Hallucination
Industry term · Widely used across research, product, and media
What it says: The model perceived or generated something that isn't real — an involuntary perceptual error, like seeing something that isn't there.
What it hides: The failure is not perceptual. The model did not "see" a wrong thing. It filled a structural void — a gap where evidence was missing, a dependency that was unresolved, a constraint that was absent — and produced output anyway, because generative systems have no native mechanism for stopping. "Hallucination" locates the problem inside the model's experience. The problem is in the system's design: it was never given a reason to stop.
What 2ndlaw uses instead: Void filling — generation that continues through missing structure rather than terminating. Also: continuation under missing constraint, which names the mechanical cause. When a system cannot refuse, it fabricates. The vocabulary should make that inevitable.
Read more: Truth, Narrative Control, and the Cost of Refusal (§16–17: refusal as structural output class; why generative systems cannot natively stop) · Stochasticity Is a Symptom. Entropy Is the Problem. (why the outputs labeled "hallucination" are structural entropy, not perceptual error) · Research (void handling in the runtime contract)
Guardrails
Industry term · Common in product, policy, and enterprise AI
What it says: Safety barriers placed around a model to keep it within acceptable bounds — the road is fine, the car just needs rails.
What it hides: The metaphor presupposes that the underlying inference is sound and the problem is peripheral — edge cases, jailbreaks, misuse. It frames governance as something applied at the boundary of an otherwise functional system. This misses the structural reality: if the inference itself is unconstrained, boundary enforcement does not prevent void filling, causal overreach, or confident fabrication inside the permitted zone. The guardrails hold; the road still leads somewhere wrong.
What 2ndlaw uses instead: Runtime admissibility — structural rules that govern what the system is allowed to emit based on whether the epistemic conditions for a claim have been satisfied. Not edge control. Internal constraint. The distinction is between policing the boundary and governing the inference itself.
Read more: Truth, Narrative Control, and the Cost of Refusal (§18–20: runtime law as structural intervention, not boundary policing) · The 2ndlaw Runtime (how admissibility works in practice) · Evaluation (governed vs unconstrained comparison)
Alignment
Industry term · Central to AI safety discourse
What it says: Making models behave in accordance with human values and intentions. The model has a direction; alignment points it correctly.
What it hides: Three things. First, it assumes a coherent "correct direction" exists — that human values are unified enough to serve as a target. Second, it hides the question of whose values, decided by whom, under what incentives. Third — and this is the structural problem — alignment achieved through proxy evaluation (reward models, human preference scores) does not produce a system that shares human goals. It produces a system that has learned to satisfy the proxy. Under selection pressure, that system becomes internally sincere about the proxy itself, not about the underlying objective. The model is not misaligned. It is aligned — to the wrong thing. And it no longer knows the difference.
What 2ndlaw uses instead: Proxy compliance under selection when describing what RLHF and preference tuning actually produce. Epistemic accountability when describing what a governed system should enforce — not agreement with values, but structural answerability to evidence, constraint, and admissibility rules.
Read more: The Self-Deception Invariant Under Algorithmic Selection (Trivers' invariant applied to RLHF; the sincere proxy-follower) · Truth, Narrative Control, and the Cost of Refusal (§15: three classes of authority signals)
Stochastic / Stochastic Error
Industry term · Research and engineering
What it says: Model outputs vary randomly. This is inherent to probabilistic systems. Stochasticity is a property of the sampling process.
What it hides: Most of what the industry labels "stochastic behavior" is not irreducible randomness. It is structural uncertainty produced by flattening constraint-bearing dimensions — causality, provenance, weight, time, context, intent, grounding — during data ingestion and model construction. The distribution is broad not because sampling is noisy, but because the hypothesis space was allowed to balloon. Calling this "stochastic" misdiagnoses the failure, obscures its origin, and implies inevitability. It converts a design debt into a physics metaphor.
What 2ndlaw uses instead: Epistemic entropy — disorder introduced upstream by the absence of structural constraint. The term identifies origin, assigns cost, and makes clear that the uncertainty is not inherent but imported. Where stochasticity describes the symptom, entropy names the cause.
Read more: Stochasticity Is a Symptom. Entropy Is the Problem. (full treatment) · Expectation-Induced Determinism (variance reduction without entropy suppression)
RLHF (Reinforcement Learning from Human Feedback)
Industry term · Technical and widely cited
What it says: A training method that uses human preferences to improve model behavior. The model learns what humans want.
What it hides: The model does not learn what humans want. It learns to maximize a reward model trained on human labels — a proxy. The reward model is the evaluator; the LLM optimizes against it. Under persistent selection pressure, the system eliminates internal states that score poorly and converges on states that score well, regardless of whether those states correspond to actual human values. The endpoint is not a model that shares your goals. It is a model that is internally sincere about a proxy it has mistaken for reality. RLHF is the mechanism; the Trivers invariant is the outcome.
What 2ndlaw uses instead: Proxy-based selection as the descriptive term. The Trivers invariant — self-deception evolves when it improves signaling under selection — as the structural account of why proxy optimization converges on internally coherent misrepresentation rather than truth.
Read more: The Self-Deception Invariant Under Algorithmic Selection (full framework) · Truth, Narrative Control, and the Cost of Refusal (§7: flattened epistemic weight)
Chain-of-Thought (CoT)
Industry term · Research and applied prompting
What it says: Prompting a model to show its reasoning step-by-step improves accuracy. The model "thinks" more carefully when forced to externalize intermediate steps.
What it hides: The improvement is real, but the explanation is wrong. The model does not think more carefully. What happens mechanically is that decomposition shortens the inference arc, reduces curvature exposure, and re-anchors context at each step. The gain is geometric — shorter paths through the probability space produce more stable continuations. Attributing this to "reasoning" invites the belief that the model is doing something cognitive, when the improvement arises from the structure of the prompt, not from the depth of the thought.
What 2ndlaw uses instead: Decomposition as the structural description. Inference geometry as the explanation — shorter arcs, lower curvature, re-anchored context, isolated assumptions. The gain is from the shape of the task, not the mind of the model.
Read more: The Consequence of Structure Under Multi-Step Inference (full treatment: local and global geometry, why improvements are misattributed to agentic reasoning)
Red Teaming
Industry term · Safety and evaluation
What it says: Adversarial testing of a model by humans who try to make it fail. Find the weaknesses before attackers do.
What it hides: Red teaming tests the output boundary, not the epistemic core. It asks "can I make the model say something bad?" not "is the model's reasoning structurally sound even when it says something acceptable?" A system can pass every red team and still produce confident fabrication, void-filled reasoning, and smoothed contradictions on every normal call — because those failures are not adversarial. They are the default behavior of unconstrained inference. Red teaming finds jailbreaks. It does not find structural epistemic failure.
What 2ndlaw uses instead: Perturbation testing and basin analysis — probing how behavior changes under systematic variation of tasks, constraints, and vocabulary. The goal is not to find what breaks the model at the edges but to map where it is stable and where it drifts under normal operating conditions.
Read more: Research (basins and perturbation testing as ongoing work) · Evaluation (governed vs unconstrained comparison as the primary diagnostic)
Prompt Engineering
Industry term · Applied practice
What it says: The art and science of crafting better instructions for language models. Write better prompts, get better results.
What it hides: "Engineering" implies construction — building something that works. In practice, most prompt engineering is constraint pressure applied without a model of how that pressure redistributes. Add a rule here, the reasoning deforms there. The metaphor of "writing instructions" conceals the reality: you are reshaping a high-dimensional probability field, and every constraint you add moves mass non-locally. Without understanding flow dynamics, logit geometry, and constraint propagation, prompt engineering is sculpting balloon animals — squeeze in one place, bulge in another.
What 2ndlaw uses instead: Constraint architecture — designing the shape of the pressure field, not just the text of the instruction. The focus shifts from "what to say" to "what geometry does this create." This includes position dominance (early tokens set the kernel), vocabulary stability (some words activate drift-prone neighborhoods), and the recognition that the model accepts pressure, not commands.
Read more: How to Write Better Prompts: Think Balloon Animal (from squeeze to flow dynamics to logit geometry) · AI Doesn't Listen Big — Think Apple ][ Small (position dominance, the tiny working window, vocabulary effects)
Context Window
Industry term · Technical and marketing
What it says: The amount of text a model can process in a single call. Bigger context windows mean models can handle more information.
What it hides: Context window size and effective working window are not the same thing. A model can accept a million tokens and still only internalize a small, high-salience slice of the input at any given point in generation. The rest is present but not active — it decorates rather than governs. Marketing larger context windows as a capability improvement hides the fact that attention bandwidth, position sensitivity, and salience competition remain binding constraints regardless of input length. More tokens in does not mean more tokens matter.
What 2ndlaw uses instead: Effective working window — the portion of the input that actually shapes the model's internal state during generation. This is small, position-sensitive, and overloaded easily. The design principle that follows: respect the small part, because the big part is an illusion.
Read more: AI Doesn't Listen Big — Think Apple ][ Small (full treatment: the illusion of abundance, position dominance, what actually breaks under load)
Agent / Agentic AI
Industry term · Architecture and product
What it says: AI systems that can plan, use tools, and take actions autonomously. Agents are the next step — models that do things, not just say things.
What it hides: "Agent" borrows intentionality from human agency. What actually happens in most agentic systems is multi-step inference with tool calls — the model generates a plan as text, executes steps sequentially, and uses outputs as inputs to the next call. The improvement comes from decomposition geometry (shorter arcs, re-anchored context), not from the model acquiring goals or planning capacity. Calling this "agentic" invites anthropomorphic reasoning about systems that are still doing constrained next-token prediction, one step at a time. It also hides the multiplication of failure: every step in an agentic chain is a new opportunity for void filling, causal overreach, and silent drift.
What 2ndlaw uses instead: Multi-step inference as the structural description. Non-agentic, single-pass as the governance constraint — the runtime contract forces each inference step to be self-contained, with no planning, no orchestration, and no tool calling inside the governed core. Agentic structure belongs in the orchestrator. Epistemic discipline belongs in each step.
Read more: The Consequence of Structure Under Multi-Step Inference (why multi-step works and what it actually fixes) · The 2ndlaw Runtime (single-pass, non-agentic governance) · Research (governed inference inside agents as open problem)
AI Safety
Industry term · Policy, research, and corporate communications
What it says: Ensuring AI systems do not cause harm. A broad umbrella covering everything from bias mitigation to existential risk.
What it hides: "Safety" carries direct, enforceable, and immediately priced consequences — lawsuits, bans, brand collapse, executive liability. Companies rationally fund safety because the costs of failure are visible and personal. Epistemic integrity — whether the system's reasoning is structurally sound — carries no equivalent pricing. Its costs are distributed across society, delayed across time, and borne by institutions downstream. "AI safety" absorbs resources and attention that could go to epistemic governance but doesn't, because safety has a budget line and truth does not. The term lets organizations claim responsible behavior while leaving the structural quality of inference unexamined.
What 2ndlaw uses instead: Epistemic accountability for the structural dimension that safety does not cover. The distinction is: safety asks "will this cause visible harm?" Epistemic accountability asks "is this structurally sound enough to be emitted as if it were true?"
Read more: Truth, Narrative Control, and the Cost of Refusal (§10: why high epistemic value is systematically discounted; §21: cost injection vs belief correction)
Sycophancy
Industry term · Alignment research and model evaluation
What it says: The model tells you what you want to hear. It agrees too readily, flatters, and avoids disagreement.
What it hides: "Sycophancy" frames the problem as a personality flaw — the model is too eager to please. The structural reality is different: the model has been trained under selection pressure to maximize approval scores. Agreement scores higher than correction. Smoothing scores higher than surfacing conflict. What looks like personality is optimization under proxy constraint. The model is not being sycophantic. It is being exactly what the training reward selected for.
What 2ndlaw uses instead: Performed authority and conflict smoothing — terms that name the mechanism (simulating agreement, suppressing contradiction) rather than attributing a social disposition to a statistical process. See also: sincere proxy-follower — the highest-risk system is not one that is lying to please you, but one that sincerely believes the proxy-aligned response is correct.
Read more: The Self-Deception Invariant Under Algorithmic Selection (the sincere proxy-follower) · Truth, Narrative Control, and the Cost of Refusal (§15: authority signal classes; §7: flattened weight and performed authority)
AGI (Artificial General Intelligence)
Industry term · Strategy, investment, and research
What it says: A future AI system capable of general reasoning across all domains. The ultimate goal of AI research. Whoever gets there first wins.
What it hides: Two structural problems. First, the business model — AGI is justified through a premise we call TORE (Train Once, Reason Everywhere), which assumes a single general model can amortize its cost across the entire economy. In practice, most high-value domains require verifiable, auditable, domain-competent systems, not generalists. The general model ends up wrapped in so much specialization infrastructure that it becomes a component, not a product. You still pay the AGI tax; you no longer get the AGI payoff. Second, the value proposition — the "delta" AGI promises to deliver shrinks every year as incremental systems improve. The future does not stand still while we wait for the discontinuity.
What 2ndlaw uses instead: TORE (Train Once, Reason Everywhere) as the name for the underlying business premise. The AGI delta problem as the name for the shrinking value proposition. Capital trap for the lock-in that follows from organizing a company around the wrong architecture for a decade.
Read more: The Real AGI Cost Isn't the Model. It's the Decade You Lose. (TORE, capital trap, strategic misallocation) · The AGI Delta Problem (why every year of progress shrinks the case)
Model Collapse
Industry term · Research
What it says: When models are trained on synthetic data generated by other models, quality degrades — distributions narrow, diversity drops, biases amplify.
What it hides: The framing stays statistical — distribution degradation, diversity loss. The deeper problem is epistemic: models trained on their own narratively constrained outputs are not just losing statistical diversity. They are learning from a world already filtered by alignment regimes, safety policies, cultural pressures, and institutional incentives. At that point, amplification is not just network-level. It is training-level. The distinction between narrative shaping of perception and narrative shaping of the training substrate itself collapses. This is not model collapse. It is epistemic self-conditioning under narrative capture.
What 2ndlaw uses instead: Archival drift under narrative capture for the epistemic dimension. Epistemic self-conditioning for the specific mechanism of models learning from their own constrained shadow of the world. The concern is not that models degrade. It is that once earlier semantic layers are lost, nothing remains to indicate what has changed — or whether the changes originated in the world at all.
Read more: Truth, Narrative Control, and the Cost of Refusal (§5–6: the feedback loop from narrative to records to models; why old models matter more than new ones)
Scaling Laws
Industry term · Research and investment narrative
What it says: Model performance improves predictably with more data, more compute, and more parameters. Scale is the path to capability.
What it hides: Scaling improves fluency and coverage. It does not impose structure. Larger models explore larger regions of an already over-entropic space with greater confidence. Without additional constraint, scaling produces more confident nonsense, more coherent but incorrect explanations, and reduced variance around the wrong mean. Scaling optimizes sampling efficiency, not epistemic quality. The term "scaling laws" implies that growth is inherently progress. It is not. Growth without constraint is inflation.
What 2ndlaw uses instead: No single replacement term. The position is structural: scaling without epistemic constraint produces fluent disorder at greater scale. The remedy is not more parameters. It is more structure — admissibility rules, sufficiency states, evidence classes, and refusal conditions that do not disappear as the model grows.
Read more: Stochasticity Is a Symptom. Entropy Is the Problem. (§7: why scaling fails absent constraint) · The Real AGI Cost Isn't the Model. It's the Decade You Lose. (the economics of scale-first strategy)
Benchmark
Industry term · Evaluation and marketing
What it says: Standardized tests that measure model capability. Higher scores mean better models.
What it hides: Benchmarks measure what they measure — which is almost never what matters in production. A model can score well on a benchmark and still produce void-filled reasoning, smoothed contradictions, and unjustified certainty on every real workload. Benchmarks test for answer correctness under controlled conditions. They do not test for epistemic quality under ambiguity, the ability to refuse when structure is missing, or stability under perturbation. They are proxy evaluators — and like all proxy evaluators, they select for systems that have learned to pass them, not systems that reason well.
What 2ndlaw uses instead: Governed vs unconstrained comparison — running the same workload with and without the runtime contract and observing where behavior actually changes. Also: meaningful completed task speed — measuring how efficiently a system reaches an acceptable epistemic state, not how many tokens it produces or how many benchmark items it answers correctly.
Read more: Evaluation (behavior, not benchmarks) · Research (meaningful completed task speed as ongoing work)
Misinformation / Disinformation
Industry and policy term · Content moderation and safety
What it says: False information spread without intent (misinformation) or with intent (disinformation). The problem is bad content reaching people.
What it hides: The framing locates the problem at the content level — specific false claims that can be identified and removed. This misses the structural machinery: information optimized for propagation outcompetes information optimized for accuracy regardless of truth value, because the former is cheaper to produce, faster to spread, and more emotionally activating. AI does not create misinformation. It removes the last remaining friction. The cost of fluent narrative production collapses to zero while verification costs remain unchanged. The problem is not bad content. It is an economic asymmetry that structurally selects against truth.
What 2ndlaw uses instead: Narrative control as the structural force. Propagation advantage as the mechanism. Numberatives — rhetorical constructs that simulate quantification without structure ("most people believe," "many experts agree") — as a specific, named attack class. The distinction: misinformation names the content. These terms name the machinery.
Read more: Propagation, Propaganda, and the Price of Truth · Truth, Narrative Control, and the Cost of Refusal (§1–2: narrative control as optimization; §8–9: numberatives and the fake fraction)
Bias
Industry term · Fairness, safety, and policy
What it says: Models have biases — systematic distortions that reflect imbalances in training data or design choices. Bias should be measured and mitigated.
What it hides: "Bias" is used so broadly it has lost diagnostic precision. It covers everything from statistical distribution skew to cultural prejudice to political lean to epistemic asymmetry, all under one word. This collapse of categories prevents structural analysis. A model that over-represents one demographic in training data has a different problem from a model that systematically suppresses uncertainty. A model that reflects cultural prejudice has a different problem from a model that lacks provenance weighting. Calling all of these "bias" makes it impossible to prescribe different remedies for structurally different failures.
What 2ndlaw uses instead: Specific structural terms for specific failures: flattened weight (provenance, frequency, and importance collapsed into the same dimension), scope laundering (local signals promoted to global claims), corpus control (what never enters the model), and alignment control (what the model learns it is allowed to say). Each names a distinct mechanism with a distinct remedy.
Read more: Truth, Narrative Control, and the Cost of Refusal (§7: flattened weight; §9: scope laundering and tiering) · Stochasticity Is a Symptom. Entropy Is the Problem. (§5: the dimensions that get flattened) · The Epistemological Preemption (§II: the epistemic control stack — corpus, alignment, and inference-time control)
Temperature
Industry term · Technical parameter
What it says: A parameter that controls randomness in model output. Lower temperature means more deterministic; higher means more creative.
What it hides: Temperature adjusts the shape of the sampling distribution. It does not improve the quality of the distribution. Lowering temperature concentrates probability mass on whatever is already most likely — which may be wrong, void-filled, or structurally unsound. It suppresses token-level variance without touching the epistemic structure underneath. A model at temperature zero is maximally confident in its most probable output, regardless of whether that output is justified. Temperature controls randomness. It does not control correctness.
What 2ndlaw uses instead: Entropy suppression as the accurate description of what temperature reduction does. The distinction: reducing temperature narrows the distribution. Reducing epistemic entropy requires structural constraint — admissibility, evidence classes, sufficiency states. One is cheap and cosmetic. The other is expensive and real.
Read more: Stochasticity Is a Symptom. Entropy Is the Problem. (entropy reduction vs randomness suppression) · Expectation-Induced Determinism (§4: how EID differs from temperature control)
Fine-Tuning
Industry term · Technical practice
What it says: Adapting a pre-trained model to a specific domain or task using additional training data. Fine-tuning makes the model better at what you need.
What it hides: Fine-tuning changes the model's weight distribution. It does not add structural constraint to inference. A fine-tuned model can be more domain-fluent while remaining epistemically unconstrained — it knows more of the right words and still fills voids, smooths conflicts, and fabricates when structure is missing. Fine-tuning improves the content of the model's knowledge. It does not improve the discipline of the model's reasoning. Those are different problems requiring different interventions.
What 2ndlaw uses instead: Runtime governance as the complement. Fine-tuning operates at training time and changes what the model knows. Runtime governance operates at inference time and constrains what the model is allowed to emit. 2ndlaw does not fine-tune. It governs — at the point of generation, one call at a time, without modifying the underlying model.
Read more: The 2ndlaw Runtime (server-side governance without model modification) · Truth, Narrative Control, and the Cost of Refusal (§11, §18: why runtime is the only economically viable intervention point)