A Zero-Leak Socratic Examiner: Server-Side Isolation of Answers and Bayesian Mastery Probing Trees: chapter answers
When an online academy uses a language model as a gatekeeping examiner, it often writes the model answer into the prompt and instructs the model never to reveal it. This book tests the opposite founding claim: with the answer in context, the cumulative probability of extraction, 1-(1-p)^k, tends to 1 as attempts accumulate and an instruction changes only the speed; only when the answer stays out of the context is content leakage zero; and mastery is booked with the Bayesian knowledge tracing of Corbett and Anderson, published in User Modeling and User-Adapted Interaction in 1995. Checking the…
DOI pending · monograph page goes live after publication · 5DL-B-262 · 中文版
- Chapter 1 · Telling It Not to Tell: Answers in Context and the Cumulative Leak Probability
Chapters 1 and 2 quantify why instructions fail: in an illustrative course of 40 items with 3 attempts per item, a class of only 30 leaks at least one item with probability 0.9727 when p is 0.001. They then draw informa… - Chapter 2 · Physical Isolation: Server-Side Grading, Information-Flow Graphs and the One-Bit Oracle
Program p21 draws four architectures as information-flow graphs. Only architecture C, with grading on the server and only bits for the model, pinches every path to the browser down to one bit, and a single hint writer t… - Chapter 3 · Bayesian Knowledge Tracing: Four Parameters, the Mastery Line and Identifiability
Chapters 3 and 4 cover the four-parameter update of knowledge tracing, the mastery line and identifiability, and show that when a fifth of learners hold the answers, false certificates make up 0.2534 of all certificates… - Chapter 4 · How Leaks Forge Mastery: Guess Rates, False Certificates and the Blind Spot of Fitting
Program p41 shows that a leak lets learners who hold the answers pass the examiner with certainty, and that a leak reaching a fifth of the class is enough to make one certificate in four false. Program p42 shows that a… - Chapter 5 · The Probing Tree: Information-Gain Item Selection, Greedy Trees and Stopping Rules
Chapter 5 builds probing trees from information gain, whose benefit comes mainly from the stopping rule: 7.71 items on average against 9.59 for a fixed order, with better accuracy. - Chapter 6 · Defence in Depth against Prompt Extraction: Correlated Layer Failure, Digest-Only Output Checks and Item Exposure
Program p61 shows that the layers of defence in depth can be multiplied only when they fail independently, whereas the failure mechanisms revealed by current research make them fail together; only isolation moves the li… - Chapter 7 · Classical Sources: "No Opening without Striving" in the Analects and the Slave-Boy Geometry of Plato's Meno
Program p71 writes "no opening without striving" and "one corner shown, three returned" as rules that can be computed: prompting after the attempt preserves the evidence, three correct transfer items give a posterior of… - Chapter 8 · The Empirical Loop: Red-Teaming the Academy Examiner for Leaks, the Mastery Probing Tree and a Pre-Registered Test
Chapter 8 gives the owner's eight-step red-team loop, the distilled Mastery Probing Tree and a pre-registered test; vouching for a term's safety of a build that keeps the answer in context would take 1437950 clean red-t…
Illustrative model calculation, not a live quote