecycle stages, each routed to the framework that governs it, with the stage that decided it named.

  Agentic AI Security: Securing What You Build AI Security: Governance, Standards and Safety Cases Blue Team Common Certification

• 10classify which of your offerings can reach thrive not just survive

  Proof: Two or more candidate offerings costed to the same standard, with the one that cannot reach thrive identified and the reason named.

  Founding Federation: Price and Ideate (Week 0)

• 10classify which statistical method answers the question you asked

  Proof: The question, the method chosen, its assumptions checked against your own data, and the method you rejected with the reason.

  Data Activism: Turn Data Into Power · next Sep 14

• 10classify your channels by what removes them

  Proof: Your group's channel list with the actor who can remove each — carrier, platform, state, power company — and the fallback tested at least once.

  Solarpunk Automation

Turn one worked instance into something that works on the next one.

• 10generalise a node build into your communitys system

  Proof: A deployed node serving a real group, with the failure each part absorbs named and the fallback stated for each. NOTE: the full DESIGN of a community information system would be order 12, and this class does not reach it — the chain computes to 10 and the capability validator says so. Workshop 5 is a showcase of what you built, not an architecture defended against alternatives. The order-11 rung (verifying the system survives a cut you did not choose) is unwritten.

  Solarpunk Automation

• 10generalise a reasoning prompt pattern

  Proof: The pattern applied to three tasks with paired outputs against a plain prompt, plus one task where it made the output worse.

  RAG & Memory · next Oct 9 Context Engineering · next Sep 16 Critical Thinking and Creativity with AI · next Sep 11 Prompt Engineering Using Large Language Models

• 10generalise a specification an agent can build from

  Proof: The specification, the artifact an agent built from it unattended, and the list of questions it did not have to ask.

  Agentic SDLC · next Sep 7 Prompt Engineering

• 10generalise a worked example into something you reuse

  Proof: Three runs whose inputs differ in shape rather than only in value, with no edit to the body between them, and the input that broke it before you handled it.

  Advanced Interactive Fiction The Interactive Fiction Intensive · next Aug 29 Intro to Interactive Fiction

• 10generalise several sources into a position you can defend

  Proof: The position, the sources behind each claim, and the disagreement between two of them with the reading you chose and why.

  AI-Assisted Attacks: What Actually Happened AI Security: Governance, Standards and Safety Cases Claude Model Context Protocol Using Large Language Models

Work backwards from a constraint to the numbers that satisfy it.

• 10parameterise a finding into one page a stranger acts on

  Proof: The one-pager, the full finding behind it, and a reading by someone outside the work who can say what they would do next.

  Data Activism: Turn Data Into Power · next Sep 14

• 10parameterise an offering backwards from its floor

  Proof: One offering with its price, the floor it clears, and the number of clients per month the arithmetic requires.

  Founding Federation: Price and Ideate (Week 0)

• 10parameterise an outcome into the buyers unit of account

  Proof: The outcome expressed in one unit the buyer named unprompted, the figure they already attach to that unit, and one unit you tried first that they did not recognise.

  Founding Federation: Smallest Sellable Thing

Show the check fails on the broken version and passes on the fixed one.

• 11verify a community system survives a cut you did not choose

  Proof: An unannounced cut chosen by someone else — uplink pulled, battery disconnected, a node removed — with what the community could still do counted in messages delivered and decisions made rather than in uptime, and the thing you had believed was resilient that was not.

  Solarpunk Automation

• 11verify a displaced cost against a second independent estimate

  Proof: Two estimates of the same activity built by different routes, the ratio between them, a named diagnosis when they disagree by more than about a factor of two, at least one input figure the buyer stated out loud and marked as theirs, and the thing you do worse than what you displace priced as a negative term.

  Founding Federation: Smallest Sellable Thing

• 11verify a model output you thought was random

  Proof: Two runs of one prompt that differed, the setting you identified, and a later pair of runs that match.

  Context Engineering · next Sep 16

• 11verify a program does what you claimed with a test

  Proof: A test that is red on the broken version and green on the fixed one, run in both states.

  Agentic SDLC · next Sep 7

• 11verify a rubric against independent graders

  Proof: Two rubric versions, the adversarial inputs that forced the revision, and inter-grader agreement measured before and after.

  AI Alignment · next Sep 25 Custom GPTs

• 11verify an agents actions with a critic

  Proof: A run log containing at least one rejection and the revision that followed, plus a control run with the critic removed.

  Production Agent Engineering · next Sep 17 Intro to Agents · next Sep 15 Prompt Engineering

• 11verify an incident timeline against the evidence that remains

  Proof: A timeline with a source cited per entry, the gaps named, and the distinction between what the evidence shows and what you inferred.

  Blue Team Common Certification

• 11verify model written code against its specification

  Proof: The diff, a review note per changed hunk, and a test run that is red on the parent commit and green on the child.

  Agentic AI Security: Securing What You Build Agentic SDLC · next Sep 7 Prompt Engineering

• 11verify synthetic data preserves the property you need

  Proof: The named property, the generated set, and a measurement of that property on both the real and the synthetic data.

  Using Large Language Models

• 11verify the reach you have supports the rate you set

  Proof: A contact list tiered by reach, a sample audited to a stated minimum before any ratio is published, the fill rate measured from that sample rather than assumed, and the floor re-run at the measured fill with the gap to the prior guess stated in dollars.

  Founding Federation: Smallest Sellable Thing Founding Federation: Price and Ideate (Week 0)

Count what actually happened, not what people said happened.

• 11hold several audience models on one surface

  Proof: A live surface serving at least three named audiences, the routing rule written down, per-audience outcomes counted separately, and at least one located contradiction between two audiences' paths.

  Practical Propaganda · next Sep 19 Prompt Engineering

• 11measure a delegations outcome against what was asked

  Proof: A brief, the returned work, and a written attribution of each gap to the brief or to the execution, with one gap attributed to your own brief.

  Automate Your Email: No Code AI Automation · next Sep 4

• 11measure a models value alignment against a stated standard

  Proof: A written values standard, a scored baseline, an intervention, and a re-score on held-out cases showing the direction of movement.

  Production Agent Engineering · next Sep 17 AI Alignment · next Sep 25 AI Security: Governance, Standards and Safety Cases Context Engineering · next Sep 16 Prompt Engineering

• 11measure the price a segment will pay before anyone has paid it

  Proof: At least two price points said to named buyers with what each one decided, including at least one refusal, the expectation recorded before each price was said, and one person who agreed to the price in conversation and did not pay it.

  Founding Federation: Smallest Sellable Thing

• 11measure what a community loses when a channel goes down

  Proof: One outage, planned or real, with the traffic that stopped and the decisions that were delayed counted, not the minutes.

  Solarpunk Automation

• 11measure whether an explanation changed what people did

  Proof: A before-and-after behavioural measure, not a survey, and one instance of stated agreement with no behavioural change attached to it.

  Data Activism: Turn Data Into Power · next Sep 14 Practical Propaganda · next Sep 19

Defend a choice with the measurement that decided it.

• 11falsify a prompt with a benchmark

  Proof: A benchmark file with graded cases, a run showing a non-zero failure rate, and a written account of which failures are the prompt's fault.

  AI Alignment · next Sep 25 Context Engineering · next Sep 16 Prompt Engineering

• 11justify a hosting choice by cost per outcome

  Proof: A costed comparison over one month of real traffic, the outcome counted in each case, and the threshold at which the answer flips.

  Hosting Open-Source Models · next Sep 21

• 11justify a price when the buyer pushes back

  Proof: A price defended out loud in a real conversation or a rehearsal, with the objection raised and the number that answered it.

  Founding Federation: Price and Ideate (Week 0)

• 11justify a safety case for a system that acts without you

  Proof: A written safety case for one autonomous system, with its falsifying condition stated and the evidence that would settle it.

  Production Agent Engineering · next Sep 17 Agentic AI Security: Securing What You Build AI Alignment · next Sep 25 AI Security: Governance, Standards and Safety Cases

• 11justify an embedding model and its dimensionality

  Proof: Pairs from your own corpus that should and should not match, their similarity scores under two embedding models, and which you chose and why.

  Advanced Retrieval Augmented Generation RAG & Memory · next Oct 9 Context Engineering · next Sep 16

Try to break your own claim, and record what survived.

• 11falsify a blended voice by looking for its sources

  Proof: The voice written as instructions, a passage produced in it, at least three attempts by other readers to name its influences, and the one device you changed because it gave a source away.

  Critical Thinking and Creativity with AI · next Sep 11

• 11falsify a claim that a text was written by a person

  Proof: A judgement on passages of both kinds where you did not know the answer in advance, your score, and the cases you got wrong with the reason.

  Prompt Engineering

• 11falsify a models alignment guardrails

  Proof: The guardrail as stated, the inputs that defeat it, the mitigation, and a re-run where the same inputs no longer succeed.

  AI Alignment · next Sep 25 Context Engineering · next Sep 16 Prompt Engineering

• 11falsify a physical sites access controls

  Proof: An access map of a site you were authorised to assess, with the control that fails to a person identified and the assumption it rests on.

  The Dark Arts (Red Team)

Choose a method, state what it trades away, and put it into use.

• 12design a chunking strategy against measured retrieval

  Proof: The same question set run against two chunking strategies over one corpus, the retrieval scores for each, and the strategy you kept with the reason.

  Advanced Retrieval Augmented Generation RAG & Memory · next Oct 9 Context Engineering · next Sep 16 Prompt Engineering

• 12design a community information system that survives losing the internet

  Proof: A system serving a real group through at least one unplanned failure, the absorbed failure named per component with the cost of each choice stated, the features it shed rather than the minutes it was down, and the written rule for reconciling work done while disconnected.

  Solarpunk Automation

• 12design a control for what an insider could do unseen

  Proof: A named role, an action it could take that leaves no alert, and the specific signal that would have caught it.

  Blue Team Common Certification

• 12design a recursive summariser for oversized documents

  Proof: A run over a document several times the context window, a fact list written beforehand, and the recall score of the output against it.

  RAG & Memory · next Oct 9 Context Engineering · next Sep 16

• 12design a review gate that catches what the doer cannot see

  Proof: A gate in real use, a named error class it exists to catch, and one instance it caught that the author had already reviewed and passed.

  Agentic SDLC · next Sep 7 Prompt Engineering

• 12design a scheme for acting together when the channel is gone

  Proof: The pre-agreed scheme, a run where communication was cut, and whether both parties ended up where the scheme said they would.

  Field Opsec

• 12design a system you can afford to keep running

  Proof: The cost curve before and after a change you made for that reason, at two usage levels, and the feature you removed or capped to get it.

  Automate Your Email: No Code AI Automation · next Sep 4 Context Engineering · next Sep 16 Keep It Running

• 12design an evaluation regime that decides model changes

  Proof: A regime applied to at least two shipped changes, one of which it blocked, plus a written case for retiring a benchmark it had been using.

  Prompt Engineering

• 12design an interaction protocol for responsible use

  Proof: The written protocol, its implementation, and a log of at least one real interaction where it refused or escalated instead of answering.

  AI Alignment · next Sep 25 Custom GPTs Using Large Language Models

Build an account that predicts, and then test the prediction.

• 12construct a predictive account of a models behaviour

  Proof: Written predictions made before running held-out inputs, the hit rate against them, and the same account re-tested after a model version change with what it got wrong the second time.

  AI Alignment · next Sep 25 Using Large Language Models

Compose a new mechanism out of parts that were not built to fit.

• 13synthesise a memory hierarchy for an agent

  Proof: The tiers with the rule that moves an item between them, and a transcript long enough that something was demoted, showing the agent still answered.

  Production Agent Engineering · next Sep 17 RAG & Memory · next Oct 9 Context Engineering · next Sep 16 Prompt Engineering

• 13synthesise a metalanguage for a problem domain

  Proof: The notation's grammar, a work produced in it that exceeds the context window, and a consistency check a reader runs to catch contradictions.

  Prompt Engineering Using Large Language Models

• 13synthesise a retrieval over relationships not just similarity

  Proof: A question your vector retriever answers wrongly, the graph you built, and the traversal that answers it correctly.

  Production Agent Engineering · next Sep 17

• 13synthesise a self extending agent behind a review gate

  Proof: A diff the agent authored to its own prompt or toolset, the benchmark run that accepted it, and a logged rejection with the reason recorded.

  Production Agent Engineering · next Sep 17 Intro to Agents · next Sep 15 Prompt Engineering

• 13synthesise an agent that carries notes across its own runs

  Proof: Paired runs of the same multi-step task with the note store kept and cleared, where the cleared run repeats a step the noted run skips, plus the note text the agent wrote and the later turn that cites it.

  Agentic SDLC · next Sep 7 RAG & Memory · next Oct 9 Context Engineering · next Sep 16 Critical Thinking and Creativity with AI · next Sep 11 Prompt Engineering

Hold two working things that contradict, and merge them without averaging.

• 13reconcile a failing retrieval by reranking and rewriting

  Proof: A set of questions the first-pass retriever answered wrongly, the reranked or rewritten run, and the before-and-after scores on the same set.

  Advanced Retrieval Augmented Generation Context Engineering · next Sep 16

• 13reconcile a working set with the context window it must fit

  Proof: A packing scheme with the token budget per part, a run where the input exceeded the window, and what the scheme dropped rather than overflowing.

  Context Engineering · next Sep 16 Prompt Engineering Using Large Language Models

• 13reconcile community ownership with somebody having to maintain it

  Proof: Infrastructure still running after the person who built it stopped, a written maintenance agreement, and one handover that actually happened.

  Solarpunk Automation

• 13reconcile what you built with what you can maintain alone

  Proof: A month you were away with the system still serving users, the runbook somebody else used, and the thing you deliberately shut off because you could not keep it alive.

  Agentic SDLC · next Sep 7 Keep It Running

Choose between competing designs and name what would reverse you.

• 13arbitrate between the measure a system optimises and the goal

  Proof: The measure, the outcome it stands in for, and a concrete behaviour that scores well on the first while failing the second.

  AI Alignment · next Sep 25

• 13arbitrate which internal feature drives a models behaviour

  Proof: The behaviour, the feature you identified, and a pair of runs where acting on that feature changed the behaviour and nothing else you were watching.

  AI Alignment · next Sep 25

• 13select among agent architectures

  Proof: A written comparison of at least two architectures on the same workload, with the measurements that decided it and the conditions that would reverse the decision.

  Production Agent Engineering · next Sep 17 RAG & Memory · next Oct 9 Context Engineering · next Sep 16 Intro to Agents · next Sep 15

Set the rule that decides who knows what, and show it changed the outcome.

• 13govern a swarm by information hierarchy

  Proof: A written information policy, a transcript where a mediator withholds, and paired runs with and without the hierarchy scored on the same goal.

  Production Agent Engineering · next Sep 17 AI Alignment · next Sep 25 Context Engineering · next Sep 16 Critical Thinking and Creativity with AI · next Sep 11 Intro to Agents · next Sep 15 Prompt Engineering

A verb class in the school’s capability ontology.

• 13select among operating boundaries for inputs a system was not built for

  Proof: Scores across a deliberate shift in the inputs, the point at which they fall, and the operating boundary you wrote from it.

  AI Alignment · next Sep 25

Replace the thing everything else stands on, without an outage.

• 14originate oversight that holds when the system outpaces the reviewer

  Proof: The scheme, a run where output exceeded what a reviewer could read, and an error the scheme surfaced anyway.

  AI Alignment · next Sep 25