pattern_references.md

Pattern References for CAK R&D

This is not a dependency list. This is not a required-tool list. This is not an endorsed stack. These references are inspected for reusable patterns that may inspire CAK R&D.

Seed references are not accepted evidence until inspected and recorded in a source ledger.

How to use this file

Structure each reference as:

## Reference name

Inspect for:
- ...

Why it matters:
- ...

Questions it informs:
- ...

Caution:
- ...

Potential CAK-native adaptation:
- ...

Evidence status:
- seed | inspected | source-ledgered | rejected

Skill and procedural-memory anchors

Voyager

Inspect for:

• executable code skills;
• external skill library without weight updates;
• self-verification before library admission;
• iterative prompting with environment feedback and execution errors;
• compositional reuse;
• transfer to new tasks / another agent benefiting from the skill library.

Why it matters:

• Voyager suggests that agent skills can be executable external artifacts, not model-weight updates.

Questions it informs:

• What is an agent-native skill?
• How should skill admission use execution evidence?
• When should skills remain external and auditable?

Caution:

• Stable API and clear environment feedback make Voyager easier than open web or enterprise tool governance.

Potential CAK-native adaptation:

• Skill artifact admission should require executable validation, verifier evidence, trace/replay, and compositional dependency tracking.

Evidence status:

• seed

Agent Workflow Memory

Inspect for:

• workflow induction from trajectories;
• offline and online memory;
• abstract sub-routines;
• website/domain generalization;
• workflow action-space variant;
• brittleness in dynamic UI;
• need to diverge from workflow guidelines.

Why it matters:

• AWM shows workflow memory can improve web agents, but also shows why linear workflows and macro-actions are not enough.

Questions it informs:

• Should CAK represent workflows as lists, state machines, behavior trees, or guarded plans?
• How should retrieval account for runtime state?

Caution:

• Web workflows can break when UI state diverges from the learned routine.

Potential CAK-native adaptation:

• Represent workflows as guarded state machines or stage-aware plans with preconditions/postconditions, not fixed step lists.

Evidence status:

• seed

HASP

Inspect for:

• Program Functions;
• should_activate / intervene interface;
• action override;
• context injection;
• structured intervention records;
• four intervention signals: timing, mode, correctness, outcome;
• executable validation;
• teacher review;
• strict filtering;
• library pollution risk.

Why it matters:

• HASP is a strong anchor for active runtime-control skills rather than passive prompt advice.

Questions it informs:

• When should a skill actively intervene?
• How should active skills be validated, recorded, and quarantined?

Caution:

• Active interventions can overblock, override valid behavior, or become unsafe without strong gates.

Potential CAK-native adaptation:

• Treat active skills as runtime control objects with hook semantics, audit records, rate limits, validation gates, and rollback/quarantine.

Evidence status:

• seed

Agent Skills format

Inspect for:

• portable skill package ergonomics;
• SKILL.md-like metadata and instructions;
• scripts/references/assets separation;
• progressive disclosure;
• cross-agent reuse.

Why it matters:

• This is a baseline for portable skill packaging.

Questions it informs:

• Which package fields are distribution concerns?
• Which package fields need compiled runtime or verifier semantics?

Caution:

• Folder-based human-readable packages may not be enough for agent-native runtime control, verifier-gated admission, or safety.

Potential CAK-native adaptation:

• Use portable packaging only as a distribution layer; compile into CAK-native artifacts such as ContractSpec, SkillSpec, VerifierPlan, PolicySpec, and tests.

Evidence status:

• seed

last30days-skill

Inspect for:

• horizon scanning;
• multi-source discovery;
• source scoring;
• entity-aware pre-research;
• cross-source cluster merging;
• grounded synthesis;
• shareable brief artifact generation;
• social/practitioner signal collection.

Why it matters:

• CAK R&D needs a discovery process richer than "ask one model."

Questions it informs:

• How should CAK run source discovery?
• How should a Scout separate leads from accepted evidence?

Caution:

• Social engagement is not truth. Discovery output must be audited through source_ledger.yaml.

Potential CAK-native adaptation:

• Horizon Scout role for research runs: discover sources, repos, discussions, and recent signals; extract candidate patterns; never convert discovery into accepted evidence without source ledger.

Evidence status:

• seed

VASO

Inspect for:

• semantic skill contracts;
• formal/planner-facing interfaces;
• model-checker filtering;
• temporal safety specifications;
• counterexample traces used to improve skill contracts;
• frozen model weights with evolving external skill contracts.

Why it matters:

• VASO is close to CAK's ContractSpec/type-system direction.

Questions it informs:

• Should ContractSpec expose runtime predicates and verifier obligations?
• How should counterexample traces update contracts?

Caution:

• Formal interfaces may require assumptions that current agent environments do not expose cleanly.

Potential CAK-native adaptation:

• ContractSpec may need dual interfaces: runtime-facing predicates for the agent loop; verifier-facing formal obligations for safety or correctness.

Evidence status:

• seed

Programmatic Skill Networks

Inspect for:

• executable symbolic skill programs;
• compositional skill networks;
• structured fault localization;
• maturity-aware update gating;
• rollback validation;
• structural refactoring.

Why it matters:

• PSN may offer better lifecycle semantics than a flat skill registry.

Questions it informs:

• Should CAK skills have dependency graphs, maturity levels, and rollback validation?

Caution:

• Symbolic networks may understate ambiguity in LLM agent state and tool environments.

Potential CAK-native adaptation:

• Skill libraries may need maturity levels, dependency graphs, refactoring operations, and rollback validation.

Evidence status:

• seed

Hierarchical Memory Tree

Inspect for:

• decoupling logical planning from action execution;
• intent/stage/action hierarchy;
• observable preconditions and postconditions;
• stage-aware Planner/Actor split;
• preventing workflow mismatch across unseen websites.

Why it matters:

• HMT is highly relevant to CAK's state-conditioned retrieval and workflow-memory questions.

Questions it informs:

• How should CAK separate task intent, semantic stage, action grounding, and verifier conditions?

Caution:

• Website-specific grounding may not transfer without environment-specific selectors and postconditions.

Potential CAK-native adaptation:

• CAK memory should distinguish: task intent; semantic stage; action grounding; verifier conditions; environment-specific selectors.

Evidence status:

• seed

SkillWiki

Inspect for:

• skill infrastructure;
• provenance-aware exploration;
• grounding reusable skills in originating evidence;
• continuous evolution of skill assets.

Why it matters:

• SkillWiki may be a closer analogue to a future SkillPack ecosystem than ordinary package managers.

Questions it informs:

• What provenance does a skill registry need?
• When should infrastructure follow runtime evidence rather than precede it?

Caution:

• Do not assume infrastructure/registry should come before runtime/verifier evidence.

Potential CAK-native adaptation:

• Treat provenance and continuous evolution as required metadata, but defer registry architecture until active runtime and verifier evidence exists.

Evidence status:

• seed

SkillRevise

Inspect for:

• trace-conditioned skill revision;
• execution-grounded diagnosis;
• iterative skill repair;
• empirical utility measurement;
• cross-model transfer of revised skills.

Why it matters:

• SkillRevise may help CAK avoid one-shot skill authoring and instead revise skills through traces.

Questions it informs:

• How should CAK revise skill candidates after failed traces?
• Which utility metrics should gate revised skill admission?

Caution:

• Revision can overfit to recent traces or poison a shared library if admission gates are weak.

Potential CAK-native adaptation:

• Skill evolution should be trace-conditioned, versioned, verifier-gated, and empirically compared before admission.

Evidence status:

• seed

Agent Skills security analyses

Inspect for:

• prompt injection through skill files;
• malicious skill classification;
• repository-context analysis;
• abandoned repository hijacking;
• approval widening / "don't ask again" hazards.

Why it matters:

• A skill ecosystem is also a supply-chain and prompt-injection surface.

Questions it informs:

• What trust metadata does a SkillPack need?
• How should CAK limit permission widening and self-poisoning?

Caution:

• Human-readable skills can be executable influence channels even when they are "just documentation."

Potential CAK-native adaptation:

• Every skill package needs trust metadata, provenance, sandboxing, permission declarations, exact-scope approvals, and quarantine/rollback.

Evidence status:

• seed

Debate / multi-review / research-harness pattern references

nitpicker

Inspect for:

• multiple independent reviewers;
• actor/critic debate;
• aggregator / meta-reviewer;
• reviewer agents with file/grep/git tools;
• transcript and trajectory logs;
• provider diversity;
• machine-readable output.

Why it matters:

• CAK R&D needs structured disagreement, not one-pass synthesis.

Questions it informs:

• Which debate roles and artifacts should CAK require?
• How should reviewers update source ledgers, pattern matrices, and unknowns?

Caution:

• A review harness can find disagreements without making the evidence true.

Potential CAK-native adaptation:

• R&D debate packet: Scout, Builder, Skeptic, Alienist, Security Reviewer, Evaluator, Judge. Judge cannot introduce unsupported claims. Every debate output must update source_ledger, pattern_matrix, or unknowns.

Evidence status:

• seed

STORM-like systems

Inspect for:

• multi-perspective question asking;
• outline-first research;
• retrieval-grounded synthesis;
• citation discipline;
• article/survey generation from structured exploration.

Why it matters:

• Useful for research-plan and question-generation mechanics.

Questions it informs:

• How should CAK expand one research question into subquestions?
• Which source classes and perspectives are missing?

Caution:

• Do not assume article generation equals decision-grade architecture research.

Potential CAK-native adaptation:

• Use multi-perspective question generation before source discovery, then force all claims through source ledger and adversarial review.

Evidence status:

• seed

AutoSurvey-like systems

Inspect for:

• paper search agents;
• topic mining and clustering;
• survey writer / quality evaluator split;
• citation coverage;
• multi-agent survey generation;
• structured quality evaluation.

Why it matters:

• Useful for literature synthesis and source coverage, especially for rapidly evolving agent-skill research.

Questions it informs:

• How should CAK separate source discovery, clustering, synthesis, and quality review?

Caution:

• Generated surveys can still miss negative results or implementation realities. Require source ledger and adversarial review.

Potential CAK-native adaptation:

• Use survey-style clustering to organize source discovery, but require implementation artifacts, negative evidence, and pattern transfer checks before decision-ready status.

Evidence status:

• seed

SurveyG / citation-graph approaches

Inspect for:

• hierarchical citation graph;
• foundation/development/frontier layers;
• horizontal and vertical traversal;
• multi-agent validation.

Why it matters:

• Useful for avoiding flat bibliography lists and understanding research lineage.

Questions it informs:

• Which claims are foundation concepts, recent developments, or frontier claims?
• Where is counterevidence missing in the graph?

Caution:

• Citation structure can overvalue popular lineage and miss implementation failures.

Potential CAK-native adaptation:

• Research runs may maintain a reference graph: foundation concepts; recent developments; frontier claims; counterevidence.

Evidence status:

• seed

Older planning / cognitive architecture references

STRIPS / ADL / HTN

Inspect for:

• preconditions;
• postconditions;
• action languages;
• conditional effects;
• hierarchical decomposition.

Why it matters:

• These are human/AI planning ancestors of ContractSpec, workflow memory, and action schemas.

Questions it informs:

• Which planning abstractions transfer to agent runtime contracts?
• Where do explicit world-model assumptions break down?

Caution:

• Classical planning assumes more explicit world models than LLM agents usually have.

Potential CAK-native adaptation:

• Use preconditions, postconditions, and decomposition as reference patterns, but bind them to observable state, verifier checks, and failure traces.

Evidence status:

• seed

Soar / production systems

Inspect for:

• procedural memory;
• working memory;
• if-then production rules;
• operator proposal/evaluation/application;
• chunking.

Why it matters:

• Useful historical analogue for agent-native procedural memory and runtime rule activation.

Questions it informs:

• What should CAK learn from production-rule activation without copying a whole cognitive architecture?

Caution:

• Do not copy cognitive architectures directly; extract mechanisms for runtime control and learning.

Potential CAK-native adaptation:

• Study production activation, operator selection, and chunking as mechanisms for state-conditioned skills and trace-grounded learning.

Evidence status:

• seed

Pattern extraction rule

For every pattern reference, record:

pattern:
source:
what to copy:
what not to copy:
agent-native adaptation:
risk:
evidence status: