Find your next good question

Three ways through the book

For a first visit, read pages 4–10, choose a pocket map from the atlas, then work three calling-price drills. At the live table, the study cue links to examples from your current street. You can browse at your own pace and keep a separate note about the decision you would choose.

For strategy practice, choose one opponent and one street. Cover the forecast table with a sheet of paper, name your preferred action and explain its price. Then reveal the model rankings. The useful comparison is between the assumptions behind your choice and the information represented by the model.

For experiment review, start at page 292. The original training run and the fresh evaluation are separate receipts. The controlled ablation asks what the zone representation added; the served-checkpoint result asks how the deployed player performed. Both are part of the story.

What this edition stands on

The model is real WSRF: pytoolzx.wsrf.WSRFRegressor, fitted to completed-return targets and exported for browser inference. Model predictions, simulator parity checks and final match results have separate artifacts. The browser serves the selected checkpoint with the recorded hash; its public HUD observations update as you play.

The fresh evaluation was registered before running its seeds. It compares the served checkpoint, the final-round zone fit, the zone-free fit and a conventional reference. Opponents use the same public-information rules and action engine. Every match block balances both seats and both stack depths, and begins with fresh observation priors.

The primary contrast is the final-round zone model against the zone-free model trained on the same replay rows with equal total tree count. Removing distance features and zone routing together measures that combined representation choice. It does not isolate every possible use of the author’s method. Different fits and future learning designs can be evaluated separately.

Served-model SHA-256
1c50583926cc485d9ee009742d58b503af8774d450d7613d7131d3fbba77fe3d
Full receipts: training/report.json and training/extension.json. Study snapshots: training/studies.json.

Three kinds of useful knowledge

A good field guide lets you tell a definition, a measurement and an experiment apart. Here, a definition tells you exactly how to calculate a distance map. A measurement tells you what a frozen player did against a named opponent under recorded conditions. An experiment gives you an adjustment to try next, with a way to keep score. Each can be useful without borrowing the authority of the others.

The likely and unlikely labels in this edition have a specific meaning: they describe how widely the near region covers thirteen ranks. They are the author’s compact language for a structure in the cards. Confidence in an action is a separate question. A common zone can still contain a rare betting sequence, an unfamiliar stack depth or a poorly estimated opponent range. Keep both ideas available rather than merging them into a single green light.

Read a worked decision in three passes. First, hide the recommendation and describe the public situation in your own words. Next, calculate the price and name the zone. Finally, compare the forest’s alternatives and ask what would have to be true for each to earn its cost. This turns an answer into a reusable decision process. The best study result is a clearer explanation you can carry to another hand.

The 120 decision studies are selected by street, opponent and dynamic zone before outcomes. They contain actual predictions from the served WSRF model, with a larger equity sample than the live table. They are examples of the model’s reasoning, rather than independent proof that every recommendation is optimal. The fresh-match evaluation later in the book addresses performance; the studies help you inspect the machinery.

Worked question: Can a zone map, a price calculation and an empirical result all inform the same decision while keeping their meanings distinct? Yes. Write each beside its own label. That habit preserves the imaginative usefulness of the method while allowing new evidence to improve it.

Build the thirteen-rank compass

Start with the rank sequence 2, 3, 4, 5, 6, 7, 8, 9, T, J, Q, K, A. Assign coordinates zero through twelve. For two reference ranks a and b, form the set containing a, b, their sum modulo thirteen, and their absolute difference. Repeated references count once. The atlas prints these references for every unordered pair of ranks, so you can check the entire construction by hand.

For any target rank r, measure the shorter circular distance to each reference. The distance between r and t is the smaller of their ordinary absolute difference and thirteen minus that difference. Keep the smallest distance across the reference set. Distance zero marks a reference itself. Distance one marks a neighbor on the circle. Distances two and above describe progressively farther locations in this particular map.

The circle is a chosen coordinate system. An ace sits next to a deuce in this construction because the endpoints wrap. That convention is different from the rules used to score poker straights, where the ace can complete a wheel or an ace-high straight under specific conditions. The hand evaluator and the zone map therefore remain separate functions. Each has a clear job.

Try a small example with the deuce and trey. Their coordinates are zero and one. Sum and difference both contribute one, leaving two distinct references. The near ranks are A, 2, 3 and 4. That gives four near ranks out of thirteen. Now change only the trey to a six and recompute every reference. A one-card change can reshape several regions because the derived sum and difference move together.

Worked question: Can you reconstruct every printed zone from the two ranks and this rule? Yes. The atlas includes all ninety-one unordered rank pairs. Suits affect poker equity and physical card combinations, while this reference construction uses ranks. Keep a separate suit annotation beside your map when studying a real hand.

Give each zone a precise meaning

Count how many of the thirteen rank positions have distance zero or one. Seven or more near ranks make the map LIKELY. Six or fewer make it UNLIKELY. The threshold is a rank-coverage rule: at least seven thirteenths. It follows the executable source used in this project, which was checked across all 169 ordered input pairs. This book uses that single rule consistently.

The label compresses the shape of the map into one bit, but the forest also receives richer features. It can inspect pocket coverage, dynamic coverage, four distance-bucket proportions, a distance mean, a distance standard deviation and whether the zone changed. A binary name is a useful doorway into the representation, while the underlying numbers retain detail that the name alone cannot carry.

Three percentages deserve separate spaces in your notes. Rank coverage counts near rank positions out of thirteen. Unseen-card near frequency counts physical unseen cards in those positions. Range equity estimates your share of showdowns against a modeled opponent. They use different denominators and answer different questions. A high value in one column can coexist with a lower value in another.

This matters when adjusting commitment. Your proposed approach seeks more productive large commitments and smaller losses in unfavorable situations. Translate that aim into tests that combine zone, price, range strength and action size. A zone-only all-in rule would discard information the trained player already has. The richer experiment asks whether distance features improve an otherwise comparable decision system.

Worked question: Can an UNLIKELY zone still accompany a strong hand and a profitable value opportunity? Yes. The word describes the map’s rank coverage. The actual cards, opponent range and available price determine the opportunity to be evaluated. Treat an unlikely map as a prompt to inspect carefully, and let the complete decision earn its stake.

Two maps, one developing hand

The pocket map begins with your two hole-card ranks and keeps those reference ranks throughout the hand. Its physical unseen-card counts change as community cards arrive, but its thirteen-rank coverage remains fixed. This makes it a useful label for whole-hand reporting: every hand enters one pocket group before the first betting decision.

The dynamic map uses your hole cards before the flop. From the flop onward it uses the final two public board cards in dealing order. On the flop those are the second and third flop cards; on the turn they are the third flop card and the turn; on the river they are the turn and river. This exact ordering is part of the method implemented here. A sorted picture of the board would lose that information.

Both maps remove your hole cards and all public board cards when counting unseen physical cards. Neither removes a secretly revealed opponent pocket from the policy inputs. The peek option is a practice display. Keeping the counting boundary fixed ensures that the coach, the model and the conventional opponents receive the information they would have had at the decision.

A transition is a descriptive event: the pocket and dynamic labels differ, or the dynamic map changes between streets. It can help organize study. It does not by itself tell you whether the new board helped your hand, helped your opponent or changed the profitability of a raise. Recalculate those questions using suits, made hands, draws and the public betting line.

Worked question: Can you preserve a stable whole-hand label while also following the changing board? Yes. Record pocket zone once, then dynamic zone at every decision. When reviewing results, keep hand-level returns attached to the pocket label and action-level commitment attached to the dynamic label. That avoids counting the same entire hand result repeatedly as if it were several independent observations.

Count the cards still available

With two hole cards and no board, fifty physical cards remain unseen to you. On the flop the public-information pool contains forty-seven cards; on the turn forty-six; on the river forty-five. These counts include hidden opponent cards and burn cards because their identities are unavailable. The policy samples possible worlds compatible with the visible information instead of using the actual hidden deal.

For a simple uniform next-card calculation, let N be the unseen pool and K the cards classified near. Then K divided by N is the near frequency. For two distinct cards sampled without replacement, the chance of two near cards is K(K−1) divided by N(N−1). The chance of at least one near card is one minus (N−K)(N−K−1) divided by N(N−1). These are exact counting statements under the uniform-pool assumption.

Betting information changes the interpretation. A range model can assign different probabilities to hidden opponent pockets. Once those pockets are weighted, future-board probabilities may change through card removal. The atlas deliberately reports the simpler uniform unknown-pool calculation and labels it that way. The equity estimator separately uses smoothed public-action-conditioned pocket proposals.

A card can be near in the zone map and still improve your opponent more than you. Conversely, a far card may give you the winning hand. Distance buckets describe where cards sit relative to references; poker outs describe cards that achieve a particular hand or winning condition. When you count outs, specify the condition and account for redraws, ties and opponent improvements.

Worked question: Can exact removal arithmetic make the zone map more informative without treating every near card as a winning out? Yes. Put the bucket probability beside a separate hand-improvement analysis. This lets the compressed map retain its meaning while the ordinary card accounting supplies the detail needed for an action.

Distance, spread and information

The mean distance is the card-count-weighted average of the distance values. The distance standard deviation describes how far those values spread around their own mean. Both are properties of the chosen map and the visible-card removal state. Standard deviation here is measured in rank-distance units. It is separate from the standard deviation of chip returns measured in big blinds.

The atlas also summarizes the four coarse buckets: zero away, one away, two away and three-or-more away. Their entropy is minus the sum of each bucket probability multiplied by its base-two logarithm. A concentrated bucket distribution has less bucket uncertainty than a balanced distribution. The calculation describes the bucket label of a uniform unseen card, rather than an ability to identify the card itself.

For example, learning that a card belongs to a large near bucket still leaves many identities possible. The representation compresses an answer to a particular question. It can be useful if that answer helps a decision maker generalize across related states. The usefulness of the representation is measured by task performance, while its accounting remains consistent with the information retained.

Chip returns need their own distribution. A policy can win many small pots and lose occasional large ones, or take the opposite shape. Its mean, spread, drawdown and commitment patterns tell different parts of that story. A smooth bell-shaped drawing is a helpful illustration only when its assumptions match the quantity being plotted. This guide therefore prints empirical block results and intervals rather than imposing a bell curve on every decision.

Worked question: Can a compact distance vocabulary help a model organize decisions while preserving clear uncertainty about the next card? Yes. That is a representation-learning hypothesis that we can implement and test. The controlled zone comparison asks whether this particular representation improved returns in this particular player and environment.

Let the price sharpen the read

At your turn, write the pot already in the middle, including the opponent’s current bet, as P. Write the additional amount needed to call as C. If calling closes the action and realizes all remaining equity, the break-even share is C divided by P plus C. Calling wins the existing pot on a win and spends C on a loss. This gives expected incremental value e(P+C)−C when e is your expected showdown share.

This formula is exact for the stated simplified payoff, including split-pot equity if e counts a tie as the appropriate share. On earlier streets, calling may lead to further bets and selective folds. Then raw showdown equity is only one ingredient: position, future prices, realization and the opponent’s continuing behavior matter. Use the price as a clear starting point, then describe the continuation.

Keep already-paid chips separate. A fold gives zero additional return from the present choice because earlier contributions are sunk. The training target follows this decision-time convention. Whole-hand profit uses the original starting stack, so a fold can have a negative whole-hand result while remaining the best available action at the current price.

Apply the same arithmetic inside both zone labels. A likely map with a poor price can invite patience. An unlikely map with a favorable price can invite continuation. The interesting zone policy is one that learns a useful adjustment after accounting for this shared baseline. It should earn its extra complexity by improving decisions or explaining them more clearly.

Worked question: Can a fold preserve future opportunities even after you have invested chips in the hand? Yes. Compare the options available now, using the remaining price and the range you face. Record the whole-hand result separately so your session accounting stays complete. The price drills later in the book make this distinction concrete.

Size the decision you actually face

A displayed raise-to amount includes your chips already placed on the current street. The additional cost is the target minus your current street bet. A call also costs only the unmatched difference, capped by your remaining stack. These distinctions prevent a small opening bet from being confused with a full-stack commitment.

Effective stack is the amount the two players can still contest. If you have more chips than your opponent, the excess cannot be won from that opponent during this hand. The engine returns uncalled excess through pot settlement. The model’s commitment features therefore look at both stacks and current street contributions rather than reading your stack in isolation.

Stack-to-pot ratio is a compact way to describe how much betting room remains. A forty-big-blind start and a hundred-big-blind start can lead to the same made hand but different decisions. With a shallow effective stack, a large bet may nearly close the action. With deep stacks, the same fraction of the pot can leave several meaningful future decisions.

The lab explicitly balances both button positions at both starting depths. That balance matters because an accidental pairing of deep stacks with one seat could make a model appear to learn stack strategy when it was actually learning position. Independent evaluation rotates the button on repeated decks and reports the four seat-depth cell counts.

Worked question: Can you compare commitment across different pots by expressing additional cost relative to the contestable money? Yes. Record both the actual big-blind cost and the normalized fraction. The first keeps consequences tangible; the second lets a forest compare related decisions at different scales. A model estimate remains a forecast, even when the units are carefully matched.

Make a large commitment earn its place

An all-in combines at least two outcomes: the opponent folds, or the opponent continues with some range. The profitability of the shove depends on both. Equity against the opponent’s starting range may differ sharply from equity against the hands that call a large bet. Treat the continuing range as a fresh question rather than carrying an opening equity estimate forward unchanged.

For an unopened river bet of B into a pot P, with matching stacks and called equity e, the simplified incremental value is fP + (1−f)[e(P+2B)−B], where f is the fold frequency. This expression assumes the bet closes the action after a call, a matching call of B, and no rake. A raise over an existing bet needs its own accounting of the extra call and raise amounts.

Your zone hypothesis asks whether stronger commitment in selected likely situations and smaller exposure in selected unlikely situations can improve the overall return distribution. Test that as an action policy, not merely as two outcome buckets. Hands entering the two zones have different card compositions. A positive likely-group return by itself does not tell us what another action would have earned.

A useful study record contains the public state, both zone labels, each legal action, the forecast for each action, the chosen cost and the eventual complete-hand result. A stronger causal comparison uses independently tested policies or counterfactual continuations under a specified simulator. This edition reports policy comparisons and labels model rankings as forecasts.

Worked question: Can a confident large commitment be justified by several agreeing signals rather than a zone name alone? Yes. Check the continuing range, price, effective stack, fold opportunity and model stability. Then use fresh results to judge the rule. The aim is purposeful aggression whose costs and rewards can both be explained.

Build value with a continuation plan

A value bet seeks calls from a range that gives you favorable expected returns. Start by naming plausible worse hands that continue and better hands that also continue. A board that looks safe for your current pair may still change which portion of the opponent’s range pays a bet. A larger size can earn more per call while narrowing the calling range.

Protection is the value of changing future participation and prices. A bet can deny some hands a free opportunity to improve, but that alone does not establish the best size. Compare the extra folds, the calls you welcome, the calls you dislike and the options you retain on the next street. A check can also preserve profitable future action.

Use the dynamic zone as a compact annotation of the new public board. Then add suit concentration, connected ranks, paired boards and the interaction with your hole cards. The model receives features for those ordinary poker structures alongside distance features. The guide’s map is an additional lens within a richer state description.

The live model has a finite set of action presets: fold when facing a price, check or call, minimum raise, approximately half-pot raise, pot raise and all-in when legal. Human players can choose other legal sizes with the slider. Predictions printed for the presets do not automatically price every intermediate slider position. A new size asks a new question about opponent responses.

Worked question: Can checking a strong hand belong to a thoughtful value plan? Yes. Explain the future calls or bluffs it may preserve, then compare that plan with immediate betting. The best explanation connects this street to the next decision. Save the public line in your hand history so later review can examine the entire plan.

Price the story your bluff tells

A pure river bluff of B into an existing pot P breaks even at a fold frequency of B divided by P plus B, assuming the bluff wins nothing when called and the bet closes the action. A half-pot bluff therefore needs one-third folds under those assumptions. A pot-sized bluff needs one-half. The threshold is a starting calculation, not a prediction of the opponent’s response.

To estimate response, describe the range arriving at the river. Which hands reached this point by checking, calling or raising? Which of those can comfortably continue for the chosen size? Your own cards can remove combinations from that range. A useful blocker changes the relative availability of relevant calling or folding hands; a zone distance by itself is a different description.

A semi-bluff has possible equity when called as well as immediate fold value. Its future continuation also matters. A draw that can make a second-best hand has a different payoff distribution from a draw that tends to make the strongest hand. Counting nominal improvements is only the beginning of the comparison.

Fern and Rook both mix value and bluffs, with different thresholds and frequencies. Their actions use public history, range equity, price and position. Their code deliberately excludes the distance-zone columns. This gives the trained zone player conventional sparring partners while keeping the experiment inspectable. Their strength remains local to this test environment.

Worked question: Can a carefully chosen small bluff outperform a dramatic shove? Yes. The comparison depends on folds earned per chip risked and on the value of getting called. Calculate the threshold for each size, name the response assumption and test it. The drill pages ask affirmative questions about correct calculations rather than asking you to endorse every aggressive action.

Study Fern’s selective pressure

Fern is the tight-aggressive reference opponent. Relative to Rook, Fern needs stronger estimates for several aggressive actions and uses lower bluff frequencies. Fern still responds to price, position and public betting history; tight describes a relative policy tendency rather than a promise that every bet means a premium hand.

Begin by observing which opportunities produce folds, calls and raises. VPIP measures voluntary preflop participation; PFR measures preflop raising. Their gap can suggest how much participation comes through calls. The displayed statistics use prior counts so the first few hands do not swing the estimates to extremes. An early percentage is partly a prior, and the hands-seen display helps you interpret it.

Against a selective continuing range, a hand that dominated the opening range can face a much closer decision after a large raise. Revisit the range instead of defending an earlier conclusion. A cheaper continuation may preserve flexibility; a fold may avoid a poor additional investment; a strong value hand may welcome the narrower range. The public line distinguishes these situations.

The first completed experiment gave the served zone player a positive point estimate against Fern with an interval spanning zero. Its controlled final-round ablation favored the zone-free fit against Fern. Those results make Fern an especially useful training target. The extended evaluation uses fresh seeded blocks to examine whether the pattern persists, rather than treating more hands as an automatic improvement.

Worked question: Can a difficult matchup show us where the next training improvement should focus? Yes. Group decisions by street, price, stack depth and zone, then compare returns with the same information available to competing models. Let a candidate earn promotion on separate validation data and reserve new matches for its final examination.

Study Rook’s wider pressure

Rook is the loose-aggressive reference opponent. Wider participation and more frequent bluffs put more pressure on passive responses. Wider also means a continuing range can contain more hands you beat. The productive adjustment depends on which branches are wide: opening, calling, raising and bluffing can each have different distributions.

A simple reaction of calling every bet gives aggression too much control over price. A simple reaction of raising every likely map discards range strength and commitment cost. Instead, compare the actual offered price with the hands represented by the line. Identify where Rook’s pressure adds bluffs and where a repeated raise filters the range toward value.

The original fresh-deal test gave the selected checkpoint a positive interval against Rook. The controlled final-round zone ablation against Rook was less decisive. These are different findings. One concerns the served policy’s performance; the other asks what the distance features and zone routing added to a comparable fit. Keeping both visible makes the result more informative.

Try a review session with ten hands chosen before seeing their outcomes. Write one sentence about the opponent range and one about your commitment. Then reveal the result and add a separate outcome sentence. This protects useful learning from the temptation to rewrite the original read after a surprising river. A memorable loss can become an informative example without dominating the entire dataset.

Worked question: Can patience and selective aggression work together against a wider player? Yes. Patience means choosing a justified price, while aggression means charging or pressuring the range when the opportunity supports it. Use the dynamic map as one feature in that explanation and keep the complete line available for review.

Measure the confidence you mean

The table’s equity estimate comes from sampled opponent pockets and board completions. More samples reduce one source of Monte Carlo noise. They do not automatically repair an inaccurate opponent-range assumption. The worked studies use 8,192 samples per state to make the illustration steadier, while the live game uses 128 for responsive play.

A forest’s disagreement between trees is another quantity. It can reveal variation among fitted predictors, but this implementation does not calibrate that spread into a probability that the top action is correct. The book therefore labels predictions as estimated incremental returns and treats action gaps as model preferences. A large gap is useful to inspect, rather than a certified confidence bound.

Each worked study also shifts its equity input by four percentage points in both directions and ranks the same actions again. This is a local input-sensitivity probe. It holds other features fixed, so it is neither a new dealt hand nor a confidence interval. When the favorite action changes, the example deserves closer inspection of the equity and range assumptions.

Performance uncertainty belongs to completed matches. The extended test starts independent match blocks with fresh HUD priors, then summarizes their average returns with Student t intervals across blocks. This captures block-to-block variability more directly than treating every action as independent. Eight blocks still leave uncertainty, and the sensitivity pass uses only four blocks at the larger sampling budget.

Worked question: Can we gain confidence by asking which source of uncertainty is being reduced? Yes. Separate sampling noise, range misspecification, model fit and match variation. A well-targeted experiment can improve one of them without pretending that all the others disappeared. That distinction keeps the next training investment purposeful.

Let adaptation follow recorded outcomes

The training run uses the actual pytoolzx WSRFRegressor. It fits forests specialized by street and dynamic zone. Every retained training row includes public decision features, an action description and a completed return measured from immediately before that action. Guide labels organize the representation; they are not substituted for outcomes.

Six rounds each generated 500,000 complete hands. Early rounds explored more often, and later rounds reduced exploration while using the current validated champion. The replay reservoir retained a bounded sample from recent rounds. The next round allocated more exposure to the harder opponent identified by validation. This is an explicit offline adaptive loop.

A new fit can be worse than an earlier fit. The selected checkpoint in this run was round three, trained through 1.5 million hands, even though the complete experiment continued to three million. That distinction is printed throughout the book and website. More experience enlarges the experiment; promotion still depends on the selection rule.

The browser serves a frozen model. It updates public opponent statistics during the session but does not refit WSRF in the page. Saving the session preserves your play and observations. A future training revision can use a separately curated dataset and new evaluation plan. Keeping these stages explicit prevents a changing interface from being mistaken for a newly validated model.

Worked question: Can we keep learning from a longer run while serving an earlier checkpoint that earned the best validation score? Yes. Preserve every checkpoint and its receipt, then compare new candidates under a fixed protocol. That creates room for ambitious experimentation while giving the deployed player a clear identity.

Turn a hand into a useful lesson

Before looking at the outcome, record the visible cards, stack depth, seat, public betting line, pocket zone and dynamic zone. Add the offered price and the action you prefer. A single sentence explaining why is enough to make your thinking reviewable. The hand-history download supplies much of the public sequence; your sentence adds the hypothesis.

After the hand, record the complete-hand return in a separate field. Then compare your action with the forest’s rankings. A disagreement is an invitation to inspect assumptions. The forest may have a useful pattern, or it may be poorly supported in this state. The outcome alone settles neither question. A winning call can still be expensive on average, and a losing value bet can still have been well chosen.

Cheap calls sometimes reveal an opponent’s hand, but a call on an early street does not guarantee a showdown. Information also arrives through folds, sizes and later public actions. In a controlled experiment, include the cost of exploration in the full return, and define whether opponent adaptation or repeated encounters give that information future value.

Choose review samples deliberately. A balanced notebook includes quiet folds, ordinary calls, successful value bets and dramatic confrontations. If you save only memorable losses, the notebook will overrepresent one slice of play. For performance claims, use the complete scheduled experiment rather than a collection of favorite examples.

Worked question: Can an uncomfortable result become a constructive search prompt without deciding its explanation in advance? Yes. Preserve the decision-time record, ask what alternatives were available and use comparable data to evaluate them. The process gives curiosity somewhere concrete to go: a clearer range, a better price comparison or a candidate policy to test.

Keep curiosity in the driver’s seat

This guide uses worked questions whose truthful answer is Yes. The purpose is to invite a constructive action: calculate, compare, inspect, preserve a useful distinction or try a reproducible exercise. A positive question can make a difficult concept approachable while keeping the answer grounded in the situation described.

For example, ask whether a smaller commitment can retain flexibility when the evidence is mixed. Ask whether a likely map can be combined with price and range strength. Ask whether a fresh test can challenge a favorite idea productively. Each question points toward something you can do. It leaves space for the result to be surprising.

At the table, use the coach as a study companion. Read the short cue, form your own plan and open the action rankings when you want more detail. The linked study pages let you move from a live street to related worked examples. The model’s action remains visible as a recommendation with an estimated value, so human judgment has something concrete to inspect.

For a relaxed session, choose one theme and play ten hands. A useful theme might be naming the dynamic references, distinguishing call cost from raise target, or describing Fern’s continuing range after a raise. Keep the session total visible without making it the only measure of progress. A more precise explanation is a real learning outcome, even in a short sample with noisy returns.

Worked question: Can an imaginative method grow stronger through both enthusiasm and careful records? Yes. The two support each other. A clear definition makes an idea executable; a transparent experiment gives it feedback; a friendly interface makes people want to explore it. That is the spirit of this field guide and the next round of the lab.