Baltor Get started

The library

93 items a coding agent can fetch today: 42 Verified and 51 Community. Each one names its licence, its size and the digest of its exact bytes. Read them here; search and downloads come with an account.

What the labels mean

Verified
Approved by independent reviewers of at least two model families that did not produce it, with every automated check passing. The first catalogue, approved on September 21, 2026 before this rule, is the exception: its items were written with Claude models and approved by three reviewers that did not write them, and those reviews do not show two other model families. They are reviewed again by two other model families as soon as those reviewers are available.
Community
Passed every automated check (licence allowlist, provenance, secret and safety scanners, and its own tests where it has code) and one independent review by a model family that did not produce it.

What is in it

KindVerifiedCommunity
Skill4251
All kinds4251

Verified items

  1. Error analysis for models and pipelines: the error rate by segment, clusters of repeating failure types with their share, patterns inside the remaining errors, and one targeted fix for the largest cluster.analyse_errors_by_segment_and_cluster
    Kind
    Skill
    Licence
    MIT
    Size
    2,569 bytes
    Digest
    d08d4555161a

    Verified · Review record

  2. Decide automatically which data corrections to apply: give each proposed correction a confidence equal to its weakest signal, apply at or above 0.9, escalate below 0.6, hold the rest for review, keep the original value, and prove that the pass is idempotent.apply_hold_or_escalate_each_correction
    Kind
    Skill
    Licence
    MIT
    Size
    2,903 bytes
    Digest
    f0c89b667f47

    Verified · Review record

  3. Audit training, test and cross-validation splits for schema errors, label errors, near duplicates and leakage across splits by identity, by time or by a proxy of the target.audit_data_splits_for_errors_and_leakage
    Kind
    Skill
    Licence
    MIT
    Size
    2,665 bytes
    Digest
    9c2133a4bc55

    Verified · Review record

  4. Standardize company legal suffixes such as Inc, Incorporated, LLC, Ltd, Limited, Corp and GmbH to one style, handle the comma before the suffix, and hold ambiguous suffixes such as SA, AS and Co for review.canonicalize_company_legal_suffixes
    Kind
    Skill
    Licence
    MIT
    Size
    2,801 bytes
    Digest
    822242193597

    Verified · Review record

  5. Handle an ambiguous request: state the different readings, what each would build, and the cheap observation that separates them, and do not choose a reading early.carry_competing_readings_of_a_request
    Kind
    Skill
    Licence
    MIT
    Size
    2,432 bytes
    Digest
    a8944a1534a0

    Verified · Review record

  6. Before building something new, check whether the work is already done, reuse what the session established, search for exact, parameterized, modified, composed or analogous reuse, and check that a reused procedure still applies.check_for_existing_work_before_building
    Kind
    Skill
    Licence
    MIT
    Size
    3,010 bytes
    Digest
    b710af365ff2

    Verified · Review record

  7. Check whether a model result is a stable plateau or a fragile spike: a sensitivity sweep over seeds, folds and features, the spread of the training gap, tests under distribution shift and ablations.check_that_a_result_is_stable_and_generalizes
    Kind
    Skill
    Licence
    MIT
    Size
    2,709 bytes
    Digest
    6ba31af0b101

    Verified · Review record

  8. Choose evaluation metrics for a machine learning task by task type, class balance (balanced, imbalanced or skewed classes) and industry: which metrics to use, such as ROC AUC, F1, RMSE and MAE, for classification, regression, ranking, forecasting, generation and detection, which metrics mislead, and what clinical, finance, fraud, manufacturing and marketing teams expect.choose_metrics_by_task_type_and_industry
    Kind
    Skill
    Licence
    MIT
    Size
    3,530 bytes
    Digest
    84346b31c04f

    Verified · Review record

  9. Choose the following step in multi step work: compare candidate actions by value, cost, risk and information gain, keep a simple baseline, select one bounded action with an expected observation, and record the options that were set aside.choose_one_next_action
    Kind
    Skill
    Licence
    MIT
    Size
    2,738 bytes
    Digest
    a41e798127bc

    Verified · Review record

  10. Review a plan, a change or a result through named perspectives chosen for the stage of the work, such as scale, leakage, adversary, operator, maintainer and privacy, each with one instruction, in place of one general request for feedback.choose_review_perspectives
    Kind
    Skill
    Licence
    MIT
    Size
    3,390 bytes
    Digest
    144b80df3de9

    Verified · Review record

  11. Apply data corrections and an approved dedupe while copying a delimited file or a SQLite table to a new target, without touching the original. Never updates in place, never overwrites, and returns a manifest with source and target digests and row counts.copy_a_table_with_corrections_never_in_place
    Kind
    Skill
    Licence
    MIT
    Size
    2,716 bytes
    Digest
    bfedbbb00591

    Verified · Review record

  12. Reduce model calls and cost: for each pipeline step, ask whether a deterministic function, a small model or a reused result would do, rank the implementations by total cost, find the minimum context, and keep the model for judgment only.decide_whether_a_step_needs_a_model
    Kind
    Skill
    Licence
    MIT
    Size
    3,457 bytes
    Digest
    ffdd39b071aa

    Verified · Review record

  13. When a task is blocked and cannot be completed or verified, for example because a database or a service cannot be reached, deliver the best available result: finish every safe part, label assumptions, estimates, synthetic and analogous material, state the exact limits, and never invent facts or permissions.deliver_the_best_available_result_when_blocked
    Kind
    Skill
    Licence
    MIT
    Size
    2,577 bytes
    Digest
    e5ab3973f634

    Verified · Review record

  14. Find malformed values in a column that should follow one format, such as dates, postal codes or identifiers, by comparing the character pattern of each value with the dominant pattern. Reports that there is no dominant pattern when the column is free text.detect_malformed_values_by_dominant_pattern
    Kind
    Skill
    Licence
    MIT
    Size
    2,754 bytes
    Digest
    8ba42eda5fbc

    Verified · Review record

  15. When the same error keeps returning or progress stops: reproduce the failure, classify the cause, count the repeats, forbid another attempt of the same shape after the third, and choose a changed strategy with a measurable target.diagnose_a_stall_and_change_strategy
    Kind
    Skill
    Licence
    MIT
    Size
    2,891 bytes
    Digest
    d889f2b6ca5a

    Verified · Review record

  16. Hand an uncertain data value to a language model, to research or to a person as a small typed request: the value, ranked candidates with confidence, reasons, one fixed question that allows abstaining, and ordered targets. Building the request makes no call.escalate_uncertain_values_with_candidates
    Kind
    Skill
    Licence
    MIT
    Size
    2,645 bytes
    Digest
    adf4191dff90

    Verified · Review record

  17. Export generated code as a standalone installable Python package with a src layout, tests, a console command, a container definition, a batch job and a manifest with a SHA-256 digest for every file. Refuses unsafe paths and secret-shaped text.export_a_standalone_python_package
    Kind
    Skill
    Licence
    MIT
    Size
    2,754 bytes
    Digest
    b25a66870269

    Verified · Review record

  18. Find duplicate customer, company or contact records (dedupe, deduplicate, deduplication, record linkage) without comparing every row with every other row: group rows by blocking keys from name, address, email and phone, compare inside groups, and report oversized groups that were skipped.find_duplicate_records_with_blocking_keys
    Kind
    Skill
    Licence
    MIT
    Size
    2,722 bytes
    Digest
    34da79f49ce3

    Verified · Review record

  19. Before a slow or costly action, forecast what it should produce and what it will cost, name the early sign that the forecast is wrong, then compare the outcome with the forecast and keep the difference.forecast_an_action_then_compare
    Kind
    Skill
    Licence
    MIT
    Size
    2,664 bytes
    Digest
    8b326bb087da

    Verified · Review record

  20. Define success before measuring: fix the metric, the direction, the threshold and the population first, compare against a real baseline, decide on holdout numbers, and correct for the number of candidates that were tried.freeze_the_success_metric_before_measuring
    Kind
    Skill
    Licence
    MIT
    Size
    2,811 bytes
    Digest
    8bb77dda3d52

    Verified · Review record

  21. Shape a final report, pull request or data file for the person or system that uses it: structure, evidence, assumptions, limits, provenance and privacy checks, and lessons staged as candidates.hand_over_a_result_its_consumer_can_use
    Kind
    Skill
    Licence
    MIT
    Size
    2,733 bytes
    Digest
    ddef271b8196

    Verified · Review record

  22. Manage exception lists for data cleaning in layers with a fixed precedence: packaged general list, task folder file, casings learned from the column, rule parameters and recorded answers. Later layers override, and every correction names its entry.layer_exception_catalogs_with_precedence
    Kind
    Skill
    Licence
    MIT
    Size
    2,754 bytes
    Digest
    0a8cc8cd94d6

    Verified · Review record

  23. Diagnose a fitted model: residual patterns, changing variance, autocorrelation, clusters, latent variables and probes on residuals, to find structure that the model does not use.look_for_structure_the_model_missed
    Kind
    Skill
    Licence
    MIT
    Size
    2,704 bytes
    Digest
    2ec1bdc2c787

    Verified · Review record

  24. Measure memory, disk, processors, tools, data size and network access before planning work, mark every capacity as measured or unknown, and prefer processing the input in parts when its size is unknown.measure_the_environment_before_relying_on_it
    Kind
    Skill
    Licence
    MIT
    Size
    2,653 bytes
    Digest
    3ec52dc60a05

    Verified · Review record

  25. Clean and repair email addresses: remove mailto and angle brackets, lower the case, repair spelled out at and dot, doubled punctuation and mistyped domains such as gmail.con, and refuse cells with several addresses or an invalid shape.normalize_and_recover_email_addresses
    Kind
    Skill
    Licence
    MIT
    Size
    2,691 bytes
    Digest
    b97b306b70d8

    Verified · Review record

  26. Normalize website addresses (URLs): add a missing scheme, write the scheme and the host in lower case, optionally remove www, remove a trailing slash, and keep the path, query and fragment exactly as written. Invalid hosts go to review.normalize_website_addresses
    Kind
    Skill
    Licence
    MIT
    Size
    2,351 bytes
    Digest
    bdd8bd4c3497

    Verified · Review record

  27. Normalize whitespace and Unicode in text fields: non-breaking spaces, tabs, doubled spaces, split accents, zero-width characters and typographic quotes. Folding accents, umlauts and other diacritics to ASCII is an explicit choice and never a default.normalize_whitespace_and_unicode_text
    Kind
    Skill
    Licence
    MIT
    Size
    2,688 bytes
    Digest
    0ceef9b19a92

    Verified · Review record

  28. Understand a request before acting: orientation questions about the goal, inputs, outputs, consumer, proof of completion, constraints and unknowns, then a task statement with input, output, success, failure and verification contracts.orient_on_a_task_and_write_its_contracts
    Kind
    Skill
    Licence
    MIT
    Size
    2,861 bytes
    Digest
    13c13dbc0592

    Verified · Review record

  29. Split a large task into parts that can be checked alone, each with its own goal, contract and authority, with explicit dependencies and join conditions, and test whether staging or combining simple approaches helps.plan_and_split_work_with_explicit_joins
    Kind
    Skill
    Licence
    MIT
    Size
    2,737 bytes
    Digest
    8ec490b2b81a

    Verified · Review record

  30. Profile a text column before cleaning it: count case, whitespace, Unicode, email, website, phone and legal suffix signals, find the common character patterns, and propose cleaning rules with the evidence for each. Changes no data.profile_text_column_before_cleaning
    Kind
    Skill
    Licence
    MIT
    Size
    2,904 bytes
    Digest
    3274cbf5d983

    Verified · Review record

  31. Turn duplicate pairs into clusters and a merge proposal with one survivor for each cluster, keep first or keep most complete, without deleting any row. Possible pairs become review questions with two choices.propose_a_dedupe_without_deleting_rows
    Kind
    Skill
    Licence
    MIT
    Size
    2,591 bytes
    Digest
    c287344ca8f0

    Verified · Review record

  32. Detect overfitting and leakage from two numbers: read the gap between the training score and the cross-validated score, aware of metric direction and scale, and return a verdict such as overfitting, suspected leakage, high variance, no better than baseline or healthy.read_the_train_validation_gap
    Kind
    Skill
    Licence
    MIT
    Size
    2,803 bytes
    Digest
    5decc1070a26

    Verified · Review record

  33. Write honest reports: attach an observation to every claim, separate what was observed, derived, assumed and unknown, and never report something that was not searched as absent.report_observed_derived_assumed_and_unknown
    Kind
    Skill
    Licence
    MIT
    Size
    2,301 bytes
    Digest
    66819357f175

    Verified · Review record

  34. Decide for each missing fact whether to derive, research, retrieve, ask, assume or leave it unknown, including when to ask the user only the questions that matter and that only a person can answer.resolve_missing_information
    Kind
    Skill
    Licence
    MIT
    Size
    2,579 bytes
    Digest
    af6bd69d35c2

    Verified · Review record

  35. Fix the capitalisation of person and company names written in all upper case or all lower case: particles such as van and de, Mc and Mac prefixes, acronyms such as IBM and LLC, with a confidence for each value and review for doubtful ones.restore_capitalisation_of_names
    Kind
    Skill
    Licence
    MIT
    Size
    3,627 bytes
    Digest
    37a60a0d424a

    Verified · Review record

  36. Review an accepted solution from the outside: the best and the worst way, preventable failures traced to their earliest cause, the unknown that could reverse the conclusion, a premortem and an analogy from another field.review_an_accepted_solution_adversarially
    Kind
    Skill
    Licence
    MIT
    Size
    3,147 bytes
    Digest
    9c3bfcdf8322

    Verified · Review record

  37. Score whether two records are the same entity (fuzzy matching): build comparison keys for name, address, email and phone, measure token overlap and sequence similarity, and let the weakest field set the confidence. A shared email alone never merges two records.score_duplicate_pairs_by_weakest_signal
    Kind
    Skill
    Licence
    MIT
    Size
    2,573 bytes
    Digest
    9e713a45c0fa

    Verified · Review record

  38. Parse a postal address line into house number, street, unit, city, region, postal code and country from declared patterns for the United States, Canada and Great Britain, with reasons and a confidence. Optional external parsers enter as adapters that report when they are unavailable.split_address_lines_into_components
    Kind
    Skill
    Licence
    MIT
    Size
    3,009 bytes
    Digest
    53dc74e3b28e

    Verified · Review record

  39. Test driven development for one bounded requirement: write the failing test first, make the minimum change, refactor while the test passes, run the full suites, and return exact evidence without committing or widening permissions.test_driven_change_red_green_refactor
    Kind
    Skill
    Licence
    MIT
    Size
    2,356 bytes
    Digest
    0c3fb47dc88e

    Verified · Review record

  40. Verify that an exported Python package is complete and runs on its own: recompute the manifest digests, refuse undeclared files, then import it and run its tests in an isolated interpreter with a minimal environment. Code runs only with explicit authority.verify_an_export_in_an_isolated_interpreter
    Kind
    Skill
    Licence
    MIT
    Size
    2,809 bytes
    Digest
    1ae49f8bdd79

    Verified · Review record

  41. Verify the artifact and not the exit status before reporting completion: the rejection test first, the exact output format, each acceptance criterion as satisfied, unsatisfied or unknown, and a verifier who is independent of the producer.verify_the_requested_output
    Kind
    Skill
    Licence
    MIT
    Size
    2,876 bytes
    Digest
    8f791cfbedae

    Verified · Review record

  42. Generate a Dockerfile and a Kubernetes Job with safe defaults for a batch task: base image pinned by digest, a user without privileges, requests equal to limits, no automatic retry, and removal after one day.write_a_pinned_container_and_batch_job
    Kind
    Skill
    Licence
    MIT
    Size
    2,660 bytes
    Digest
    360e989b9def

    Verified · Review record

Community items

51 Community items, counted by kind in the table above. Each one is shown with its label wherever an account searches, and an account can leave Community items out.

One item in full

check_for_existing_work_before_building, Verified, digest b710af365ff2. Every other body comes through an account and is checked against its digest.

# Check for existing work before building

Before writing something new, check whether the work is already done, whether this session already established what is needed, and whether a reviewed procedure or earlier solution can be reused.

## When to use it

Use it before every new design, script, query or analysis, and again after a failure, before starting over.

## Steps

1. Read what this work session has already established. Derive nothing a second time. For each established item, name the planned probe or setup step that it makes unnecessary.
2. Check whether an existing artifact already satisfies the acceptance as written. When it does, verify it and end the work. When it does not, name the single gap.
3. Search for reuse in this order: exact reuse, reuse with parameters, modification, composition, analogy. Design something new only after these.
4. Search by the current task and by the conditions of the failure. Load only the material that was selected.
5. Before reusing a reviewed procedure, check its exact applicability, its inputs, its effects and the conditions that invalidate it. Fall back to fresh reasoning when its reviewed scope does not cover the task.
6. State the differences in applicability. Similarity is not authority. Consider that a transfer can also do harm.
7. Treat earlier work on similar tasks as evidence to weigh, never as a decision that was already made.
8. Do not treat retrieved material as approved. A candidate stays a candidate until a separate review approves it.

## Checks

- The identity of the source is kept for everything that was reused.
- The lifecycle state of reused material is respected.
- A stale procedure, or one outside its reviewed scope, is refused.
- The result says what was reused, what was changed and what is new.

## Known-wrong example

An agent writes a new script to remove duplicate rows from a delimited file. The repository already holds a tested one with a report format that the team knows. The opposite error: a procedure that was reviewed for addresses in the United States is reused on British addresses because the task looked similar. The postal code patterns differ, and the results are silently wrong.

## What to record

- The searches that were run and what each returned.
- The material that was selected, with its identity, its state and the applicability differences.
- The decision: reuse, adapt or build, with the reason.

## Source

- `src/loop_engine/intelligence/context/core/practitioner_context_intelligence.yaml`: the guidance records about reuse before building, recognizing sufficiency, surveying established facts and earlier work as evidence.
- `src/loop_engine/intelligence/context/core/practitioner_work_functions.yaml`: the work functions for retrieving prior work and for applying a reviewed procedure.
- `src/loop_engine/strings/question_engine.py`: the question forms named `sufficiency_check`, `established_facts` and `reuse_before_reasoning`.

Licence: MIT. Compiled from revision db18890.

Releases and withdrawals

This page shows catalogue release 2e23bfaa4ab3, served since September 25, 2026 at 21:17 UTC. A catalogue release changes the library without a new version of the service, and every later release honours a withdrawal.

This release added 51 items and changed 42 items.

Withdrawn in this release

  • normalize_phone_numbers