Tiny GPT From Scratch

Formulate the problem before choosing the machinery

Tiny GPT From Scratch begins with a decision problem, not a framework. Grow a character-level language model from a bigram baseline into a working GPT with multi-head attention and sampling. Restate that sentence as an observable input, a desired output, and a criterion for preferring one output over another. Identify who or what supplies supervision, whether feedback is immediate or delayed, and whether examples can be considered independent. These choices determine what can be learned and what remains an assumption. The implementation is honest only when those assumptions are visible near the data contract rather than buried in training code.

The raw material becomes a token representation. Representation decides which distinctions the system can express and which distinctions disappear. List categorical domains, numerical units, missing-value semantics, sequence or spatial axes, masks, player or client perspective, and precision. Then consider invariances: should translation, permutation, rescaling, token position, client identity, or board symmetry change the answer? An architecture that ignores the required invariance wastes data; one that imposes the wrong invariance makes the target impossible to represent.

Finally define the baseline and the abstention point. A baseline can be a constant predictor, random policy, linear rule, naive kernel, synchronous algorithm, or human heuristic. It anchors complexity in evidence. The abstention point describes inputs for which the system lacks support and should decline, defer, or fall back. Together they prevent Tiny GPT From Scratch from being judged only by an impressive end-to-end demonstration while basic correctness, calibration, robustness, or operational usefulness remains unknown.

Connect the objective to the behavior you actually want

An objective compresses preferences into a scalar, but no scalar captures every product or scientific goal. For Tiny GPT From Scratch, distinguish the training objective from the evaluation metric and the deployment utility. The training objective must provide a usable signal to parameters or state; evaluation must estimate generalization under a controlled protocol; deployment utility includes latency, cost, safety, and the consequence of errors. When these three disagree, optimization can succeed while the system becomes less useful.

Study each term dimensionally and statistically. Ask what happens if one term is multiplied by ten, one class becomes rare, a sequence becomes longer, a client contributes more samples, or rewards are shifted. Determine whether averages are per token, example, client, action, spatial position, or batch. Regularization is not decorative: it encodes a preference over solutions and changes units unless normalized consistently. A correct derivation names the population quantity of interest, its finite-sample estimator, and the approximation introduced by minibatches, replay, sampling, or surrogate losses.

Identifiability is the deeper constraint. Data may not contain enough information to separate competing explanations. Causal attention, Adam, Sampling can improve computation or inductive bias, but they cannot manufacture missing evidence. State causal assumptions, observability limits, support conditions, and equivalence classes of solutions. Use sensitivity analysis and targeted interventions where possible. When identification is impossible, report uncertainty or a set of plausible answers rather than converting an arbitrary modeling choice into unwarranted confidence.

Make mathematical equivalence survive finite precision

Paper algebra assumes exact real numbers; the implementation uses finite precision, bounded memory, and discrete execution order. In Tiny GPT From Scratch, audit exponentials, logarithms, divisions, reductions, norms, probabilities, recursive values, and accumulated updates. Rewrite unstable expressions with max subtraction, log-sum-exp, compensated accumulation, safe denominators, or higher-precision reductions. Track where a mathematically harmless reordering changes rounding and where mixed precision needs scaling or master copies.

Shapes are part of the proof. Annotate each intermediate with semantic axes rather than only dimensions: batch, token, head, channel, client, action, expert, feature, row, column, or sample. Broadcasting should be intentional and verified with asymmetric dimensions so an accidental match cannot hide. Record contiguous layout and stride assumptions when performance code depends on them. For every reshape or transpose, write both the precondition and the inverse operation needed during backward, decoding, aggregation, or reconstruction.

Build a numerical ladder: scalar example, tiny vector or matrix example, batched reference, optimized path, then realistic workload. At each rung compare values and invariants before increasing scale. This catches defects while they are still interpretable. The acceptance test should specify absolute and relative error, exceptional values, deterministic modes, and the hardware or library versions used. Numerical stability is not a final cleanup task; it is part of the algorithm’s definition.

Design evidence that can falsify the implementation

Evaluation is an experiment. For Tiny GPT From Scratch, specify the unit of analysis, split strategy, temporal boundary, randomization, baseline, metric, and uncertainty before viewing final results. Prevent duplicates, transformed copies, future information, opponent leakage, and shared-client information from crossing the boundary. A single aggregate score can hide subgroup collapse, unstable seeds, poor calibration, tail latency, or rare catastrophic behavior, so pair it with distributions and stratified slices.

Ablations connect outcomes to mechanisms. Remove or replace Causal attention, Adam, Sampling one at a time while controlling data, compute, and evaluation. Compare equal wall-clock or equal resource budgets when efficiency is part of the claim. Repeat stochastic runs and report variation rather than selecting the best seed. Inspect learning curves and intermediate metrics because two systems with the same final score may differ radically in sample efficiency, stability, or cost.

The test suite and the benchmark answer different questions. Unit and property tests prove local contracts; integration tests prove components agree; benchmarks estimate behavior at scale; task evaluation estimates usefulness. Preserve all four. A benchmark that bypasses validation or uses a different code path from production is weak evidence. The strongest release gate reruns the exact packaged implementation with recorded configuration and produces an artifact that another person can inspect.

Turn the learning artifact into an operable system

Production structure separates pure computation from orchestration, configuration, persistence, and interfaces. Package the core of Tiny GPT From Scratch behind typed contracts. Keep data loading, model or state construction, training, evaluation, serialization, and serving independently invocable. Configuration should be validated, versioned, and printable. Random seeds, data identifiers, source commit, dependency lock, hardware, and metric definitions belong in the run record so an apparent regression can be reproduced instead of guessed at.

Capacity planning follows the critical path. Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage across representative input sizes and concurrency. Report warm-up separately, distinguish throughput from latency, and include tail percentiles. Define memory ownership and lifetime so caches, activations, buffers, replay, or optimizer state cannot grow without a bound. Backpressure and admission control are preferable to unpredictable collapse. Where hardware-specific acceleration exists, preserve a portable reference path for correctness and degraded operation.

Observability must explain decisions and failures without exposing sensitive content. Log stable identifiers, shapes, versions, summary statistics, timings, and error categories. Monitor input drift, output distribution, task quality, saturation, retries, and fallback rate. Establish rollback and shadow-evaluation procedures before the first risky change. A production-grade implementation is not merely more abstract than a notebook; it makes dependencies, state, failure, and evidence explicit enough for another engineer to operate safely.

Read claims as reproducible hypotheses

The research surrounding Tiny GPT From Scratch improves representations, objectives, algorithms, systems, or evaluation protocols. Classify each paper by which lever it changes. Then identify the comparison budget: data, parameters, tokens, environment steps, hardware, communication, wall-clock time, and tuning effort. A claimed improvement may disappear when budgets are normalized or when the baseline receives equal tuning. Read methods and appendices for details that determine reproducibility, not only the abstract and headline table.

Reproduction begins with the smallest claim. Recreate one table row or ablation before attempting the entire system. Preserve the authors’ preprocessing and metric definitions, then deliberately vary one assumption. Document deviations, failed attempts, and environment details. When a result does not reproduce, distinguish an implementation defect from missing procedural knowledge, stochastic uncertainty, and genuine sensitivity. Negative evidence is useful when it narrows the conditions under which the method works.

Extension should start from a mechanism and a falsifiable prediction. The skills developed here—Language modeling, Autoregression, Tokenization—suggest multiple directions, but change one major factor at a time. Predict which metric and intermediate signal should move if the explanation is correct. Use confidence intervals and preregistered stopping rules for expensive experiments where possible. Publish code, configuration, data provenance, and failure cases so the work contributes more than another isolated score.

Maintain a chain of evidence from equation to outcome

A proof ledger for Tiny GPT From Scratch links each important claim to the smallest evidence that could disprove it. For a mathematical claim, keep a hand-worked example and a high-precision reference. For a software contract, keep unit and property tests. For an optimization claim, keep profiler traces and equal-budget baselines. For a learning claim, keep per-seed results, confidence intervals, and ablations. For a production claim, keep load tests, failure injection, monitoring queries, and rollback evidence. This structure prevents one successful end-to-end run from being treated as proof of every layer beneath it.

Record evidence beside the versioned artifact it evaluates. A metric without its dataset revision, configuration, dependency lock, hardware, and commit cannot reliably settle a regression. Likewise, a screenshot or generated sample is qualitative evidence, not a distribution. Name the claim, evidence type, acceptance threshold, owner, and date. When the implementation changes, rerun the smallest affected evidence first and then the downstream integration gates. The ledger becomes a map of confidence: it shows what is known, what is assumed, what has become stale, and where another experiment is required.

Use the ledger during review. Ask whether each test would fail for a realistic defect, whether each benchmark measures the packaged code path, whether every aggregate retains inspectable raw values, and whether uncertainty is reported at the correct independent unit. Include counterexamples and failed experiments because they define the boundary of the method. Over time this habit turns Language modeling, Autoregression, Tokenization from isolated implementation skills into a reproducible engineering practice that survives new data, new hardware, new collaborators, and changing product constraints.

Build Vocab — from contract to production evidence

Build Vocab is the construction at milestone 1 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between tokenizer and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Build Vocab as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Build Vocab depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Build Vocab needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Build Vocab can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Build Vocab changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: Tokenizer. Build a tiny character-level tokenizer with vocab, stoi/itos, and encode/decode helpers.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Get Array Dtype — from contract to production evidence

Get Array Dtype is the pipeline boundary at milestone 10 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between numpy and softmax foundations and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Get Array Dtype as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Get Array Dtype depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Get Array Dtype needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Get Array Dtype can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Get Array Dtype changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: NumPy and Softmax Foundations. Get fluent with NumPy arrays, indexing, broadcasting, reductions, and numerically stable softmax.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Scalar Broadcast Add — from contract to production evidence

Scalar Broadcast Add is the pipeline boundary at milestone 19 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between numpy and softmax foundations and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Scalar Broadcast Add as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Scalar Broadcast Add depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Scalar Broadcast Add needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Scalar Broadcast Add can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Scalar Broadcast Add changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: NumPy and Softmax Foundations. Get fluent with NumPy arrays, indexing, broadcasting, reductions, and numerically stable softmax.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Transpose Matrix — from contract to production evidence

Transpose Matrix is the pipeline boundary at milestone 28 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between numpy and softmax foundations and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Transpose Matrix as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Transpose Matrix depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Transpose Matrix needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Transpose Matrix can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Transpose Matrix changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: NumPy and Softmax Foundations. Get fluent with NumPy arrays, indexing, broadcasting, reductions, and numerically stable softmax.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Slice Train And Val — from contract to production evidence

Slice Train And Val is the learning update at milestone 37 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between data pipeline and bigram baseline and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Slice Train And Val as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Slice Train And Val depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Slice Train And Val needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Slice Train And Val can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Slice Train And Val changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: Data Pipeline and Bigram Baseline. Load the corpus, build batched (X, Y) sequences, and train a counting-based bigram model as a baseline.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Vectorize Counts Add At — from contract to production evidence

Vectorize Counts Add At is the pipeline boundary at milestone 47 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between data pipeline and bigram baseline and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Vectorize Counts Add At as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Vectorize Counts Add At depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Vectorize Counts Add At needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Vectorize Counts Add At can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Vectorize Counts Add At changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: Data Pipeline and Bigram Baseline. Load the corpus, build batched (X, Y) sequences, and train a counting-based bigram model as a baseline.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Average Nll — from contract to production evidence

Average Nll is the pipeline boundary at milestone 56 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between data pipeline and bigram baseline and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Average Nll as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Average Nll depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Average Nll needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Average Nll can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Average Nll changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: Data Pipeline and Bigram Baseline. Load the corpus, build batched (X, Y) sequences, and train a counting-based bigram model as a baseline.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Cross Entropy Loss — from contract to production evidence

Cross Entropy Loss is the measurement at milestone 65 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between single-layer neural bigram and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Cross Entropy Loss as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Cross Entropy Loss depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Cross Entropy Loss needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Cross Entropy Loss can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Cross Entropy Loss changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: Single-Layer Neural Bigram. Replace the count table with a learned weight matrix and derive cross-entropy, gradients, and SGD updates.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Linear Forward — from contract to production evidence

Linear Forward is the transformation at milestone 74 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between layer primitives and backprop and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Linear Forward as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Linear Forward depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Linear Forward needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Linear Forward can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Linear Forward changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: Layer Primitives and Backprop. Implement forward and backward passes for linear, bias, ReLU, softmax+CE, and LayerNorm building blocks.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Layernorm Forward Mean — from contract to production evidence

Layernorm Forward Mean is the transformation at milestone 84 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between layer primitives and backprop and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Layernorm Forward Mean as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Layernorm Forward Mean depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Layernorm Forward Mean needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Layernorm Forward Mean can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Layernorm Forward Mean changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: Layer Primitives and Backprop. Implement forward and backward passes for linear, bias, ReLU, softmax+CE, and LayerNorm building blocks.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Token Embedding Forward — from contract to production evidence

Token Embedding Forward is the transformation at milestone 93 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between embeddings and self-attention and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Token Embedding Forward as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Token Embedding Forward depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Token Embedding Forward needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Token Embedding Forward can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Token Embedding Forward changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: Embeddings and Self-Attention. Add token and positional embeddings, then build masked single-head and multi-head self-attention with full backward passes.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Compute Value — from contract to production evidence

Compute Value is the measurement at milestone 102 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between embeddings and self-attention and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Compute Value as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Compute Value depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Compute Value needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Compute Value can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Compute Value changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: Embeddings and Self-Attention. Add token and positional embeddings, then build masked single-head and multi-head self-attention with full backward passes.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Attention Value Backward — from contract to production evidence

Attention Value Backward is the measurement at milestone 111 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between embeddings and self-attention and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Attention Value Backward as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Attention Value Backward depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Attention Value Backward needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Attention Value Backward can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Attention Value Backward changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: Embeddings and Self-Attention. Add token and positional embeddings, then build masked single-head and multi-head self-attention with full backward passes.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Transpose Heads To Front — from contract to production evidence

Transpose Heads To Front is the pipeline boundary at milestone 120 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between embeddings and self-attention and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Transpose Heads To Front as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Transpose Heads To Front depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Transpose Heads To Front needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Transpose Heads To Front can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Transpose Heads To Front changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: Embeddings and Self-Attention. Add token and positional embeddings, then build masked single-head and multi-head self-attention with full backward passes.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Multihead Reshape Transpose Backward — from contract to production evidence

Multihead Reshape Transpose Backward is the learning update at milestone 130 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between embeddings and self-attention and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Multihead Reshape Transpose Backward as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Multihead Reshape Transpose Backward depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Multihead Reshape Transpose Backward needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Multihead Reshape Transpose Backward can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Multihead Reshape Transpose Backward changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: Embeddings and Self-Attention. Add token and positional embeddings, then build masked single-head and multi-head self-attention with full backward passes.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Transformer Block Backward — from contract to production evidence

Transformer Block Backward is the learning update at milestone 139 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between ffn, blocks, and full model and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Transformer Block Backward as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Transformer Block Backward depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Transformer Block Backward needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Transformer Block Backward can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Transformer Block Backward changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: FFN, Blocks, and Full Model. Compose feed-forward networks, residual connections, and pre-LN Transformer blocks into the complete GPT forward/backward.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Initialize Adam Step Counter — from contract to production evidence

Initialize Adam Step Counter is the learning update at milestone 148 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between adam, training loop, and generation and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Initialize Adam Step Counter as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Initialize Adam Step Counter depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Initialize Adam Step Counter needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Initialize Adam Step Counter can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Initialize Adam Step Counter changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: Adam, Training Loop, and Generation. Implement Adam, wire up the full training and validation loop, then sample text with temperature and top-k decoding.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.

Crop Context To Block Size — from contract to production evidence

Crop Context To Block Size is the pipeline boundary at milestone 157 of Tiny GPT From Scratch. Its purpose is not merely to make the next function run. It establishes a contract between adam, training loop, and generation and every downstream stage. Begin by naming the accepted inputs, their axes, units, legal ranges, ownership rules, and whether mutation is permitted. Then name the output with the same precision. In this project the surrounding ideas—Causal attention, Adam, Sampling—only compose correctly when this boundary preserves those invariants. A useful implementation note records one representative shape, one smallest valid example, one boundary example, and one invalid example before any optimization is attempted.

From first principles, treat Crop Context To Block Size as a mapping from available information to a new token representation. Ask which information is genuinely known at this point and which information would leak from the future, evaluation set, opposing player, held-out client, or later pipeline stage. Write the transformation symbolically before translating it into array operations. Every reduction must state its axis; every probability must state its normalization set; every random choice must state its distribution and seed; every learned quantity must state the objective that changes it. This discipline turns an appealing formula into an executable specification that can be challenged with small counterexamples.

The reference implementation should favor clarity over cleverness. Separate validation, the mathematical core, and state updates so each can be tested independently. Use explicit intermediate names that correspond to the derivation rather than compressing the work into one expression. Confirm dtype promotion, broadcasting, device placement, and empty-input behavior. If Crop Context To Block Size depends on randomness, pass a generator instead of reading hidden global state. If it owns mutable state, return or document the updated state explicitly. The optimized implementation may later fuse operations or reuse buffers, but it must remain numerically comparable with this small version on deterministic fixtures.

Verification for Crop Context To Block Size needs more than a happy-path assertion. Prove a hand-computable normal case, a boundary case, an invalid case, and at least one invariant. Compare against loss curves, held-out generations, ablations, and human or automated evaluations. Add metamorphic tests when an exact answer is awkward: permutation, scaling, symmetry, conservation, monotonicity, or equivalence under a harmless representation change. Run the test repeatedly under fixed seeds to distinguish deterministic defects from statistical variation. When floating-point arithmetic is involved, justify tolerances from expected rounding error instead of choosing a loose threshold simply because the test passes.

Failure analysis asks how Crop Context To Block Size can look plausible while being wrong. Inspect data leakage, exposure bias, hallucination, unstable preference signals, and unsafe deployment behavior. Trace one example through every intermediate value and preserve enough logging to reproduce it. Distinguish a contract violation from an optimization failure and from an evaluation-design failure; each requires a different repair. A numerical answer within range is not automatically meaningful, and a rising training metric is not proof that the intended signal is being learned. The strongest debugging move is usually to shrink the input until the complete computation fits on paper, then compare the paper trace with the program line by line.

Productionizing Crop Context To Block Size changes the question from “does it work once?” to “does it remain trustworthy under load and change?” Measure tokens, parameter memory, attention work, decoding latency, and evaluation coverage. Define observability for inputs, outputs, latency, failures, drift, and resource saturation. Decide what happens on malformed data, cancellation, partial worker failure, unavailable accelerators, or a distribution outside the training envelope. Version configuration and schemas with the code, preserve reproducible seeds where appropriate, and expose a safe fallback. Optimization is accepted only when the reference tests, numerical comparisons, and task-level metrics remain within an explicitly documented budget.

• Part: Adam, Training Loop, and Generation. Implement Adam, wire up the full training and validation loop, then sample text with temperature and top-k decoding.
• Normal case: choose the smallest input that exercises the intended transformation.
• Boundary case: use an empty, singleton, saturated, masked, terminal, or maximum-size input as appropriate.
• Invariant: verify shape, range, conservation, normalization, symmetry, immutability, or monotonicity.
• Production evidence: record correctness, latency, memory or cost, and the exact configuration.