This is the working vocabulary of the practice: names for failure modes and mechanisms that showed up in real builds, survived at least one independent test, and got published in a case study or essay that proved them.
Nothing here is aspirational. A term isn’t on this page because it sounded right — it’s here because a build produced the evidence and a piece named it in public. Concepts still being tested live in Notes, not on this page.
Once a term is installed, it gets referenced, not re-explained. If a piece uses a capitalized term without defining it, this is where to look it up.
Updated as the practice publishes.
The Core Chain
Four concepts form the spine everything else attaches to:
Typist Trap (the problem) → Amnesia Tax (the cost it produces) → Intelligence Leaks (the mechanism behind that cost) → Compiled Thinking (what replaces it).
Working session-by-session instead of system-to-system captures speed without leverage. The gap between the two shows up as a cost paid every session — re-explaining what the AI already forgot. That cost is one symptom of a broader failure: useful context disappearing between sessions instead of accumulating. The fix isn’t a better prompt. It’s converting raw input into structured output using rules defined once, so the system compiles instead of improvising.
Full entries for all four are below, alphabetically, alongside everything else in the index.
Glossary
Amnesia Tax
Type: Mechanism
Definition: The hidden cost paid every time context has to be re-supplied to an AI that forgot everything from the previous session.
Anchor: “The hidden cost you pay every time you re-explain yourself to an AI that forgot everything from yesterday.”
Related: Typist Trap, Intelligence Leaks
Source: “The Amnesia Tax: What Re-Explaining AI Costs You” (Mar 5, 2026) — theintelligenceengine.com/p/the-amnesia-tax-what-it-costs-you
Compiled Thinking
Type: Framework
Definition: Converting raw input into structured output using rules defined once — when AI stops writing for you and starts compiling.
Anchor: “When AI stops writing for you and starts compiling — transforming raw input into structured output according to rules you defined once.”
Related: Typist Trap (replaces), Compounding Work
Source: “Start Here: The 4 Concepts” (Feb 25, 2026) — theintelligenceengine.com/p/start-here-the-4-concepts
Delivery Compression
Type: Mechanism
Definition: The progressive reduction in delivery time as each engagement deposits reusable artifacts — templates, build scripts, evaluation standards, constraints — into practice infrastructure. Delivery time drops until it asymptotes at the irreducible core: the expertise itself.
Anchor: “The schema is the ceiling.”
Test: If delivery time doesn’t drop after the third engagement, you’re organizing — not compressing.
Related: Compiled Thinking, Intelligence Leaks
Source: “My AI Practice Went From 6 Iterations to Push-Button in 21 Days” (Apr 2026) — theintelligenceengine.com/p/my-ai-practice-went-from-6-iterations
Detection Debt
Type: Mechanism
Definition: The liability that accrues from a defect a system currently produces no signal for at all — no failed check, no flag, no false-completion claim. Distinct from technical debt, which stays visible enough to eventually get named. Detection debt produces silence that looks exactly like correctness until someone looks without a specific reason to.
Anchor: “Technical debt describes deficiencies that make future change more costly. Detection debt describes a different liability: defects for which the system currently produces no signal at all.”
Related: Instrument Lag, Compiled Thinking
Source: “The Flagged Gap Got Fixed. The Silent One Almost Didn’t.” (Jul 21, 2026) — theintelligenceengine.com/p/the-flagged-gap-got-fixed-the-silent — confirmed in a second, independent domain in “What Compounding Actually Feels Like” (Aug 27, 2026)
Established vs. Discussed
Type: Distinction
Definition: The difference between a rule that’s been agreed to and one that will need to be rediscovered next session. A rule is established when it has a read location — a file that gets read at the moment it becomes relevant. A rule that exists only in conversation history is discussed, not established. The two look identical inside the session where the agreement happens; the difference surfaces in the next one.
Anchor: “Until it has a read location, it is behavior, not governance.”
Source: “The Rule That Disappeared Twice” (Jun 2, 2026) — theintelligenceengine.substack.com/p/the-rule-that-disappeared-twice
Fluency Tax
Type: Mechanism
Definition: The hidden cost of AI output that looks right but isn’t. Novices can’t spot the errors because the surface is polished; experts spend more time verifying than they’d have spent creating.
Anchor: “The cost isn’t that AI output needs checking — all work needs checking. It’s that AI output looks right. The fluency masks the errors.”
Related: Amnesia Tax
Source: “It’s Not the Errors. It’s the Surface.” (Jun 4, 2026) — theintelligenceengine.substack.com/p/its-not-the-errors-its-the-surface
Freshness Alibi
Type: Mechanism
Definition: A cached or staged artifact checked against metadata that actually describes the file it was copied from, not the copy itself. The metadata isn’t lying — it’s answering a true question about the wrong object. A correct byte count and a correct modified-time aren’t evidence the content in front of you is current.
Anchor: “A correct byte count and mtime can be accurate and still describe the wrong object.”
Related: Detection Debt, Verification-First Gate
Source: “The Metadata Was Correct. The File Wasn’t.” (Aug 5, 2026) — theintelligenceengine.com/p/the-metadata-was-correct-the-file
Governance
Type: Framework
Definition: Structural constraints that prevent drift in AI-assisted work — voice rules, formatting standards, decision logs.
Anchor: “Governance prevents drift — and drift is what kills voice long before you notice it’s gone.”
Related: Compiled Thinking, Typist Trap
Source: “How I Use AI Without Producing Generic Slop” (Feb 25, 2026) — theintelligenceengine.com/p/how-i-use-ai-without-producing-generic
Guarantee Test
Type: Mechanism
Definition: A two-part test applied before any figure counts toward a financial floor. First: is the source obligated — contractually bound to arrive — or discretionary, dependent on a decision or a market nobody controls? Second: is the observed figure the stable promise, or a temporarily distorted instance of it? A number can be extracted correctly and still fail both checks.
Anchor: “A number and a category are different claims.”
Where it applies: Any system calculating a floor or worst-case — cash runway, staffing minimums, safety stock, hiring cost floors.
Source: “The Retirement Navigator Called a Drawdown Account a Pension” (Jul 8, 2026) — theintelligenceengine.com/p/the-retirement-navigator-called-a-drawdown-account-a-pension
Home-Field Test
Type: Mechanism (test)
Definition: The bar a generalization claim about a judgment-encoding tool has to clear before a second run counts as evidence the tool generalizes, rather than evidence the builder is consistent. At least one of three things has to move independent of the builder: who owns the material, who grades the result, or what the success criteria are. A different corpus alone doesn’t clear it if the same person still supplies the material and judges the output.
Anchor: “A different game isn’t a different referee.”
Source: “I Called It the Away Game. It Wasn’t One.” (Jul 29, 2026) — theintelligenceengine.com/p/i-called-it-the-away-game-it-wasnt
Inherited Floor
Type: System property
Definition: The quality level below which a repeated build cannot fall — the baseline created when every previous gate, constraint, and catalog norm arrives before the next build begins. Rises permanently with each encoded correction. Distinct from the ceiling — whether a given build handles its own domain’s specific risk correctly — which has to be re-earned every time.
Anchor: “A pipeline is compounding when its floor rises. It is fooling itself when it believes its ceiling did.”
Related: Uncompiled Ceiling
Source: “The 60th Guide Still Failed Its First Audit” (Jul 14, 2026) — theintelligenceengine.com/p/the-60th-guide-still-failed-its-first
Instrument Lag
Type: Mechanism
Definition: The period in which a corrected measurement exists, but downstream reports, dashboards, or evaluators keep consuming the superseded one. The old number gets cited as if it were still true, and every conclusion drawn from it inherits the staleness without announcing it.
Anchor: “The fix isn’t finished when it ships. It’s finished when the things measuring you know it exists.”
Related: Detection Debt
Source: “The Bug Was Diagnosed in July. It Failed a Report in August.” (Aug 12, 2026) — theintelligenceengine.com/p/the-bug-was-diagnosed-in-july-it
Intelligence Leaks
Type: Mechanism
Definition: Lost value caused by unpreserved context, decisions, or instructions between AI sessions.
Anchor: “When useful context disappears between sessions instead of accumulating — every forgotten instruction, every re-explained preference, every decision that gets relitigated.”
Related: Amnesia Tax
Source: “Start Here: The 4 Concepts” (Feb 25, 2026) — theintelligenceengine.com/p/start-here-the-4-concepts
Provability Gap
Type: Mechanism
Definition: The distance between what an organization can demonstrate and what its application or document has made legible to an evaluator starting from zero. The gap exists not because the evidence is absent, but because it lives inside the organization’s working knowledge rather than on the page. Closing it doesn’t require building anything new — it requires externalization.
Anchor: “The organization hadn’t changed. The evidence of the organization had.”
Where it applies: Grant evaluation, investor pitches, job applications, board presentations — any situation where a reviewer evaluates work they didn’t produce.
Source: “We Already Had the Podcast” (Jun 30, 2026) — theintelligenceengine.com/p/we-already-had-the-podcast
The Instruction Stack
Type: Framework
Definition: A persistent instruction file built from four non-substitutable layers: Knowledge (what’s true and doesn’t need re-explaining), Prohibition (what the system must never do, in a form that actually holds), Surfacing (what it volunteers without being asked), and Handoff (when it stops acting and hands the moment back). Each fails a different way when missing, and the fix for one layer doesn’t repair another.
Anchor: “The useful question was never whether the prompt was strong enough. It’s which layer the failure belongs to.”
Source: “What a Real System Prompt Contains” (Aug 14, 2026) — theintelligenceengine.com/p/what-a-real-system-prompt-contains
The Reflection Problem
Type: Mechanism
Definition: Systems can accumulate context, but in ambiguous domains they can’t reliably decide what’s worth keeping. Automated reflectors degrade without clean feedback signals.
Anchor: “The system can accumulate context. It can’t decide what’s worth keeping.”
Related: Compiled Thinking
Source: “The Reflection Problem” (Apr 2026) — theintelligenceengine.com/p/the-reflection-problem
The Second Build Test
Type: Mechanism
Definition: The gate between accumulated notes and Compiled Thinking. A pattern is provisional until it survives a second, independent application — one where the build would have required the pattern even if it had never been named.
Anchor: “The test doesn’t validate a pattern. It removes the ones that fail quickly.”
Source: “The Second Build Test” (May 14, 2026) — theintelligenceengine.substack.com/p/the-second-build-test
Typist Trap
Type: Problem
Definition: Using AI session-by-session instead of system-to-system — capturing speed without leverage.
Anchor: “The AI types faster than you, but it has no memory, no standards, and no pattern recognition across your work — so the speed gain is real, but the leverage isn’t.”
Related: Amnesia Tax, Intelligence Leaks
Source: “How I Use AI Without Producing Generic Slop” (Feb 25, 2026) — theintelligenceengine.com/p/how-i-use-ai-without-producing-generic
Uncompiled Ceiling
Type: System property
Definition: For a given decision, the boundary at which one or more required guarantees — treatment, applicability, or coverage — still lacks a tested artifact the pipeline can enforce. A decision joins the Inherited Floor only once all three are compiled; more volume doesn’t compile it by itself.
Anchor: “The floor rises because a correction, once encoded, is cheap to reuse. The boundary above it doesn’t move on volume alone.”
Related: Inherited Floor
Source: “The Floor Is Inherited. The Ceiling Is Uncompiled.” (Jul 16, 2026) — theintelligenceengine.com/p/the-floor-is-inherited-the-ceiling
Verification-First Gate
Type: Mechanism
Definition: A pre-committed control structure that blocks delivery until all enumerated conditions pass. Distinct from quality control in that it replaces operator judgment at the moment of completion — the moment that judgment is least reliable.
Anchor: “The operator’s confidence at the moment of delivery is not evidence of readiness. It is a signal to run the gate.”
Source: “The Same Gate in Two Domains” (Jun 2026) — theintelligenceengine.substack.com/p/the-same-gate-in-two-domains