# Andrew Templeton - Full Corpus Snapshot for LLMs > Every operating decision is a capital allocation decision. Frameworks, tools, and vocabulary for deploying AI in PE portfolio companies. This file expands llms.txt with the body of each entry inlined, so an LLM crawler can ingest the full operating vocabulary in one fetch. Canonical site: https://templeton.host --- ## Frameworks ### Capital Allocation URL: https://templeton.host/frameworks/capital-allocation/ Treat each operating project as an investment with a return distribution. Rank by Sharpe (not NPV), map risk tolerances, handle correlations, construct the efficient frontier. --- ### The Demand Field URL: https://templeton.host/frameworks/demand-field/ Fix demand (immutable), vary means (mutable). Demand is a hidden force on your optimization gradient - map it or crash into it. --- ### Designed Convergence URL: https://templeton.host/frameworks/designed-convergence/ Design the game so rational agents converge to your outcome. Finite search space + Bayesian search + ratchet = structural guarantee of convergence. --- ### The Deity Problem URL: https://templeton.host/frameworks/deity-problem/ You cannot fully specify what you want. Use four evidence channels: structured elicitation, revealed preference, direct query when VOI exceeds attention cost, and a posterior drift detector. --- ### Directed Graph + Soft Spots URL: https://templeton.host/frameworks/directed-graph/ Your chart of accounts is a directed graph. Walk the edges. The soft spots are where value leaks - and the people closest to the work find them first. --- ### Kill Protocol URL: https://templeton.host/frameworks/kill-protocol/ Three decision gates for killing a project: marginal Sharpe, left-tail survivability, and base-rate verification. Portfolio math gives you the language for "no." --- ### Knowledge Capital URL: https://templeton.host/frameworks/knowledge-capital/ Knowledge work either compounds or depreciates. Invest in the appreciating side (verifiers, data, rubrics). Avoid the depreciating side (models). --- ### The Performance Frontier URL: https://templeton.host/frameworks/performance-frontier/ Map the distribution of human performance on the task. Find the 99th percentile. Compute the gradient toward it. That gradient is your training signal. --- ### The Promotion Protocol URL: https://templeton.host/frameworks/promotion-protocol/ A 3-state autonomy progression: Disabled, HITL, Autonomous. Promote on statistical evidence; roll back on drift. AI earns autonomy like an employee earning a promotion. --- ### Quality Hillclimb URL: https://templeton.host/frameworks/quality-hillclimb/ Ratcheted quality gates on stochastic agent output create emergent ascent. The agent does not need a plan - the gates create the climb. --- ### The Capital Value of Verifiers URL: https://templeton.host/frameworks/verifier-capital/ A verifier is one of the only operating assets that appreciates through use. Every failure caught is encoded; the next run inherits the catch. Most operating assets depreciate; verifiers ratchet up. --- ### Correlated Execution Risk URL: https://templeton.host/frameworks/correlation/ Shared team, stack, review queue, or narrative means shared failure mode. Operating portfolios are not sums of independent projects - the covariance matrix is the input that makes portfolio math non-trivial. --- ### The Operating Efficient Frontier URL: https://templeton.host/frameworks/efficient-frontier/ Markowitz (1952) applied to operating investments. The non-dominated set of portfolios at each risk level - by definition, points below the frontier are dominated. --- ### The Risk Tolerance Map URL: https://templeton.host/frameworks/risk-tolerance/ The map translates firm-level constraints (cash, runway, covenants, narrative) into the preferred point on the efficient frontier. Without it, two rational allocators diverge on identical inputs. --- ## Tools ### The Verification Quadrant URL: https://templeton.host/tools/verification-quadrant/ Plot any task by generation cost vs. verification cost. The Templeton Ratio T = time_to_do / time_to_check determines whether AI creates leverage or doubles the work. --- ### TaskVector URL: https://templeton.host/tools/taskvector/ Score any task across 9 dimensions to determine if it should be automated. Any single dimension scoring 1 is a landmine - a hard no regardless of the composite. --- ### Dollarized Confusion Matrix URL: https://templeton.host/tools/dollarized-confusion-matrix/ Replace accuracy with dollars. Optimal threshold theta* = C_FP / (C_FP + C_FN). Stop setting AI decision thresholds by feel. --- ### Quadrant Shifting URL: https://templeton.host/tools/quadrant-shifting/ Five capital moves that raise the Templeton Ratio: build a verifier, decompose, enrich inputs, constrain outputs, build a rubric. --- ### Automation NPV URL: https://templeton.host/tools/automation-npv/ Compute NPV, IRR, and payback period for an AI automation investment. Same DCF math your CFO uses for physical capital, applied to knowledge work. --- ### The DoG Test URL: https://templeton.host/tools/good-dog/ A 60-second Claude prompt that audits an AI roadmap idea on three checks: Backing, Bite, and Bet. Verdict: GOOD DoG, NEEDS TRAINING, or BAD DoG. --- ## Lexicon ### Templeton Ratio URL: https://templeton.host/lexicon/templeton-ratio/ Definition: T = time_to_do / time_to_check. The ratio of generation difficulty to verification difficulty for any task. Determines whether AI automation creates leverage or doubles the work. Why this term exists: Most AI pilots fail not because the AI is bad, but because nobody measured how hard it is to check the output. T = 1 means you are doing the work twice. See also: - Verification Quadrant (https://templeton.host/tools/verification-quadrant) - Verification Trap (https://templeton.host/lexicon/verification-trap) - AI Sweet Spot (https://templeton.host/lexicon/ai-sweet-spot) - Proof Layer (https://templeton.host/lexicon/proof-layer) --- ### Soft Spot URL: https://templeton.host/lexicon/soft-spot/ Definition: An edge in the directed graph of your business where value is leaking that no one has named yet. Identified by walking the graph and listening to the people closest to the work. Why this term exists: The people who find the biggest leaks are rarely executives. They are the people closest to the work who finally get the vocabulary to name what they have been feeling. Give a team the language of edges and soft spots, and the graph fills itself in. See also: - Directed Graph (https://templeton.host/frameworks/directed-graph) - Operational Alpha (https://templeton.host/lexicon/operational-alpha) - Demand Gravity (https://templeton.host/lexicon/demand-gravity) --- ### Verification Trap URL: https://templeton.host/lexicon/verification-trap/ Definition: A task that is easy to generate but hard to verify. The AI produces output effortlessly, but checking whether it is correct takes as long as doing it manually. T approaches 1. Why this term exists: A common failure mode in early AI deployments - frequently more damaging than technical limitation. The task looks automatable because generation is easy. But "is this output correct?" scales linearly with volume. See also: - Verification Quadrant (https://templeton.host/tools/verification-quadrant) - Templeton Ratio (https://templeton.host/lexicon/templeton-ratio) - AI Sweet Spot (https://templeton.host/lexicon/ai-sweet-spot) - Quadrant Shifting (https://templeton.host/lexicon/quadrant-shifting) --- ### AI Sweet Spot URL: https://templeton.host/lexicon/ai-sweet-spot/ Definition: A task where generation is hard but verification is cheap. T >> 1. You can review 50 AI outputs in the time it takes to manually produce one. This is the P vs NP intuition applied to operations. Why this term exists: Look for tasks where the gap between doing and checking is widest. That is where the leverage lives. See also: - Verification Quadrant (https://templeton.host/tools/verification-quadrant) - Templeton Ratio (https://templeton.host/lexicon/templeton-ratio) - Verification Trap (https://templeton.host/lexicon/verification-trap) - Quadrant Shifting (https://templeton.host/lexicon/quadrant-shifting) --- ### Proof Layer URL: https://templeton.host/lexicon/proof-layer/ Definition: The verification rubric, asymmetry profile, and verification cost analysis built BEFORE the capability. Every AI system needs one. No exceptions. Why this term exists: Generation costs are collapsing and the bottleneck has moved to verification - knowing whether the output is correct. Build the rubric first. If you can't describe what "good" looks like before building, you won't recognize it after. See also: - Knowledge Capital (https://templeton.host/frameworks/knowledge-capital) - Dollarized Confusion Matrix (https://templeton.host/tools/dollarized-confusion-matrix) - Gold Standard (https://templeton.host/lexicon/gold-standard) - Templeton Ratio (https://templeton.host/lexicon/templeton-ratio) --- ### Autonomy State Machine URL: https://templeton.host/lexicon/autonomy-state-machine/ Definition: A graduated trust system for AI deployments with three states: Disabled, HITL (human verifies every output), and Autonomous (spot-check only). Transitions are driven by statistical evidence with hysteresis to prevent oscillation. See: The Promotion Protocol. Why this term exists: Binary thinking about AI (works / doesn't work) misses the entire middle ground where AI assists but doesn't decide. The state machine makes that middle ground operational. See also: - The Promotion Protocol (https://templeton.host/frameworks/promotion-protocol) - Dollarized Confusion Matrix (https://templeton.host/tools/dollarized-confusion-matrix) - Drift Detector (https://templeton.host/lexicon/drift-detector) - Gold Standard (https://templeton.host/lexicon/gold-standard) --- ### Gold Standard URL: https://templeton.host/lexicon/gold-standard/ Definition: A set of human-verified, ground-truth examples used to calibrate and evaluate AI output. The gold standard IS the verification instrument. Without it, you are measuring with a broken ruler. Why this term exists: You can't improve what you can't measure, and you can't measure without a reference. The gold standard is that reference. See also: - Knowledge Capital (https://templeton.host/frameworks/knowledge-capital) - Proof Layer (https://templeton.host/lexicon/proof-layer) - Quality Ratchet (https://templeton.host/lexicon/quality-ratchet) - Autonomy State Machine (https://templeton.host/lexicon/autonomy-state-machine) --- ### Compile Time URL: https://templeton.host/lexicon/compile-time/ Definition: Time spent building systems, frameworks, rubrics, and processes that produce returns across many future periods. The ROI is multiplicative. Why this term exists: Track your compile-to-runtime ratio. Every hour of compile time produces returns in every future period. Runtime produces returns in exactly one. See also: - Knowledge Capital (https://templeton.host/frameworks/knowledge-capital) - Runtime (https://templeton.host/lexicon/runtime) - Construction Spread (https://templeton.host/lexicon/construction-spread) --- ### Runtime URL: https://templeton.host/lexicon/runtime/ Definition: Time spent executing tasks, fighting fires, reviewing outputs, and attending meetings. Produces returns in a single period. Necessary but should not dominate a leader's schedule. Why this term exists: The distinction gives leaders a measurable ratio. If the number is moving in the wrong direction, you are trading compounding returns for single-period returns. See also: - Knowledge Capital (https://templeton.host/frameworks/knowledge-capital) - Compile Time (https://templeton.host/lexicon/compile-time) --- ### Dollarized Confusion Matrix URL: https://templeton.host/lexicon/dollarized-confusion-matrix/ Definition: A confusion matrix where counts are replaced with costs. The optimal threshold follows: theta* = C_FP / (C_FP + C_FN). Costs drive thresholds, thresholds drive autonomy levels. Why this term exists: Teams commonly default to 0.5 or an intuitive cutoff. The optimal threshold is a calculation, and the inputs are the actual dollar cost of being wrong in each direction. See also: - Dollarized Confusion Matrix Tool (https://templeton.host/tools/dollarized-confusion-matrix) - Verification Trap (https://templeton.host/lexicon/verification-trap) - Autonomy State Machine (https://templeton.host/lexicon/autonomy-state-machine) --- ### Quadrant Shifting URL: https://templeton.host/lexicon/quadrant-shifting/ Definition: Capital investments that move a task to a better position on the Verification Quadrant. Five moves: build a verifier, decompose the task, enrich inputs, constrain outputs, build a rubric. Why this term exists: The quadrant isn't destiny. If you can build a verifier that makes verification cheap, a task that was "don't automate" becomes "automate now." That verifier is a capital investment with compounding returns. See also: - Quadrant Shifting Tool (https://templeton.host/tools/quadrant-shifting) - Verification Quadrant (https://templeton.host/tools/verification-quadrant) - AI Sweet Spot (https://templeton.host/lexicon/ai-sweet-spot) - Verification Trap (https://templeton.host/lexicon/verification-trap) - Construction Spread (https://templeton.host/lexicon/construction-spread) --- ### Operational Alpha URL: https://templeton.host/lexicon/operational-alpha/ Definition: Excess return on enterprise value generated through systematic identification and capture of mispriced edges in business operations. The operational equivalent of alpha in financial markets. Why this term exists: The same discipline that finds mispriced securities can find mispriced business processes. The directed graph is the model. The tools are the execution engine. The alpha is the return. See also: - Directed Graph (https://templeton.host/frameworks/directed-graph) - AI Operations Tools (https://templeton.host/tools) - Soft Spot (https://templeton.host/lexicon/soft-spot) - Construction Spread (https://templeton.host/lexicon/construction-spread) --- ### Construction Spread URL: https://templeton.host/lexicon/construction-spread/ Definition: S = (annual_value x P(success)) / build_cost. The risk-adjusted return on the capital deployed to build a knowledge asset. Rank opportunities by spread descending. Deploy capital top-down until budget is exhausted. Why this term exists: PE funds rank deals by risk-adjusted return on deployed capital. Knowledge assets are the same asset class with a different depreciation curve. If you can't calculate the spread, you're guessing where to invest. See also: - Automation NPV (https://templeton.host/tools/automation-npv) - Knowledge Capital (https://templeton.host/frameworks/knowledge-capital) - Compile Time (https://templeton.host/lexicon/compile-time) - Dual Curve (https://templeton.host/lexicon/dual-curve) - Operational Alpha (https://templeton.host/lexicon/operational-alpha) --- ### Dual Curve URL: https://templeton.host/lexicon/dual-curve/ Definition: The simultaneous depreciation of AI models (distribution shift, competitive erosion) and appreciation of knowledge assets (verifiers, labeled corpora, institutional rubrics) - where the appreciating side gets better through operating use, not in spite of it. The net rate determines whether an automation is a wasting asset or a compounder. Why this term exists: Physical capital allocation assumes depreciation. Knowledge capital allocation has both curves running at once - and the appreciating side is structurally unusual because it ratchets up with use rather than down. Invest in the appreciating side (verifiers, data) not the depreciating side (models). See also: - Knowledge Capital (https://templeton.host/frameworks/knowledge-capital) - Verifier Capital (https://templeton.host/frameworks/verifier-capital) - Automation NPV (https://templeton.host/tools/automation-npv) - Construction Spread (https://templeton.host/lexicon/construction-spread) - Compile Time (https://templeton.host/lexicon/compile-time) --- ### Verifier Capital URL: https://templeton.host/lexicon/verifier-capital/ Definition: A verifier is one of the only capital assets that appreciates through operating use. Every failure it catches gets encoded as a new rule, test, or rubric line, and the next run inherits the catch. Use raises the floor instead of lowering it. Why this term exists: Standard capital allocation assumes assets decay through use - that is why depreciation schedules exist. Verifiers invert the schedule, which means they are systematically mispriced by NPVs that treat them like trucks. Repricing them correctly is the highest-leverage move in an AI operating budget. See also: - The Capital Value of Verifiers (https://templeton.host/frameworks/verifier-capital) - Knowledge Capital (https://templeton.host/frameworks/knowledge-capital) - Dual Curve (https://templeton.host/lexicon/dual-curve) - Templeton Ratio (https://templeton.host/lexicon/templeton-ratio) - Quadrant Shifting (https://templeton.host/lexicon/quadrant-shifting) - Proof Layer (https://templeton.host/lexicon/proof-layer) --- ### Quality Ratchet URL: https://templeton.host/lexicon/quality-ratchet/ Definition: A CI-enforced floor that only moves up. Once a quality metric hits a threshold, the system blocks any change that drops below it. Each improvement becomes the new minimum. The sequence of baselines is monotonically non-decreasing. Formally: a monotonic ratchet. Why this term exists: Ship a 15% improvement on Tuesday. Regress to baseline by Friday. The ratchet makes this structurally impossible - the floor is enforced by CI, not by good intentions. See also: - Quality Hillclimb (https://templeton.host/frameworks/quality-hillclimb) - Designed Convergence (https://templeton.host/frameworks/designed-convergence) - Gold Standard (https://templeton.host/lexicon/gold-standard) - Oracle Gradient (https://templeton.host/lexicon/oracle-gradient) --- ### Oracle Gradient URL: https://templeton.host/lexicon/oracle-gradient/ Definition: The computed vector from your current performance toward the unobservable optimum in a domain's performance space. You never reach the oracle. You approach it. Each measurement sharpens your estimate of where it is. Why this term exists: In every domain, there's an ideal performance point you can't directly observe. The oracle gradient gives you a direction to move even when you can't see the destination - it's the operational derivative of the Performance Frontier. See also: - The Performance Frontier (https://templeton.host/frameworks/performance-frontier) - Quality Ratchet (https://templeton.host/lexicon/quality-ratchet) --- ### Structured Elicitation URL: https://templeton.host/lexicon/structured-elicitation/ Definition: A controlled experiment designed to learn the operator's preferences. Pairwise comparisons, best-worst scaling, or adaptive conjoint analysis. Highest information per query of the three Deity Problem channels, but requires operator attention. Why this term exists: Passive observation is cheap but noisy. Sometimes you need to design an experiment that asks exactly the right question - targeting the specific preference dimension where uncertainty is highest. See also: - The Deity Problem (https://templeton.host/frameworks/deity-problem) - Revealed Preference (https://templeton.host/lexicon/revealed-preference) - Direct Query (https://templeton.host/lexicon/direct-query) - Drift Detector (https://templeton.host/lexicon/drift-detector) --- ### Revealed Preference URL: https://templeton.host/lexicon/revealed-preference/ Definition: Inferring the operator's preferences by watching what they actually do - not what they say they want. Based on revealed preference theory (Afriat's theorem, GARP). Cheapest evidence channel because the operator is doing what they would do anyway. Why this term exists: People are unreliable narrators of their own preferences. What they choose when real stakes are on the line reveals what they actually value. Behavioral observation captures this without interrupting the operator. See also: - The Deity Problem (https://templeton.host/frameworks/deity-problem) - Structured Elicitation (https://templeton.host/lexicon/structured-elicitation) - Direct Query (https://templeton.host/lexicon/direct-query) --- ### Direct Query URL: https://templeton.host/lexicon/direct-query/ Definition: A question posed to the operator, used only when the expected value of the answer exceeds the cost of the operator's attention. An agent that asks too many questions isn't diligent - it's poorly calibrated. Why this term exists: Some ambiguities can't be resolved through observation alone. Direct queries are the escape hatch, but they have a cost (operator attention) that has to be weighed against the expected improvement in decision quality. See also: - The Deity Problem (https://templeton.host/frameworks/deity-problem) - Structured Elicitation (https://templeton.host/lexicon/structured-elicitation) - Revealed Preference (https://templeton.host/lexicon/revealed-preference) --- ### Drift Detector URL: https://templeton.host/lexicon/drift-detector/ Definition: A posterior predictive check that detects when the operator's preferences have drifted from the learned model. Computed as the fraction of recent decisions the model predicted incorrectly. When the drift score exceeds a threshold, the agent triggers re-elicitation. Why this term exists: Preferences are not static. People change their minds, priorities shift, new constraints emerge. The drift detector is how the agent notices that the operator changed their mind - and responds proportionally. See also: - The Deity Problem (https://templeton.host/frameworks/deity-problem) - Autonomy State Machine (https://templeton.host/lexicon/autonomy-state-machine) - Structured Elicitation (https://templeton.host/lexicon/structured-elicitation) --- ### Demand Gravity URL: https://templeton.host/lexicon/demand-gravity/ Definition: The inescapable pull of real demand on product trajectories. Demand is always there, always pulling, and you can't negotiate with it. Map it or crash into it. Why this term exists: Ask "what can we build with this?" and you get a solution looking for a problem. Demand gravity is what kills it - the invisible force pulling users toward what they actually need, whether you measured it or not. See also: - The Demand Field (https://templeton.host/frameworks/demand-field) - Soft Spot (https://templeton.host/lexicon/soft-spot) - Operational Alpha (https://templeton.host/lexicon/operational-alpha) --- ### The Designer's Seat URL: https://templeton.host/lexicon/designers-seat/ Definition: The position of designing the game rather than playing it. Every multi-agent system is a game. You can optimize your moves within existing rules (playing), or you can choose the rules so that the equilibrium of self-interested behavior is your desired outcome (designing). The CTO's job is the second one. Why this term exists: Two jobs look identical from outside: optimizing your moves within existing rules, and choosing the rules so that the equilibrium IS your desired outcome. The second one is the job. See also: - Designed Convergence (https://templeton.host/frameworks/designed-convergence) - Autonomy State Machine (https://templeton.host/lexicon/autonomy-state-machine) --- ### Serial Composition URL: https://templeton.host/lexicon/serial-composition/ Definition: A pipeline topology where each node feeds the next. The output of stage i is the input of stage i+1. End-to-end yield is the product of per-node accept rates. The math is brutally multiplicative: ten stages at 80% per-gate ships 11% of jobs. Why this term exists: Every AI POC that fails to ship fails here. The demo clears one gate. Production demands ten in series. The compounded yield collapse is invisible until you write the math down. See also: - DAG Topology widget (https://templeton.host/widgets/dag-topology) - Yield Collapse widget (https://templeton.host/widgets/yield-collapse) - Parallel Composition (https://templeton.host/lexicon/parallel-composition) - Verifier-Producer (https://templeton.host/lexicon/verifier-producer) --- ### Parallel Composition URL: https://templeton.host/lexicon/parallel-composition/ Definition: A pipeline topology where a source fans out to N independent branches that later merge. Two variants. Parallel-AND requires every branch to clear (joint yield is still the product, no improvement over serial - only latency drops). Parallel-OR accepts when any branch clears (yield approaches 1 fast, the redundant-verifier pattern). Why this term exists: Parallel-AND is what most people mean by "parallelize the pipeline" - and it does not improve correctness, only throughput. Parallel-OR is the underused redundancy pattern: when false-rejects are cheap, running two verifiers in parallel and accepting on either catches misses. See also: - DAG Topology widget (https://templeton.host/widgets/dag-topology) - Serial Composition (https://templeton.host/lexicon/serial-composition) - Verifier-Producer (https://templeton.host/lexicon/verifier-producer) --- ### Verifier-Producer URL: https://templeton.host/lexicon/verifier-producer/ Definition: A two-node topology where a producer P emits a candidate and a verifier V either accepts it (forwards to output) or rejects it (loops back to P for a rerun). With unbounded reruns and accept rate a, expected number of attempts is 1/a, so expected total work is 1/a producer-runs. Why this term exists: Most operational AI systems are this shape and nobody draws it. Naming the verifier separately from the producer is what makes the gate predicate visible, the rework cost measurable, and the routing decision (which surface verifies?) discrete. See also: - DAG Topology widget (https://templeton.host/widgets/dag-topology) - Brier Reliability Diagram (https://templeton.host/widgets/reliability-diagram) - Partial Rework (https://templeton.host/lexicon/partial-rework) - Rework Loop (https://templeton.host/lexicon/rework-loop) --- ### Rework Loop URL: https://templeton.host/lexicon/rework-loop/ Definition: The cyclical structure introduced when a verifier rejects a producer's output and the producer must retry. Each pass through the pipeline is still a DAG; the cyclicality lives in the attempt-counter, not in any single pass. Expected total work scales as 1/accept producer-runs under unbounded retries. Why this term exists: Conflating "the pipeline has a cycle" with "the pipeline is not a DAG" makes the math intractable. The right framing: every individual pass is a DAG, and the rework loop is a Markov chain over attempts. This keeps topological order, parallelism, and associative composition intact. See also: - DAG Topology widget (https://templeton.host/widgets/dag-topology) - Verifier-Producer (https://templeton.host/lexicon/verifier-producer) - Partial Rework (https://templeton.host/lexicon/partial-rework) --- ### Partial Rework URL: https://templeton.host/lexicon/partial-rework/ Definition: A pipeline design where the producer emits N independent parts and the verifier checks each part independently. Only failed parts rerun. Expected total work per pass is the sum of per-part rework costs, strictly less than rerunning the entire producer per failure. Partial rework dominates full rework at any per-part accept rate below 100%, and the gap grows fast with N and with the failure rate. Why this term exists: The corollary that turns the verifier-producer pattern from "useful" into "economically obvious". Designing pipelines so the verifier can check parts (not just the whole) is one of the highest-leverage moves available, and most teams have not noticed it because they have not drawn the DAG. See also: - DAG Topology widget (https://templeton.host/widgets/dag-topology) - Verifier-Producer (https://templeton.host/lexicon/verifier-producer) - Rework Loop (https://templeton.host/lexicon/rework-loop) - Serial Composition (https://templeton.host/lexicon/serial-composition) --- ### The Conviction Fraction URL: https://templeton.host/lexicon/conviction-fraction/ Definition: f* = (mu - r) / (gamma x sigma^2). The fraction of capital you commit to a bet: the edge over the hurdle, divided by risk aversion times variance. It is the Merton optimal risky-asset weight under CRRA utility; the Kelly criterion is the special case where risk aversion goes to 1 (log utility). A bigger edge or tighter evidence raises the size; more variance or more caution shrinks it. Why this term exists: The Scientist produces the posterior - the estimate of the edge, the variance, and the tails. The Allocator supplies the risk aversion - how much of that conviction to turn into position size. Together they answer the only question left once you have an edge: how big. Size the position to the strength of the evidence, not the strength of the feeling. See also: - The Performance Frontier (https://templeton.host/frameworks/performance-frontier) - Dollarized Confusion Matrix (https://templeton.host/tools/dollarized-confusion-matrix) - Oracle Gradient (https://templeton.host/lexicon/oracle-gradient) - Causal Yield (https://templeton.host/lexicon/causal-yield) - Portfolio Alpha (https://templeton.host/lexicon/portfolio-alpha) --- ### Causal Yield URL: https://templeton.host/lexicon/causal-yield/ Definition: yield = E[Y | do(X=1)] - E[Y | do(X=0)]. The return on an operating intervention, measured as the average treatment effect under Pearl's do-operator: the expected outcome when you set X versus when you do not. The do() is intervention, not observation - it is what separates a yield you can bank from a correlation you cannot. Why this term exists: Operators are paid for interventions that move the number, not for noticing that two numbers move together. A correlation says Y tends to be high when X is high; the do-operator says what Y becomes when you reach in and set X yourself. Confounders make those two quantities diverge, and the gap is exactly where operating capital gets burned - crediting a tailwind to a lever that never moved anything. See also: - Directed Graph (https://templeton.host/frameworks/directed-graph) - Designed Convergence (https://templeton.host/frameworks/designed-convergence) - Operational Alpha (https://templeton.host/lexicon/operational-alpha) - Soft Spot (https://templeton.host/lexicon/soft-spot) - The Conviction Fraction (https://templeton.host/lexicon/conviction-fraction) - Portfolio Alpha (https://templeton.host/lexicon/portfolio-alpha) --- ### Portfolio Alpha URL: https://templeton.host/lexicon/portfolio-alpha/ Definition: The excess return that exists only when the allocation decision and the operating execution are both right. Brinson performance attribution splits a portfolio's active return into an allocation effect (did you weight the right bets?) and a selection effect (did each bet perform?); portfolio alpha is the interaction the two lenses miss alone - you have to fund the right edges AND have operated the P&L well enough to capture them. Why this term exists: An allocator who cannot operate funds the right bets and watches the organization fail to capture them. An operator who cannot allocate executes flawlessly on bets that should never have been funded. Portfolio alpha is the interaction term: it takes knowing which bets to make and having the operating muscle to extract the return. Neither capability prices it alone. See also: - Operational Alpha (https://templeton.host/lexicon/operational-alpha) - Construction Spread (https://templeton.host/lexicon/construction-spread) - The Conviction Fraction (https://templeton.host/lexicon/conviction-fraction) - Causal Yield (https://templeton.host/lexicon/causal-yield) - Directed Graph (https://templeton.host/frameworks/directed-graph) --- ### Bayesian Alignment Kernel URL: https://templeton.host/lexicon/bayesian-alignment-kernel/ Definition: A typed contract for running an AI loop on two posteriors instead of one score: an operator-preference posterior and a reality-outcome posterior with disjoint evidence channels, a per-evaluator reliability ledger that gates every update, and a divergence statistic between the two that rings when the proxy is being gamed. Makes machine judgment underwritable by forcing every output to carry the evidence channel that produced it and the measured reliability of every grader that scored it. Why this term exists: The single number most agent loops optimize - "quality" - silently merges four different objects (what the operator wants, what reality rewards, what the system produces, how much each grader can be trusted), and while they share one number reward hacking is not even statable: there is nothing for the hacked metric to diverge from. Typing the concerns apart turns autonomy from a vibe into a balance sheet: it extends the way a lender extends credit, against measured loss rates instead of demos, with a stated alarm for the day the book is being marked to model instead of market. See also: - Alignment Kernels (framework) (https://templeton.host/frameworks/alignment-kernels) - The Capital Value of Verifiers (https://templeton.host/frameworks/verifier-capital) - The Deity Problem (https://templeton.host/frameworks/deity-problem) - Quality Hillclimb (https://templeton.host/frameworks/quality-hillclimb) - Designed Convergence (https://templeton.host/frameworks/designed-convergence) - Verification Trap (https://templeton.host/lexicon/verification-trap) --- ### Reliability Ledger URL: https://templeton.host/lexicon/reliability-ledger/ Definition: A per-evaluator audit trail of measured trust: agreement with prevalence, calibration, test-retest variance, cross-provider disagreement, recorded as a typed entry with a date and a quarantine flag. The same ρ object enters the math at every gate that consumes evaluator evidence - so an untrusted grader cannot move the system, no matter how loud it is. Why this term exists: Most pipelines treat evaluator reliability as a quarterly report you read and then ignore. The ledger collapses that two-step into a single typed object the loop already consults: an evaluator below its floor contributes nothing to any update, automatically, with no manual discounting to remember. See also: - Alignment Kernels (https://templeton.host/frameworks/alignment-kernels) - The Capital Value of Verifiers (https://templeton.host/frameworks/verifier-capital) - Verifier Capital (https://templeton.host/lexicon/verifier-capital) - Gold Standard (https://templeton.host/lexicon/gold-standard) --- ### Verifier Battery URL: https://templeton.host/lexicon/verifier-battery/ Definition: The ordered stack of checks that decide whether an output ships: deterministic gates first (cheap, sharp, false-negative-free), graded checks second (LLM or human judges, always from a different provider than the generator), each carrying its precision weight from a reliability ledger entry. Cliffs refuse; graded checks scale the q-score. Why this term exists: A single grader is a single point of failure with no second opinion. A battery composed in the right order - cheap-and-sharp before expensive-and-fuzzy, generator and adversary on different providers - turns evaluation into a structured search instead of a vibe call, and gives reliability a place to attach per-checker rather than per-pipeline. See also: - The Capital Value of Verifiers (https://templeton.host/frameworks/verifier-capital) - Reliability Ledger (https://templeton.host/lexicon/reliability-ledger) - Verifier Capital (https://templeton.host/lexicon/verifier-capital) - Alignment Kernels (https://templeton.host/frameworks/alignment-kernels) --- ### Evidence Channel URL: https://templeton.host/lexicon/evidence-channel/ Definition: A typed pipe with one source of truth and one destination posterior - preference evidence (tournaments, keep/drop verdicts) updates only the preference posterior; realized outcomes (the thing that shipped, then what happened) update only the reality posterior. Disjoint by construction; nothing else gets the wrong key. Why this term exists: A single quality score collapses preference and reality into one latent and makes the most diagnostic divergence in the system - is the proxy being gamed or just stale? - literally unstateable. Channeled evidence is the type discipline that keeps the two questions answerable. See also: - Alignment Kernels (https://templeton.host/frameworks/alignment-kernels) - The Deity Problem (https://templeton.host/frameworks/deity-problem) - Structured Elicitation (https://templeton.host/lexicon/structured-elicitation) - Revealed Preference (https://templeton.host/lexicon/revealed-preference) --- ### Divergence Alarm URL: https://templeton.host/lexicon/divergence-alarm/ Definition: A scale-free statistic between the preference posterior and the reality posterior that rings when the two stop agreeing under continued optimization. Three estimators in increasing cost: Kendall-tau between ranks, regret of the preference winner under reality, and the slope-opposition trend where the proxy climbs while realized outcomes turn down. Why this term exists: Goodhart's law is unfalsifiable when there is only one number; you cannot detect the proxy being gamed unless something else exists to compare against. The alarm is the only thing that fires on reward hacking before it shows up in the P&L - because by the time it shows up in the P&L, the optimization has already converged on the wrong target. See also: - Alignment Kernels (https://templeton.host/frameworks/alignment-kernels) - Drift Detector (https://templeton.host/lexicon/drift-detector) - Verification Trap (https://templeton.host/lexicon/verification-trap) --- ### P&L Engineering URL: https://templeton.host/lexicon/pnl-engineering/ Definition: Engineering as a discipline practiced from the financial statements down, not the tech stack up. The optimization target is per-unit margin and accumulated capital quality - not throughput, not velocity, not coverage. Every architectural decision is priced against its expected effect on the P&L over the horizon the firm is actually trying to survive. Why this term exists: Most "engineering" advice optimizes a proxy - latency, test coverage, lines of code - while leaving the connection to the income statement implicit. Naming the discipline forces the engineer to defend the price tag of each decision in the units the business measures itself in, and forces the allocator to learn enough engineering to underwrite the spend. See also: - Knowledge Capital (https://templeton.host/frameworks/knowledge-capital) - Capital Allocation (https://templeton.host/frameworks/capital-allocation) - Operational Alpha (https://templeton.host/lexicon/operational-alpha) - Construction Spread (https://templeton.host/lexicon/construction-spread) --- ## Writing ### When a Free Model Can Replace the Frontier (and How to Prove It) Published: 2026-06-29 URL: https://templeton.host/writing/free-model-replace-frontier/ Markdown: https://templeton.host/writing/free-model-replace-frontier.md --- ### The Eighth Ledger: When Intellectual Labor Becomes a Capital Asset Published: 2026-06-17 URL: https://templeton.host/writing/capitalizing-intellectual-labor/ Markdown: https://templeton.host/writing/capitalizing-intellectual-labor.md --- ### Any Sufficiently Advanced Technology: A Taxonomy of Magic Published: 2026-06-09 URL: https://templeton.host/writing/technology-magic-taxonomy/ Markdown: https://templeton.host/writing/technology-magic-taxonomy.md >- --- ### What AI Actually Made Scarce Published: 2026-06-03 URL: https://templeton.host/writing/scarcity-inversion/ Markdown: https://templeton.host/writing/scarcity-inversion.md >- ---