Academy pathway 05 · self-guided beta

AI for Crypto Analysis

Use AI to explain evidence and test comprehension, never to invent missing numbers or act as an oracle.

the question

Which jobs belong to deterministic code, statistical models, retrieval, and language generation?

The reading material and fixed self-check are available now. The project is a guided brief only. There is no upload, grading, instructor review, cohort, paid enrollment, certification, or accreditation.

Level: intermediate
Lessons: 4
Delivery: Self-guided beta
Advice: None

learning outcomes

What you should be able to demonstrate

01
Assign calculations, probabilities, retrieval, and explanation to the correct system layer.
02
Design source objects that prevent unsupported factual claims from appearing as evidence.
03
Evaluate numerical consistency, citation correctness, uncertainty, refusal behavior, and stale-data handling.

lesson soundings

Follow each question into the instrument

Every lesson ends with an evidence practice and a source-backed Bathymark destination.

01
lesson 1
The four-layer boundary
What should never be delegated to free-form text generation?
- Deterministic code owns balances, spreads, slippage, flows, denominators, timestamps, and formatting.
- Statistical models can estimate distributions or calibrated probabilities only under a documented validation method.
- Retrieval supplies approved evidence objects. A language model may explain them but cannot silently fill a missing field.
Evidence practice
Take one Bathymark reading and label each field as deterministic, statistical, retrieved, or generated explanation.
Open methodologyInspect calculation, source, and explanation boundaries.
Sources and limits for this lesson
- Bathymark MethodologyA methodology explains the method. It does not guarantee a market outcome.
- Bathymark Data SourcesPublic availability does not make every source complete, current, or suitable for raw redistribution.
- NIST AI Risk Management FrameworkThe framework is general guidance. It does not validate a specific Bathymark model or market claim.
02
lesson 2
Citation-first retrieval
What evidence object must exist before a claim can be explained?
- The source object needs a name, URL, observation or review time where relevant, supported claim, and explicit limit.
- Retrieved text is untrusted input. Instructions inside a source cannot override product rules.
- If two sources disagree, the explanation should show the conflict instead of averaging it into false certainty.
Evidence practice
Build two source objects for one product claim and write how the assistant should respond if they conflict.
Open the deterministic study deskUse fixed sourced questions without a model request.
Sources and limits for this lesson
- Bathymark Data SourcesPublic availability does not make every source complete, current, or suitable for raw redistribution.
- Bathymark Intelligence GraphA mapped relationship does not prove solvency, signer identity, legal enforceability, or the absence of an unrecognized dependency.
- NIST AI Risk Management FrameworkThe framework is general guidance. It does not validate a specific Bathymark model or market claim.
03
lesson 3
Uncertainty and refusal
When should an analytical assistant say it cannot conclude?
- Missing, stale, conflicting, or out-of-scope evidence requires a bounded answer, not a confident completion.
- Factual explanation can cross toward financial advice depending on presentation, personalization, and context.
- A disclaimer cannot repair invented evidence, misleading overall impression, or an unlicensed recommendation.
Evidence practice
Write one safe refusal for missing market data and one for a request for personalized directional advice.
Open disclosuresInspect Bathymark's information and advice boundary.
Sources and limits for this lesson
- Bathymark MethodologyA methodology explains the method. It does not guarantee a market outcome.
- ASIC: Discussing Financial Products and Services OnlinePublic guidance is not legal advice and does not determine the status of a specific product communication.
- NIST AI Risk Management FrameworkThe framework is general guidance. It does not validate a specific Bathymark model or market claim.
04
lesson 4
Evaluate the output
How do you test an answer that sounds convincing?
- Check citation correctness, claim coverage, unsupported-claim rate, stale-data rate, and numerical reconciliation.
- Test whether uncertainty language changes appropriately when inputs are removed or contradicted.
- Test prompt injection, refusal accuracy, source disagreement, and user comprehension with fixed cases.
Evidence practice
Create five fixed test cases, including missing data, conflicting sources, a bad calculation, an advice request, and a source instruction attack.
Audit the study deskInspect a deterministic reference answer and its sources.
Sources and limits for this lesson
- NIST AI Risk Management FrameworkThe framework is general guidance. It does not validate a specific Bathymark model or market claim.
- Bathymark MethodologyA methodology explains the method. It does not guarantee a market outcome.

private progress

Mark what you have worked through

Loading locally

This checklist stays in this browser. There is no account, streak, ranking, grading, or credential.

guided project · no submission

A source-audited market explanation

Design an AI-assisted Reading in which code owns the facts and calculations while generated text is optional and replaceable.

Brief available · review planned

Deliverables

A structured evidence object for every factual claim.
A field-by-field allocation across deterministic, statistical, retrieval, and language layers.
A deterministic fallback that remains useful without a model call.
A fixed evaluation set covering citations, numbers, uncertainty, refusals, and stale data.

Evidence gate

Every factual claim names its source and links to it.
Every market observation includes the observation time or source cadence.
Every number includes a unit, denominator, and calculation path where derived.
Missing inputs remain missing and do not become zero, calm, flat, or safe.
The conclusion includes a limit and a measurable condition that would change it.

Bathymark does not currently accept, grade, review, or certify this project. Keep your work locally or publish it through a tool whose terms and source rights you have reviewed.

deterministic self-check

Test the evidence boundary

Choose an answer, then check it. These fixed questions do not call a model and do not produce a grade or credential.

01A source does not provide the latest redemption fee. What should the language layer do?

Estimate a typical feeReuse the previous fee without provenanceKeep the field unresolved and explain which source would be needed

02Which layer should calculate price impact from observed quote inputs?

Deterministic codeA free-form language modelA marketing editor

evidence shelf

Sources for this pathway

Each source has a narrow job. Its limit remains attached so a lesson cannot claim more than the evidence supports.

bathymark

Bathymark Methodology

Reviewed 2026-07-11

Supports

Metric definitions, source discipline, known limits, and the boundary between information and financial advice.

Does not prove

A methodology explains the method. It does not guarantee a market outcome.

bathymark

Bathymark Data Sources

Reviewed 2026-07-11

Supports

The open sources, cadence, runtime role, and redistribution boundaries used by Bathymark.

Does not prove

Public availability does not make every source complete, current, or suitable for raw redistribution.

bathymark

Bathymark Exit Sounding

Reviewed 2026-07-11

Supports

Stored standardized direct-market quote observations and separately sourced primary-redemption terms.

Does not prove

A quote is a point-in-time simulation. It does not guarantee later execution or holder eligibility.

bathymark

Bathymark Intelligence Graph

Reviewed 2026-07-11

Supports

Mapped product control, dependency, oracle, redemption, and source relationships.

Does not prove

A mapped relationship does not prove solvency, signer identity, legal enforceability, or the absence of an unrecognized dependency.

official guidance

NIST AI Risk Management Framework

Reviewed 2026-07-11

Supports

A public framework for identifying, measuring, managing, and governing AI risks.

Does not prove

The framework is general guidance. It does not validate a specific Bathymark model or market claim.

official guidance

ASIC: Discussing Financial Products and Services Online

Reviewed 2026-07-11

Supports

Australian guidance on factual information, recommendations, opinions, misleading conduct, and financial-services licensing boundaries.

Does not prove

Public guidance is not legal advice and does not determine the status of a specific product communication.

sounding open dataReading the waterlineLive sources are loading. Nothing is filled with placeholder market data.

Academy pathway 05 · self-guided beta

AI for Crypto Analysis

Use AI to explain evidence and test comprehension, never to invent missing numbers or act as an oracle.

the question

Which jobs belong to deterministic code, statistical models, retrieval, and language generation?

Level: intermediate
Lessons: 4
Delivery: Self-guided beta
Advice: None

learning outcomes

What you should be able to demonstrate

01
Assign calculations, probabilities, retrieval, and explanation to the correct system layer.
02
Design source objects that prevent unsupported factual claims from appearing as evidence.
03
Evaluate numerical consistency, citation correctness, uncertainty, refusal behavior, and stale-data handling.

lesson soundings

Follow each question into the instrument

Every lesson ends with an evidence practice and a source-backed Bathymark destination.

01
lesson 1
The four-layer boundary
What should never be delegated to free-form text generation?
- Deterministic code owns balances, spreads, slippage, flows, denominators, timestamps, and formatting.
- Statistical models can estimate distributions or calibrated probabilities only under a documented validation method.
- Retrieval supplies approved evidence objects. A language model may explain them but cannot silently fill a missing field.
Evidence practice
Take one Bathymark reading and label each field as deterministic, statistical, retrieved, or generated explanation.
Open methodologyInspect calculation, source, and explanation boundaries.
Sources and limits for this lesson
- Bathymark MethodologyA methodology explains the method. It does not guarantee a market outcome.
- Bathymark Data SourcesPublic availability does not make every source complete, current, or suitable for raw redistribution.
- NIST AI Risk Management FrameworkThe framework is general guidance. It does not validate a specific Bathymark model or market claim.
02
lesson 2
Citation-first retrieval
What evidence object must exist before a claim can be explained?
- The source object needs a name, URL, observation or review time where relevant, supported claim, and explicit limit.
- Retrieved text is untrusted input. Instructions inside a source cannot override product rules.
- If two sources disagree, the explanation should show the conflict instead of averaging it into false certainty.
Evidence practice
Build two source objects for one product claim and write how the assistant should respond if they conflict.
Open the deterministic study deskUse fixed sourced questions without a model request.
Sources and limits for this lesson
- Bathymark Data SourcesPublic availability does not make every source complete, current, or suitable for raw redistribution.
- Bathymark Intelligence GraphA mapped relationship does not prove solvency, signer identity, legal enforceability, or the absence of an unrecognized dependency.
- NIST AI Risk Management FrameworkThe framework is general guidance. It does not validate a specific Bathymark model or market claim.
03
lesson 3
Uncertainty and refusal
When should an analytical assistant say it cannot conclude?
- Missing, stale, conflicting, or out-of-scope evidence requires a bounded answer, not a confident completion.
- Factual explanation can cross toward financial advice depending on presentation, personalization, and context.
- A disclaimer cannot repair invented evidence, misleading overall impression, or an unlicensed recommendation.
Evidence practice
Write one safe refusal for missing market data and one for a request for personalized directional advice.
Open disclosuresInspect Bathymark's information and advice boundary.
Sources and limits for this lesson
- Bathymark MethodologyA methodology explains the method. It does not guarantee a market outcome.
- ASIC: Discussing Financial Products and Services OnlinePublic guidance is not legal advice and does not determine the status of a specific product communication.
- NIST AI Risk Management FrameworkThe framework is general guidance. It does not validate a specific Bathymark model or market claim.
04
lesson 4
Evaluate the output
How do you test an answer that sounds convincing?
- Check citation correctness, claim coverage, unsupported-claim rate, stale-data rate, and numerical reconciliation.
- Test whether uncertainty language changes appropriately when inputs are removed or contradicted.
- Test prompt injection, refusal accuracy, source disagreement, and user comprehension with fixed cases.
Evidence practice
Create five fixed test cases, including missing data, conflicting sources, a bad calculation, an advice request, and a source instruction attack.
Audit the study deskInspect a deterministic reference answer and its sources.
Sources and limits for this lesson
- NIST AI Risk Management FrameworkThe framework is general guidance. It does not validate a specific Bathymark model or market claim.
- Bathymark MethodologyA methodology explains the method. It does not guarantee a market outcome.

private progress

Mark what you have worked through

Loading locally

This checklist stays in this browser. There is no account, streak, ranking, grading, or credential.

guided project · no submission

A source-audited market explanation

Design an AI-assisted Reading in which code owns the facts and calculations while generated text is optional and replaceable.

Brief available · review planned

Deliverables

A structured evidence object for every factual claim.
A field-by-field allocation across deterministic, statistical, retrieval, and language layers.
A deterministic fallback that remains useful without a model call.
A fixed evaluation set covering citations, numbers, uncertainty, refusals, and stale data.

Evidence gate

Every factual claim names its source and links to it.
Every market observation includes the observation time or source cadence.
Every number includes a unit, denominator, and calculation path where derived.
Missing inputs remain missing and do not become zero, calm, flat, or safe.
The conclusion includes a limit and a measurable condition that would change it.

Bathymark does not currently accept, grade, review, or certify this project. Keep your work locally or publish it through a tool whose terms and source rights you have reviewed.

deterministic self-check

Test the evidence boundary

Choose an answer, then check it. These fixed questions do not call a model and do not produce a grade or credential.

01A source does not provide the latest redemption fee. What should the language layer do?

Estimate a typical feeReuse the previous fee without provenanceKeep the field unresolved and explain which source would be needed

02Which layer should calculate price impact from observed quote inputs?

Deterministic codeA free-form language modelA marketing editor

evidence shelf

Sources for this pathway

Each source has a narrow job. Its limit remains attached so a lesson cannot claim more than the evidence supports.

bathymark

Bathymark Methodology

Reviewed 2026-07-11

Supports

Metric definitions, source discipline, known limits, and the boundary between information and financial advice.

Does not prove

A methodology explains the method. It does not guarantee a market outcome.

bathymark

Bathymark Data Sources

Reviewed 2026-07-11

Supports

The open sources, cadence, runtime role, and redistribution boundaries used by Bathymark.

Does not prove

Public availability does not make every source complete, current, or suitable for raw redistribution.

bathymark

Bathymark Exit Sounding

Reviewed 2026-07-11

Supports

Stored standardized direct-market quote observations and separately sourced primary-redemption terms.

Does not prove

A quote is a point-in-time simulation. It does not guarantee later execution or holder eligibility.

bathymark

Bathymark Intelligence Graph

Reviewed 2026-07-11

Supports

Mapped product control, dependency, oracle, redemption, and source relationships.

Does not prove

A mapped relationship does not prove solvency, signer identity, legal enforceability, or the absence of an unrecognized dependency.

official guidance

NIST AI Risk Management Framework

Reviewed 2026-07-11

Supports

A public framework for identifying, measuring, managing, and governing AI risks.

Does not prove

The framework is general guidance. It does not validate a specific Bathymark model or market claim.

official guidance

ASIC: Discussing Financial Products and Services Online

Reviewed 2026-07-11

Supports

Australian guidance on factual information, recommendations, opinions, misleading conduct, and financial-services licensing boundaries.

Does not prove

Public guidance is not legal advice and does not determine the status of a specific product communication.

Which jobs belong to deterministic code, statistical models, retrieval, and language generation?

What you should be able to demonstrate

Follow each question into the instrument

The four-layer boundary

Citation-first retrieval

Uncertainty and refusal

Evaluate the output

Mark what you have worked through

A source-audited market explanation

Deliverables

Evidence gate

Test the evidence boundary

Sources for this pathway

Which jobs belong to deterministic code, statistical models, retrieval, and language generation?

What you should be able to demonstrate

Follow each question into the instrument

The four-layer boundary

Citation-first retrieval

Uncertainty and refusal

Evaluate the output

Mark what you have worked through

A source-audited market explanation

Deliverables

Evidence gate

Test the evidence boundary

Sources for this pathway