# Signal Quest ML Masterclass — Figure Governance Manifest

This manifest governs 40 original teaching figures for a doctoral masterclass. The sequence is cumulative: **measurement** establishes what can be known, **decision** separates probability from permission, **representation** tests what models preserve, and **governed deployment** proves how evidence remains bounded after modeling. Every figure has one teaching claim. None is decorative, none may imply that a design target has been implemented, and none may present illustrative values as measured BTC or Polymarket results.

All manuscript-bound assets use a pure-white outer canvas, subtle role-based color inside the figure, external captions, and the canonical `chNN-figXX-kebab-name.png` convention. Lecture-only assets use `lecture-figXX-kebab-name.png`. Generation must preserve a 16:9-safe composition for slides and a legible center crop for the textbook. Charts must label axes and units; color may not be the only carrier of meaning. Existing 1200 × 520 PNGs require a 2400 × 1040 regeneration or equivalent high-resolution redraw before final slide or print delivery.

Status vocabulary is deliberately narrow. **Existing—redraw/QA** means that the canonical PNG exists but lacks recorded reproducible generation provenance or masterclass-resolution QA. **Planned—source locked** means that the teaching claim and evidence basis are approved, but the PNG does not yet exist. No status in this manifest means that empirical data or a measured model result exists.

The exact 13-file deep-dive source set referenced by the synthesis figure is `chapter-01-ml-foundations-deep-dive.html`, `chapter-02-labels-clocks-deep-dive.html`, `chapter-03-data-engineering-deep-dive.html`, `chapter-04-metrics-deep-dive.html`, `chapter-05-fair-experiments-deep-dive.html`, `chapter-06-calibration-abstention-deep-dive.html`, `chapter-07-reproducible-python-deep-dive.html`, `chapter-08-catboost-deep-dive.html`, `chapter-09-self-supervised-lob-deep-dive.html`, `chapter-10-causal-lob-transformer-deep-dive.html`, `chapter-11-llm-python-deep-dive.html`, `chapter-12-replay-paper-trading-deep-dive.html`, and `chapter-13-agentic-monitoring-deep-dive.html`, all under `docs/textbook/btc-polymarket-ml/full-textbook/`.

## Movement I — Measurement: make the evidence and denominator visible

By the end of this movement, students should be able to identify the object being measured, the clock that makes it admissible, and the denominator that makes a metric interpretable.

| Figure id | Canonical PNG filename | Single teaching claim | Visual form | Exact data/source basis | Epistemic status | Manuscript anchor | Slide destination | Textbook destination | Generation method | Accessibility description | Status |
|---|---|---|---|---|---|---|---|---|---|---|---|
| M1.01 | `ch01-fig01-learning-decision-loop.png` | A model score is only one object in a longer evidence-to-outcome chain. | Left-to-right object chain with five distinct roles | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-01-ml-foundations-deep-dive.html`, §2 “Five objects that should never be collapsed into one word” and §3 “Prediction is not decision”; no numeric data. | Design target | Chapter 1, §§2–3 | Movement I, slide M1-01 | Chapter 1, immediately after the five-object table | Redraw the existing original PNG as a flat Graphviz chain; export 2400 × 1040 PNG on white; keep caption outside. | Five labeled stages—observation, model, probability estimate, decision policy, and outcome—are connected in order; a boundary separates the model from policy authority. | Existing—redraw/QA |
| M1.02 | `ch02-fig01-causal-label-timeline.png` | Evidence received after the cutoff is forbidden even when its event time is earlier. | Dual-clock timeline with allowed and excluded events | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-02-labels-clocks-deep-dive.html`, §2 and the worked trace at 09:59:58–10:05:00; times are explicitly illustrative. | Illustrative | Chapter 2, §2 “The evidence set has a different contract” | Movement I, slide M1-02 | Chapter 2, after the five-time-field table | Redraw the existing original PNG with native slide shapes; export 2400 × 1040 PNG; use solid versus hatched event markers as well as color. | Two aligned timelines compare event time and receive time; an event occurring before but received after the 10:00 cutoff is crossed out. | Existing—redraw/QA |
| M1.03 | `ch02-fig02-five-clock-availability-contract.png` | Temporal validity depends on preserving the meaning of each clock rather than collapsing timestamps. | Five-lane clock contract | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-02-labels-clocks-deep-dive.html`, §2 table defining event, receive, process, decision-cutoff, and settlement time; no observations or latency values. | Design target | Chapter 2, §2 | Movement I, slide M1-03 | Chapter 2, immediately after Figure 2.1 | Graphviz swimlane diagram to 2400 × 1040 PNG; one icon-free lane per clock; white canvas. | Five horizontal lanes define when an event occurred, arrived, was processed, became a decision, and later settled; arrows show only lawful temporal dependencies. | Planned—source locked |
| M1.04 | `ch03-fig01-event-to-feature-contract.png` | A feature is admissible only when its lineage reaches eligible validated raw events through a versioned recipe. | Provenance pipeline with rejection branch | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-03-data-engineering-deep-dive.html`, §§1–2 and equation “raw event → validation result → eligible event set → feature recipe”; no quality-rate claim. | Design target | Chapter 3, §1 “Raw evidence comes first” | Movement I, slide M1-04 | Chapter 3, after the provenance equation | Redraw the existing original PNG in Graphviz with a quarantine branch; export 2400 × 1040 PNG. | Raw capture passes through validation, cutoff eligibility, and a versioned recipe to a feature vector; invalid records branch to preserved quarantine evidence. | Existing—redraw/QA |
| M1.05 | `ch04-fig01-confusion-matrix-denominators.png` | Every classification metric is a different question over the same four outcome counts. | Confusion matrix with denominator overlays | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-04-metrics-deep-dive.html`, §1 formulas and the illustrative worked trace with 1,000 cases, 950 Down, 50 Up, TP=30, FP=30, FN=20, TN=920 for Model B. | Illustrative | Chapter 4, §1 “A metric is a question with a denominator” | Movement I, slide M1-05 | Chapter 4, before the metric-question table | Native slide table plus equation callouts; export 2400 × 1350 PNG; label the worked counts “Illustrative.” | A two-by-two matrix shows true and predicted classes, with braces indicating the denominators for accuracy, precision, recall, and specificity. | Planned—source locked |
| M1.06 | `ch04-fig03-threshold-sweep-tradeoffs.png` | Changing the classification threshold moves errors rather than revealing one universally correct operating point. | Symbolic threshold sweep with linked confusion matrices | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-04-metrics-deep-dive.html`, §1 metric definitions; use only symbolic functions TP(τ), FP(τ), TN(τ), and FN(τ), with no empirical curve or selected threshold. | Illustrative; formula-derived | Chapter 4, after §1 | Movement I, slide M1-06 | Chapter 4, after Figure 4.1 | Vega-Lite layered schematic using symbolic threshold positions, plus three native matrix insets; export 2400 × 1350 PNG. | A threshold axis links low, middle, and high thresholds to three small confusion matrices, showing false positives falling while false negatives rise. | Planned—source locked |
| M1.07 | `ch04-fig04-roc-pr-operating-points.png` | ROC and precision–recall views answer different questions at the same threshold operating point. | Paired axis maps joined by threshold tokens | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-04-metrics-deep-dive.html`, §§1 and 3; scikit-learn model-evaluation documentation cited in `docs/textbook/btc-polymarket-ml/source-map.md`; use symbolic operating points only, not an invented AUC. | Literature-derived; illustrative rendering | Chapter 4, §3 “Ranking is not calibration” | Movement I, slide M1-07 | Chapter 4, immediately before §3 | Native slide axes and labeled points; no fitted curve or AUC number; export 2400 × 1350 PNG. | Two plots map the same threshold to false-positive-rate versus true-positive-rate and recall versus precision; a prevalence baseline is labeled but not assigned a value. | Planned—source locked |
| M1.08 | `ch04-fig02-calibration-curve.png` | Calibration asks whether forecast probabilities match observed frequencies, not whether positives rank above negatives. | Reliability diagram with perfect-calibration diagonal | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-04-metrics-deep-dive.html`, §3, and `docs/textbook/btc-polymarket-ml/full-textbook/chapter-06-calibration-abstention-deep-dive.html`, §1; any non-diagonal points remain explicitly illustrative and must not report ECE. | Illustrative | Chapter 4, §3 “Ranking is not calibration” | Movement I, slide M1-08 | Chapter 4, after the Brier and log-loss equations | Redraw the existing original PNG as a Vega-Lite reliability schematic; export 2400 × 1350 PNG; annotate “Illustrative—not fitted data.” | A diagonal shows perfect calibration while two forecast bins fall above and below it, labeled underconfidence and overconfidence. | Existing—redraw/QA |
| M1.09 | `ch04-fig05-brier-log-loss-penalties.png` | Log loss penalizes confident wrong forecasts more sharply than the Brier score. | Formula-derived loss curves for y=0 and y=1 | Brier and log-loss equations in `docs/textbook/btc-polymarket-ml/full-textbook/chapter-04-metrics-deep-dive.html`, §3; evaluate the displayed formulas over p from 0.01 to 0.99 only, with no observed dataset. | Illustrative; formula-derived | Chapter 4, §3 | Movement I, slide M1-09 | Chapter 4, immediately after Figure 4.2 | Vega-Lite line chart generated directly from the two printed scoring-rule formulas; export 2400 × 1350 PNG. | Four labeled curves compare Brier and log-loss penalties for positive and negative outcomes; the log-loss curve rises steeply near a confidently wrong endpoint. | Planned—source locked |
| M1.10 | `ch05-fig01-temporal-cross-validation.png` | Purge and embargo gaps protect evaluation only when their widths follow feature, label, and availability windows. | Chronological split with overlap brackets | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-05-fair-experiments-deep-dive.html`, §§1–2; widths are symbolic functions of maximum lookback, label horizon, and measured availability delay, not fixed durations. | Design target | Chapter 5, §2 “The job of the purge and embargo” | Movement I, slide M1-10 | Chapter 5, after the purge/embargo explanation | Redraw the existing original PNG as a proportional-but-not-to-scale timeline; Graphviz or native slide shapes; export 2400 × 1040 PNG. | Training, purge, embargo, validation, and untouched final-test blocks appear in chronological order; brackets show which feature and label windows could cross a boundary. | Existing—redraw/QA |

## Movement II — Decision: turn estimates into bounded, reviewable choices

By the end of this movement, students should be able to explain why calibration, cost, uncertainty, reproducibility, and authority gates can all convert a model score into abstention.

| Figure id | Canonical PNG filename | Single teaching claim | Visual form | Exact data/source basis | Epistemic status | Manuscript anchor | Slide destination | Textbook destination | Generation method | Accessibility description | Status |
|---|---|---|---|---|---|---|---|---|---|---|---|
| M2.01 | `ch06-fig02-ranking-calibration-decision-ladder.png` | Discrimination, calibration, and decision usefulness require separate evidence. | Three-rung evidence ladder | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-06-calibration-abstention-deep-dive.html`, §1 property table; no metric values. | Design target | Chapter 6, §1 “A score can rank well and mean the wrong thing” | Movement II, slide M2-01 | Chapter 6, after the property table | Graphviz ladder with one evidence artifact per rung; export 2400 × 1040 PNG. | Three ascending rungs are labeled ranking evidence, probability evidence, and cost-aware decision evidence; each rung names what it cannot prove. | Planned—source locked |
| M2.02 | `ch06-fig03-expected-value-abstention-boundary.png` | A positive raw probability difference can still have non-positive conservative per-unit expected value and therefore require abstention. | Symbolic expected-value number line with no-action interval | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-06-calibration-abstention-deep-dive.html`, §§2–3, including the explicitly illustrative p̂=0.58, q=0.55, c=0.04 trace; `docs/architecture/btc-polymarket-5m-prediction-architecture.html`, “Value test”; `docs/planning/btc-polymarket-5m-15m-research-backtest-replay-prompt.md`, “Execution realism,” which defines a $1/$0 resolved payout. For this teaching figure only, q is explicitly declared the executable per-unit purchase price, so `EV_unit = p̂ − q − c`; it is not treated as a generic probability. | Illustrative; formula-derived | Chapter 6, §2 “A probability is not the market’s probability” | Movement II, slide M2-02 | Chapter 6, after the conservative-difference equation | Native slide number line and decomposed per-unit expected-value equation; export 2400 × 1040 PNG; label all values and the $1/$0 payoff assumption “Illustrative.” | Under an explicitly stated $1/$0 payoff, a 0.58 estimate minus a 0.55 executable price and 0.04 allowance yields negative per-unit expected value inside the abstain region. | Planned—source locked |
| M2.03 | `ch06-fig01-abstention-decision-region.png` | Abstention is a first-class policy output rather than a failed classification. | Three-region decision map | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-06-calibration-abstention-deep-dive.html`, §3 conditions and fail-closed results; no learned boundary or frequency. | Design target | Chapter 6, §3 “Abstention is an action with a reason” | Movement II, slide M2-03 | Chapter 6, before the fail-closed condition table | Redraw the existing original PNG with native slide shapes; export 2400 × 1040 PNG; use labels and boundary patterns. | A central abstain region separates bounded positive and negative research decisions; stale data and artifact mismatch route directly into abstention. | Existing—redraw/QA |
| M2.04 | `ch06-fig04-uncertainty-cost-haircut.png` | Conservative edge is what remains after reference probability, cost, latency, and uncertainty deductions. | Waterfall decomposition ending at an eligibility gate | `docs/planning/btc-polymarket-5m-15m-research-backtest-replay-prompt.md`, “Smart risk-management doctrine” and “Execution realism”; `docs/textbook/btc-polymarket-ml/full-textbook/chapter-06-calibration-abstention-deep-dive.html`, §2; all magnitudes symbolic. | Design target | Chapter 6, §2 | Movement II, slide M2-04 | Chapter 6, immediately after Figure 6.2 | Native slide waterfall with variables rather than amounts; export 2400 × 1350 PNG. | A starting calibrated probability difference is reduced in separate steps for fees, slippage, latency, and uncertainty before meeting or missing a fixed gate. | Planned—source locked |
| M2.05 | `ch05-fig02-final-test-lock.png` | The final test ceases to be final the moment it influences threshold, feature, or model selection. | Locked temporal experiment map with prohibited feedback arrows | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-05-fair-experiments-deep-dive.html`, worked trace “the locked final month”; months 1–12 are explicitly illustrative. | Illustrative | Chapter 5, §3 “Baselines are scientific controls” | Movement II, slide M2-05 | Chapter 5, after the worked trace | Graphviz dependency diagram with months 1–7 train, 8–9 validation, and 10–12 final test; export 2400 × 1040 PNG. | Training and validation may feed model selection, while the final three-month block is padlocked and has no arrow back to development. | Planned—source locked |
| M2.06 | `ch07-fig01-reproducible-experiment-artifacts.png` | A result becomes reviewable only when code, environment, data, configuration, artifacts, and report share one run identity. | Artifact dependency graph | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-07-reproducible-python-deep-dive.html`, §§1 and 5; no claim that the system currently emits these artifacts. | Design target | Chapter 7, §1 “A notebook is not a research record” | Movement II, slide M2-06 | Chapter 7, after the reproducible-run equation | Redraw the existing original PNG as a Graphviz dependency graph; export 2400 × 1040 PNG. | Six artifact nodes converge on a run identifier, which points to a report; a notebook is shown as a caller, not the source of truth. | Existing—redraw/QA |
| M2.07 | `ch01-fig02-model-policy-authority-boundary.png` | Models may supply evidence, but only a deterministic policy may grant bounded research eligibility. | Capability boundary with one-way evidence flow | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-01-ml-foundations-deep-dive.html`, §3 eligibility conjunction; `docs/textbook/btc-polymarket-ml/full-textbook/safety-contract.md`; no implemented authorization claim. | Design target | Chapter 1, §3 | Movement II, slide M2-07 | Chapter 1, after the two-score worked trace | Graphviz trust-boundary diagram; model output crosses to policy, but no reverse authority path exists; export 2400 × 1040 PNG. | A model box sends a probability into a policy gate; failed data, calibration, and risk predicates exit as named abstention reasons, and no arrow reaches live execution. | Planned—source locked |
| M2.08 | `lecture-fig01-supervised-prediction-rl-control.png` | Supervised prediction estimates an outcome, whereas reinforcement learning selects actions to optimize cumulative reward through environment interaction. | Side-by-side objective and feedback-loop contrast | Sutton and Barto, *Reinforcement Learning: An Introduction*, 2nd ed., author-hosted edition `http://incompleteideas.net/book/RLbook2020.pdf`; local authority constraint in `docs/textbook/btc-polymarket-ml/full-textbook/safety-contract.md`; no claim that Signal Quest implements RL. | Literature-derived; lecture-only | Lecture bridge after Chapter 7; no manuscript claim | Movement II, slide M2-08 | None—lecture companion only | Graphviz: one supervised mapping and one agent–environment loop; add a structural “out of scope for Signal Quest” barrier; export 2400 × 1350 PNG. | Left panel maps labeled examples to an outcome estimate; right panel loops state, action, reward, and next state, with Signal Quest’s execution boundary blocking the action channel. | Planned—source locked |
| M2.09 | `ch06-fig05-abstention-reason-codes.png` | Distinct abstention reasons are necessary evidence because identical no-action outcomes can arise from different failures. | Decision tree terminating in reason-coded abstentions | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-06-calibration-abstention-deep-dive.html`, §3 condition table and worked trace; reason codes are the named chapter conditions, not observed frequencies. | Design target | Chapter 6, §3 | Movement II, slide M2-09 | Chapter 6, after the worked trace | Graphviz decision tree using data health, calibration, conservative difference, risk budget, and artifact compatibility predicates; export 2400 × 1350 PNG. | Five yes-or-no gates end in either bounded eligibility or a specific abstention reason such as stale data, stale calibration, insufficient difference, exhausted risk, or artifact mismatch. | Planned—source locked |

## Movement III — Representation: earn complexity without losing causality

By the end of this movement, students should be able to trace what boosted trees and sequence models represent, state the temporal boundary of self-supervision, and test that attention cannot see the future.

| Figure id | Canonical PNG filename | Single teaching claim | Visual form | Exact data/source basis | Epistemic status | Manuscript anchor | Slide destination | Textbook destination | Generation method | Accessibility description | Status |
|---|---|---|---|---|---|---|---|---|---|---|---|
| M3.01 | `ch08-fig02-boosting-residual-sequence.png` | Boosting improves a predictor by adding small trees that target the current objective’s remaining error. | Four-stage additive-model sequence | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-08-catboost-deep-dive.html`, §1 equation `score(x)=base_score+Σ η·tree_m(x)`; CatBoost primary paper cited in `docs/textbook/btc-polymarket-ml/source-map.md`; residuals remain symbolic r⁽ᵐ⁾. | Literature-derived; illustrative rendering | Chapter 8, §1 “What a boosted-tree model learns” | Movement III, slide M3-01 | Chapter 8, after the additive-score equation | Native slide sequence of base score, residual signal, next tree, and updated score; export 2400 × 1040 PNG. | Four panels show a base prediction, its remaining error, a shallow tree fitted to that error, and the updated additive prediction. | Planned—source locked |
| M3.02 | `ch08-fig03-catboost-ordered-target-statistic.png` | Ordered target statistics prevent a row’s own label from leaking into its categorical encoding. | Permutation timeline with prefix-only statistic | Prokhorenkova, Gusev, Vorobev, Dorogush, and Gulin (2018), “CatBoost: unbiased boosting with categorical features,” `https://proceedings.neurips.cc/paper/2018/hash/14491b756b3a51daac41c24863285549-Abstract.html`, as normalized in the masterclass source map; render the ordered prefix formula symbolically with prior P and smoothing a, not sample values. | Literature-derived | Chapter 8, §1 and §3 | Movement III, slide M3-02 | Chapter 8, after §1 | Native slide permutation strip plus typeset ordered-statistic formula; export 2400 × 1350 PNG. | A highlighted categorical row may use matching rows earlier in a random permutation plus a prior, while its own and later labels are visibly blocked. | Planned—source locked |
| M3.03 | `ch08-fig01-boosted-tree-feature-path.png` | A boosted-tree prediction is assembled from explicit feature questions whose validity still depends on the upstream feature contract. | Feature-to-tree path with contract gate | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-08-catboost-deep-dive.html`, §§2–3; feature names are chapter examples only and carry no importance values. | Literature-derived; illustrative rendering | Chapter 8, §3 “Feature engineering is where most hidden assumptions live” | Movement III, slide M3-03 | Chapter 8, before the feature-dictionary paragraph | Redraw the existing original PNG in Graphviz; export 2400 × 1040 PNG; show a feature-version gate before the tree. | Versioned activity, depth, missingness, and context features enter a small decision tree only after a cutoff-validity gate. | Existing—redraw/QA |
| M3.04 | `ch08-fig04-importance-is-not-causality.png` | Feature importance describes model dependence under observed data and does not establish a causal intervention effect. | Forked inference diagram: attribution versus intervention | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-08-catboost-deep-dive.html`, §3 “Importance is not explanation” and §6; no attribution values or causal effects. | Literature-derived; conceptual | Chapter 8, §3 | Movement III, slide M3-04 | Chapter 8, after the importance warning | Native slide causal-claim contrast with a shared feature node; export 2400 × 1040 PNG. | One branch ends at “model used this feature,” while a blocked branch toward “changing it changes the outcome” is labeled unsupported without causal evidence. | Planned—source locked |
| M3.05 | `ch09-fig01-self-supervised-pretraining-flow.png` | Self-supervision creates a training signal from permitted inputs but does not relax temporal evaluation rules. | Pretraining-to-downstream pipeline | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-09-self-supervised-lob-deep-dive.html`, §§1–3; SimLOB reference in `docs/textbook/btc-polymarket-ml/source-map.md`; no pretraining result. | Literature-derived; design-target pipeline | Chapter 9, §§1–3 | Movement III, slide M3-05 | Chapter 9, after the encoder equation | Redraw the existing original PNG in Graphviz with a temporal corpus boundary; export 2400 × 1040 PNG. | A permitted event window feeds a self-supervised encoder, representation, and downstream head; the untouched final period remains outside the pretraining and selection boundary. | Existing—redraw/QA |
| M3.06 | `ch09-fig02-self-supervised-objective.png` | A self-supervised objective is useful only when its transformation preserves the downstream signal without crossing the decision cutoff. | Masked-window objective with loss path | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-09-self-supervised-lob-deep-dive.html`, §2 objective-family table and worked masked-window trace; SimLOB `https://arxiv.org/abs/2406.19396`; masks and tokens are schematic, not sampled data. | Literature-derived; illustrative rendering | Chapter 9, §2 “What self-supervision changes—and what it does not” | Movement III, slide M3-06 | Chapter 9, after the objective-family table | Native slide sequence tensor with hatched masked cells, encoder, reconstruction head, and symbolic mask loss; export 2400 × 1350 PNG. | A time-ordered event window has selected attributes masked; the encoder reconstructs only those attributes while a cutoff line blocks future tokens. | Planned—source locked |
| M3.07 | `ch09-fig03-pretraining-temporal-boundary.png` | Unlabeled future data can still change the claim, so the pretraining corpus must be frozen by time. | Nested temporal partitions with allowed and prohibited pretraining arrows | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-09-self-supervised-lob-deep-dive.html`, §3 “The pretraining boundary”; no dates or corpus sizes. | Design target | Chapter 9, §3 | Movement III, slide M3-07 | Chapter 9, after the conservative teaching-default paragraph | Graphviz or native slide temporal partition map; export 2400 × 1040 PNG. | Training, validation, and untouched final periods appear in order; pretraining arrows stay inside training, selection reaches validation, and no arrow enters the final period. | Planned—source locked |
| M3.08 | `ch10-fig02-lob-tensor-model-inputs.png` | Limit-order-book models differ partly because they preserve different axes and resolutions of the same admissible event history. | Tensor schema comparison | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-10-causal-lob-transformer-deep-dive.html`, §2 design-question table; DeepLOB, TLOB, and LiT references in `docs/textbook/btc-polymarket-ml/source-map.md`; use symbolic dimensions T, L, and C only. | Literature-derived; conceptual | Chapter 10, §2 “What TLOB and LiT-style ideas contribute” | Movement III, slide M3-08 | Chapter 10, before the design-question table | Native slide tensor blocks with axes time T, book level L, and channel C; export 2400 × 1350 PNG. | Three labeled tensor views show tabular compression, a sequence embedding, and a time-by-level-by-channel order-book input without numeric market values. | Planned—source locked |
| M3.09 | `ch10-fig03-causal-attention-tensors.png` | A lower-triangular attention mask blocks future token-to-token paths but cannot block future information already embedded in an allowed token. | Q/K/V tensor flow plus masked attention matrix | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-10-causal-lob-transformer-deep-dive.html`, §1; TLOB `https://arxiv.org/abs/2502.15757` and LiT source cited in `docs/textbook/btc-polymarket-ml/source-map.md`; tensor entries are symbols, not learned weights. | Literature-derived; illustrative rendering | Chapter 10, §1 “Attention is selective memory” | Movement III, slide M3-09 | Chapter 10, after the allowed/forbidden position notation | Native slide tensor tables for Q, K, V and a patterned lower-triangular mask M; export 2400 × 1350 PNG. | Query, key, and value matrices feed an attention matrix whose upper triangle is blocked; a side warning shows a contaminated earlier token bypassing the architectural mask. | Planned—source locked |
| M3.10 | `ch10-fig04-future-token-perturbation-test.png` | A causal implementation must keep an earlier score invariant when only a future event changes. | Before/after metamorphic test | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-10-causal-lob-transformer-deep-dive.html`, §3 and the worked trace; the 100-event/101st-event example is explicitly illustrative and tolerance remains a declared symbol ε. | Illustrative test contract | Chapter 10, §3 “The causal perturbation test” | Movement III, slide M3-10 | Chapter 10, after the invariance equation | Native slide paired pipelines with identical prefixes and one changed future token; export 2400 × 1040 PNG. | Two sequences share all events through the cutoff and differ only afterward; their earlier score boxes are connected by an equality-within-tolerance mark. | Planned—source locked |
| M3.11 | `ch10-fig05-lob-model-comparison.png` | Model complexity is justified only by an out-of-sample contribution under the same temporal, calibration, latency, and replay contract. | Non-ranking comparison matrix and advancement gate | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-08-catboost-deep-dive.html`, §§2 and 6; `docs/textbook/btc-polymarket-ml/full-textbook/chapter-09-self-supervised-lob-deep-dive.html`, §§3–5; `docs/textbook/btc-polymarket-ml/full-textbook/chapter-10-causal-lob-transformer-deep-dive.html`, §§2–5; `docs/planning/btc-polymarket-5m-15m-research-backtest-replay-prompt.md`, “Required model ladder and ablations”; no winner, performance number, or latency value. | Design target | Chapter 10, §5 “Deliberate ensembles and abstention” | Movement III, slide M3-11 | Chapter 10, after §5 | Native slide comparison matrix for CatBoost, frozen SSL encoder, causal sequence model, and ensemble; use checks only for required evidence, never superiority; export 2400 × 1350 PNG. | Four model rows compare retained structure, causal test, calibration, latency evidence, and abstention obligations; an advancement gate requires the same final-period contract. | Planned—source locked |

## Movement IV — Governed deployment: replay, contain, and preserve the evidence chain

By the end of this movement, students should be able to place an LLM, replay engine, cost model, and monitoring guardian inside explicit capability boundaries and reconstruct why the system acted or abstained.

| Figure id | Canonical PNG filename | Single teaching claim | Visual form | Exact data/source basis | Epistemic status | Manuscript anchor | Slide destination | Textbook destination | Generation method | Accessibility description | Status |
|---|---|---|---|---|---|---|---|---|---|---|---|
| M4.01 | `ch11-fig01-llm-code-review-gate.png` | LLM-generated code remains untrusted until a human reviews the diff and named tests produce evidence. | Four-gate code contribution loop | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-11-llm-python-deep-dive.html`, §3 “The four-gate coding loop”; no claim about a deployed assistant or passed test. | Design target | Chapter 11, §3 | Movement IV, slide M4-01 | Chapter 11, after the four-gate equation | Redraw the existing original PNG in Graphviz; contract, patch, review, and test nodes; export 2400 × 1040 PNG. | A one-way loop moves from narrow contract to small patch, human diff review, and test evidence, then branches to accept or reject. | Existing—redraw/QA |
| M4.02 | `ch11-fig02-prompt-injection-trust-boundary.png` | Untrusted repository and web content must never become governing instructions or gain tool authority. | Trust-boundary dataflow with injection attempt | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-11-llm-python-deep-dive.html`, §2 threat table; `chapter-07-reproducible-python-deep-dive.html`, §4; no incident-frequency claim. | Design target | Chapter 11, §2 “The LLM threat model” | Movement IV, slide M4-02 | Chapter 11, after the threat table | Graphviz trust-boundary diagram; label untrusted content as data, governing instructions separately, and tools behind an allowlist; export 2400 × 1350 PNG. | Files, issues, and web text enter a data channel; an injection arrow is stopped at the instruction boundary, while bounded tools require a separate allowlist and review. | Planned—source locked |
| M4.03 | `ch12-fig01-event-driven-replay.png` | Event-driven replay exposes only evidence available at each simulated time and records abstentions as well as hypothetical actions. | Replay event loop with append-only ledger | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-12-replay-paper-trading-deep-dive.html`, §1; `docs/planning/btc-polymarket-5m-15m-research-backtest-replay-prompt.md`, “Event-driven backtest and replay requirements”; no replay result. | Design target | Chapter 12, §1 “A backtest is an executable claim about time” | Movement IV, slide M4-03 | Chapter 12, after the replay pipeline equation | Redraw the existing original PNG in Graphviz as an event loop; export 2400 × 1040 PNG. | Ordered raw events advance a clock through feature, model, policy, hypothetical fill, settlement, and append-only ledger stages, with abstention retained as an output. | Existing—redraw/QA |
| M4.04 | `ch12-fig02-replay-clock-separation.png` | Replay validity requires event, receive, decision, and settlement clocks to remain distinct. | Four-lane sequence diagram | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-12-replay-paper-trading-deep-dive.html`, §2 clock table; no delay values. | Design target | Chapter 12, §2 “The three clocks” (including the separately listed settlement time) | Movement IV, slide M4-04 | Chapter 12, after the clock table | Graphviz sequence diagram with four lanes; export 2400 × 1040 PNG. | A market event occurs, arrives later, becomes a decision after processing, and settles last; arrows make the causal order explicit without assigning durations. | Planned—source locked |
| M4.05 | `ch12-fig03-cost-sensitivity-surface.png` | Small forecast differences are not decision-relevant when plausible cost and latency assumptions erase conservative value. | Symbolic two-dimensional cost surface with abstention contour | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-12-replay-paper-trading-deep-dive.html`, §§3 and 5; `docs/planning/btc-polymarket-5m-15m-research-backtest-replay-prompt.md`, execution-realism and required-figures sections; compute only the symbolic surface `d = p_LCB − q_exec − fee − slippage − latency_haircut`, with no fitted parameters or PnL. | Illustrative; formula-derived | Chapter 12, §5 “Sensitivity analysis and paper-trading readiness” | Movement IV, slide M4-05 | Chapter 12, after the sensitivity-analysis opening paragraph | Vega-Lite heatmap from normalized symbolic parameter ranges; label axes “assumption severity,” not dollars or basis points; draw the d=0 abstention contour; export 2400 × 1350 PNG. | A heatmap transitions from eligible to abstain as combined execution cost rises and the lower-confidence-bound edge falls; a bold zero-value contour separates the regions. | Planned—source locked |
| M4.06 | `ch12-fig04-decision-ledger-anatomy.png` | A replay claim is auditable only when one decision row binds frozen inputs, artifacts, gates, assumptions, and later outcome. | Exploded ledger-record schema | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-12-replay-paper-trading-deep-dive.html`, §4; `docs/planning/btc-polymarket-5m-15m-research-backtest-replay-prompt.md`, “Deterministic replay”; fields are contract fields, not a real record. | Design target | Chapter 12, §4 “The immutable replay ledger” | Movement IV, slide M4-06 | Chapter 12, after the ledger-field paragraph | Native slide exploded-record diagram with grouped fields for provenance, model, policy, execution assumptions, and settlement; export 2400 × 1350 PNG. | One decision record expands into raw-input hash, code and configuration hashes, feature and model versions, probability, gates, abstention or hypothetical fill, costs, and later settlement. | Planned—source locked |
| M4.07 | `ch13-fig01-guardian-state-machine.png` | A guardian may contain an unsafe research state only through verified, reversible, allowlisted transitions. | Finite-state machine with fail-closed transitions | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-13-agentic-monitoring-deep-dive.html`, §§1 and 4; states derive from observe, verify, contain, verify containment, record, escalate; no deployed-agent claim. | Design target | Chapter 13, §1 “Monitoring is a control loop, not a chat window” | Movement IV, slide M4-07 | Chapter 13, after the bounded-cycle equation | Graphviz finite-state machine; every transition labeled with predicate or verification result; export 2400 × 1350 PNG. | The guardian moves from observing to verifying, containing, verifying containment, recording, and escalating; incomplete verification routes to a safe paused state. | Planned—source locked |
| M4.08 | `ch13-fig02-containment-remediation-boundary.png` | Containment limits use of suspect evidence, while remediation changes the system and therefore requires human approval. | Two-column capability boundary with handoff | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-13-agentic-monitoring-deep-dive.html`, §4 “Containment is not remediation”; examples are the chapter’s named actions. | Design target | Chapter 13, §4 | Movement IV, slide M4-08 | Chapter 13, after the containment/remediation comparison | Native slide two-column flow with an explicit human-approval handoff; export 2400 × 1040 PNG. | Reversible actions such as quarantine, pause, and block report appear on the guardian side; schema approval, model deployment, and threshold change sit behind a human gate. | Planned—source locked |
| M4.09 | `ch13-fig03-monitor-authority-matrix.png` | Capability removal is a stronger safety boundary than a prompt telling an agent not to trade. | Allowed/prohibited capability matrix with structural barrier | `docs/textbook/btc-polymarket-ml/full-textbook/chapter-13-agentic-monitoring-deep-dive.html`, §2 capability table; `docs/textbook/btc-polymarket-ml/full-textbook/safety-contract.md`; no implemented identity or credential audit. | Design target | Chapter 13, §2 “The authority boundary” | Movement IV, slide M4-09 | Chapter 13, after the capability table | Native slide matrix with text labels, check/cross shapes, and a heavy capability boundary; export 2400 × 1350 PNG. | Reading health metadata, quarantining, pausing, and rerunning validation are marked allowed; approving changes and all order-control capabilities are structurally blocked. | Planned—source locked |
| M4.10 | `ch13-fig04-final-evidence-chain.png` | A defensible prediction claim survives only when every link from source truth to human-governed outcome remains traceable and fail closed. | End-to-end evidence chain with audit and stop branches | The exact 13-file deep-dive source set enumerated above; `docs/textbook/btc-polymarket-ml/full-textbook/source-map.md`; `docs/textbook/btc-polymarket-ml/full-textbook/safety-contract.md`; `docs/planning/btc-polymarket-5m-15m-research-backtest-replay-prompt.md`; no performance or deployment claim. | Design target; synthesis | Chapter 13, §6 “The final lesson” | Movement IV, slide M4-10 | Chapter 13, immediately before review and discussion | Graphviz end-to-end chain: settlement contract, clocks, raw evidence, features, temporal evaluation, model, calibration, policy, replay, ledger, guardian, human; add named fail-closed exits; export 3200 × 1350 PNG. | A twelve-stage chain runs from settlement truth through governed human review; each major stage has a downward branch to abstain, quarantine, block, or escalate when its evidence fails. | Planned—source locked |

## Release gate for every figure

A figure may move from this manifest into the masterclass or textbook only when its single claim, source basis, and epistemic label match the row above. The PNG must exist under `assets/`, preserve a white manuscript canvas, remain legible at both slide and inserted-page size, include non-color cues, omit a baked-in caption or figure number, and pass visual inspection for clipping, overlap, ambiguous arrows, and misleading precision. Any future measured replacement requires a dated data manifest, reproducible command, configuration and code hashes, named output artifact, sample definition, and limitation; until then, the approved figure remains illustrative, literature-derived, or a design target exactly as recorded here.