crp.advanced¶

Auto-generated reference for the crp.advanced subpackage.

`advanced`¶

`crp.advanced` ¶

Advanced features - hierarchical, parallel, auto-ingest, CQS, meta-learning.

`IngestFact` `dataclass` ¶

Lightweight fact from per-chunk extraction.

`IngestResult` `dataclass` ¶

Summary returned by auto_ingest().

`ContextHungerSignal` `dataclass` ¶

Single context hunger signal detected from LLM output.

Attributes:

Name	Type	Description
`signal_type`	`str`	Signal category ("hedging", "reference_miss", or "repetition").
`strength`	`float`	Normalised signal strength in the range [0.0, 1.0].
`topic`	`str`	Short text snippet identifying the affected topic.
`window_id`	`str`	Identifier of the window that produced the generation.
`token_offset`	`int`	Token offset where the signal was observed.
`details`	`dict[str, Any]`	Additional structured details about the signal.

`CQSDetector` ¶

Detect implicit context hunger from LLM generation output.

`detect_context_hunger(generation_text, window_id='', tokens_generated=None)` ¶

Scan generation output for context hunger signals.

Parameters:

Name	Type	Description	Default
`generation_text`	`str`	Raw text produced by the LLM.	required
`window_id`	`str`	Identifier of the originating window.	`''`
`tokens_generated`	`int \| None`	Number of tokens generated, if known.	`None`

Returns:

Type	Description
`list[ContextHungerSignal]`	List of detected signals (may be empty).

`respond_to_context_hunger(signals, tokens_generated=0)` ¶

Determine action based on detected signals.

§12.4: If max(strength) >= 0.8 AND tokens < 500 → abandon + redispatch. Otherwise → enrich next window.

Parameters:

Name	Type	Description	Default
`signals`	`list[ContextHungerSignal]`	Signals returned by :meth:`detect_context_hunger`.	required
`tokens_generated`	`int`	Number of tokens already generated for the current response.	`0`

Returns:

Type	Description
`CQSResponse`	A `CQSResponse` describing the recommended action, budget, and
`CQSResponse`	enrichment topics.

`CQSResponse` `dataclass` ¶

Response after CQS processing.

Attributes:

Name	Type	Description
`action`	`str`	Recommended action ("abandon_and_redispatch", "enrich_next", or "none").
`signals`	`list[ContextHungerSignal]`	Detected hunger signals that led to the action.
`enrichment_budget`	`int`	Suggested token budget for the next enrichment pass.
`enrichment_topics`	`list[str]`	Topics to target during enrichment.

`ConsistencyIssue` `dataclass` ¶

Single consistency issue found during validation.

Attributes:

Name	Type	Description
`issue_type`	`str`	Category of issue, such as `numerical_contradiction`, `semantic_contradiction`, `undefined_reference`, or `structural_gap`.
`description`	`str`	Human-readable explanation of the issue.
`severity`	`str`	Severity label ("low", "medium", or "high").
`windows`	`list[int] \| None`	Window numbers involved in the issue, if known.
`confirmed`	`bool`	Whether the issue has been confirmed by a higher tier.
`facts`	`list[Any] \| None`	Indices or identifiers of facts involved in the issue.

`CrossWindowValidator` ¶

Three-tier cross-window consistency validation.

Parameters:

Name	Type	Description	Default
`dispatch_fn`	`Callable[[str, str], tuple[str, Any]] \| None`	Callable accepting `(prompt, context)` and returning `(output, metadata)`; used for LLM-based tiers. May be None.	`None`
`embedding_fn`	`Callable[[str], list[float]] \| None`	Callable that returns an embedding vector for a string; used for semantic contradiction detection. May be None.	`None`
`config`	`ReviewCycleConfig \| None`	`ReviewCycleConfig` overrides; defaults are used if None.	`None`

`extraction_based_validation(facts, window_outputs=None, planned_sections=None)` ¶

Zero-cost structural validation. Always runs.

Parameters:

Name	Type	Description	Default
`facts`	`list[dict[str, Any]]`	List of fact dictionaries to validate.	required
`window_outputs`	`list[str] \| None`	Optional aggregated window outputs, used for structural completeness checks.	`None`
`planned_sections`	`list[str] \| None`	Optional list of planned sections to verify.	`None`

Returns:

Type	Description
`ValidationResult`	A Tier 1 `ValidationResult`.

`targeted_llm_validation(tier1_issues)` ¶

Targeted LLM validation for Tier 1 issues.

Parameters:

Name	Type	Description	Default
`tier1_issues`	`list[ConsistencyIssue]`	Issues produced by Tier 1 validation.	required

Returns:

Type	Description
`ValidationResult`	A Tier 2 `ValidationResult` with confirmed/updated issue flags.

`full_llm_review(accumulated_output, task_intent, top_facts=None)` ¶

Dedicated review window with top facts + document map.

Parameters:

Name	Type	Description	Default
`accumulated_output`	`str`	Full generated output to review.	required
`task_intent`	`str`	Original task description.	required
`top_facts`	`list[dict[str, Any]] \| None`	Optional top facts to ground the review.	`None`

Returns:

Type	Description
`ValidationResult`	A Tier 3 `ValidationResult` parsed from the model's review.

`assess_review_capability()` ¶

Probe model to determine max validation tier (1, 2, or 3).

Returns:

Type	Description
`int`	Highest validation tier the configured model can support.

`apply_corrections(issues, task_intent='', blockers=None)` ¶

Apply corrections based on config mode.

Parameters:

Name	Type	Description	Default
`issues`	`list[ConsistencyIssue]`	Issues to process.	required
`task_intent`	`str`	Original task description (unused by current logic).	`''`
`blockers`	`list[str] \| None`	Optional list to populate with flagged issue summaries when `correction_mode` is "flag".	`None`

Returns:

Type	Description
`list[str]`	List of correction actions taken.

`should_run_tier(window_index, tier)` ¶

Check if a validation tier should run at this window index.

Parameters:

Name	Type	Description	Default
`window_index`	`int`	Current window number.	required
`tier`	`int`	Tier to evaluate (1, 2, or 3).	required

Returns:

Type	Description
`bool`	True when the tier is enabled and the window index aligns with its
`bool`	configured interval.

`ValidationResult` `dataclass` ¶

Output of a validation tier.

Attributes:

Name	Type	Description
`tier`	`int`	Tier number (1, 2, or 3).
`issues`	`list[ConsistencyIssue]`	Issues produced by this tier.
`timestamp`	`float`	Unix timestamp of the validation run.
`window_range`	`tuple[int, int]`	Inclusive window range that was validated.

`has_issues` `property` ¶

Return True when the result contains at least one issue.

`high_severity_count` `property` ¶

Return the number of high-severity issues.

`CurationConfig` `dataclass` ¶

Configuration for LLM curation.

`LLMContextCurator` ¶

LLM-driven context curation with progressive understanding.

`current_synthesis` `property` ¶

Return the current synthesis.

`evolution_count` `property` ¶

Return the current evolution count.

`should_curate(window_index)` ¶

Check if curation should run at this window.

`curate(window_index, top_facts, recent_output_summary='')` ¶

Run curation (initial or progressive).

Returns new synthesis or None if dispatch unavailable.

`format_for_envelope()` ¶

Format current synthesis for envelope injection (Section 1.5).

`to_dict()` ¶

Serialize the curator state, history and configuration to a dict.

`LLMSynthesis` `dataclass` ¶

Curated synthesis from LLM review of accumulated facts.

`to_dict()` ¶

Serialize this synthesis to a JSON-ready dict.

`from_dict(data)` `classmethod` ¶

Create a new instance from a dictionary.

Parameters:

Name	Type	Description	Default
`data`	`dict[str, Any]`	The data value.	required

Returns:

Type	Description
`LLMSynthesis`	`LLMSynthesis`.

`FeedbackEntry` `dataclass` ¶

Single human feedback action.

`FeedbackLoop` ¶

Human-in-the-loop corrections for facts in warm state.

`entry_count` `property` ¶

Return the current entry count.

`override_fact(fact_id, corrected_text, reason='')` ¶

Replace fact text with human-provided correction.

`boost_confidence(fact_id, delta=0.1, reason='')` ¶

Increase fact confidence based on human validation.

`penalize_confidence(fact_id, delta=-0.2, reason='')` ¶

Decrease fact confidence based on human rejection.

`reject_fact(fact_id, reason='')` ¶

Mark fact as rejected (confidence → 0).

`get_adjusted_confidence(fact_id, base_confidence)` ¶

Get confidence after applying all feedback adjustments.

`get_entries_for_fact(fact_id)` ¶

Return all feedback entries associated with fact_id.

`to_dict()` ¶

Serialize all feedback entries and cumulative adjustments to a dict.

`HierarchicalPlan` `dataclass` ¶

Plan for hierarchical processing.

`HierarchicalProcessor` ¶

Map-reduce-validate pattern for oversized inputs.

`plan(total_tokens, config=None)` ¶

Create a hierarchical processing plan.

`map_phase(segments, task_intent)` ¶

MAP: Process each segment independently.

`reduce_phase(syntheses, task_intent, fan_in=DEFAULT_FAN_IN)` ¶

REDUCE: Iteratively merge syntheses until ≤ fan_in remain.

`hierarchical_dispatch(task_intent, large_input, config=None)` ¶

Full map-reduce-validate pipeline for oversized input.

Returns (final_syntheses, plan).

`MetaLearningEngine` ¶

ORC + ICML + RTL meta-learning capabilities.

Parameters:

Name	Type	Description	Default
`dispatch_fn`	`Callable[[str, str], tuple[str, Any]] \| None`	Callable accepting `(prompt, context)` and returning `(output, metadata)`. Used to execute reasoning steps and probe the model. May be None for offline planning.	`None`
`model_capability`	`int`	Integer capability level of the active model.	`1`
`config`	`MetaLearningConfig \| None`	`MetaLearningConfig` overrides; defaults are used if None.	`None`

`trace_count` `property` ¶

Return the number of traces currently stored in the RTL.

`should_use_orc(task_complexity=3, resource_pressure='NONE', probe_quality=0.0)` ¶

Gate check for ORC activation.

Gate 1: resource_pressure >= HIGH → False Gate 2: model_capability >= task_complexity → False Gate 3: probe_quality >= 0.7 → False (ORC unnecessary)

Parameters:

Name	Type	Description	Default
`task_complexity`	`int`	Estimated complexity of the task on a 1–5 scale.	`3`
`resource_pressure`	`str`	Resource pressure label ("NONE", "LOW", "MODERATE", "HIGH", "CRITICAL").	`'NONE'`
`probe_quality`	`float`	Quality score from a zero-shot probe; high values indicate ORC is unnecessary.	`0.0`

Returns:

Type	Description
`bool`	True when ORC should be used to break the task into steps.

`orchestrated_reasoning(task_intent, task_complexity=3, resource_pressure='NONE')` ¶

Decompose and execute an orchestrated reasoning chain.

Parameters:

Name	Type	Description	Default
`task_intent`	`str`	Natural-language description of the task to solve.	required
`task_complexity`	`int`	Estimated complexity of the task.	`3`
`resource_pressure`	`str`	Resource pressure label; reduces the number of allowed steps under higher pressure.	`'NONE'`

Returns:

Type	Description
`ORCResult`	An `ORCResult` containing step outputs and the synthesised final
`ORCResult`	answer.

`build_reasoning_scaffold(task_intent)` ¶

Build a reasoning scaffold adapted to model capability.

Capability ≤ 1 (0.5B-1B): Full step-by-step template Capability ≤ 2 (2B-7B): Light approach Capability > 2: No scaffolding

Parameters:

Name	Type	Description	Default
`task_intent`	`str`	Task description to scaffold.	required

Returns:

Type	Description
`str`	Scaffold string to prepend to the prompt, or an empty string when
`str`	scaffolding is disabled.

`build_metacognitive_envelope(task_intent, base_envelope='', few_shot_traces=None)` ¶

Build an envelope with reasoning scaffold + few-shot examples.

Parameters:

Name	Type	Description	Default
`task_intent`	`str`	Task description.	required
`base_envelope`	`str`	Existing envelope text to preserve, if any.	`''`
`few_shot_traces`	`list[ReasoningTrace] \| None`	Optional explicit traces to include as examples. When omitted, the RTL is queried for matching traces.	`None`

Returns:

Type	Description
`str`	Combined envelope string containing the base envelope, scaffold,
`str`	and any few-shot examples.

`store_trace(trace)` ¶

Store a reasoning trace if quality meets the configured threshold.

Parameters:

Name	Type	Description	Default
`trace`	`ReasoningTrace`	`ReasoningTrace` to store.	required

Returns:

Type	Description
`bool`	True when the trace was stored, False when RTL is disabled or the
`bool`	trace quality is too low.

`to_dict()` ¶

Serialize the engine and its trace library.

Returns:

Type	Description
`dict[str, Any]`	Dictionary with the trace library and active configuration.

`ORCResult` `dataclass` ¶

Result of orchestrated reasoning chain.

Attributes:

Name	Type	Description
`steps_completed`	`int`	Number of steps that produced output.
`steps_total`	`int`	Total number of steps in the chain.
`final_output`	`str`	Synthesised final response.
`step_outputs`	`list[str]`	Raw output from each executed step.
`quality_score`	`float`	Estimated quality score for the chain.
`trace`	`ReasoningTrace \| None`	Optional `ReasoningTrace` captured for the RTL.

`ReasoningTrace` `dataclass` ¶

Complete reasoning trace for RTL storage.

Attributes:

Name	Type	Description
`trace_id`	`str`	Unique identifier for the trace.
`task_type`	`str`	Category of task this trace addresses.
`task_summary`	`str`	Short summary of the original task.
`steps`	`list[ReasoningStep]`	Ordered reasoning steps that produced the result.
`model_class`	`str`	Model capability class (e.g., "0.5B-1B", "2B-7B", "7B+").
`quality_score`	`float`	Quality score assigned to the trace.
`created_at`	`float`	Unix timestamp when the trace was created.
`usage_count`	`int`	Number of times the trace has been retrieved.

`to_dict()` ¶

Serialize the trace to a JSON-friendly dictionary.

Returns:

Type	Description
`dict[str, Any]`	Dictionary representation of the trace, including nested steps.

`from_dict(data)` `classmethod` ¶

Restore a ReasoningTrace from a dictionary.

Parameters:

Name	Type	Description	Default
`data`	`dict[str, Any]`	Serialized trace data produced by :meth:`to_dict`.	required

Returns:

Type	Description
`ReasoningTrace`	Reconstructed `ReasoningTrace` instance.

`FanOutResult` `dataclass` ¶

Result of one parallel dispatch.

`FanOutTask` `dataclass` ¶

One independent task for parallel dispatch.

`ParallelFanOut` ¶

Dispatch N independent windows and merge results.

Algorithm (§4.4): 1. Identify N independent tasks 2. Construct independent envelopes from warm_state 3. Dispatch all N windows (sequential fallback if no async) 4. Collect all N outputs 5. Extract facts from all N outputs 6. Merge facts into warm_state 7. Update DAG with fan-out edges 8. Continue with next dependent task

`fan_out(tasks)` ¶

Dispatch tasks (sequentially - async version would override).

Returns results in same order as tasks.

`merge_results(results, existing_facts=None)` ¶

Merge fan-out results. Successful results first, failures last.

`AssessmentResult` `dataclass` ¶

Output from post-generation self-assessment.

`ReviewCycleManager` ¶

Active LLM review cycles - planning, checkpoint, assessment.

`pre_generation_plan(task_intent, predicted_chain_length=0)` ¶

Generate document plan when chain > 5 windows.

Returns None if chain is short or no dispatch_fn.

`checkpoint_review(window_index, review_interval=20, task_intent='', top_facts=None, gap_summary='')` ¶

Periodic review at checkpoint windows.

Gate: model_capability < 3 → None Gate: window_index not at interval → None

`post_generation_assessment(accumulated_output, task_intent)` ¶

Score output quality and flag issues.

Weak model → basic heuristic scoring. Strong model → full LLM self-assessment.

`targeted_regeneration(issues, task_intent)` ¶

Re-generate targeted fixes for each issue (capped at max_corrections).

`ReviewGuidance` `dataclass` ¶

Output from a checkpoint review.

`QualityTier` ¶

Bases: IntEnum

Quality tiers - S (single window) through D (>1000 windows at 128K ctx).

`ScaleModeSelector` ¶

Auto-configure session based on quality tier and model capability.

`configure_session(estimated_tokens, model_capability=1)` ¶

Auto-configure session based on input size and model capability.

Parameters:

Name	Type	Description	Default
`estimated_tokens`	`int`	Total estimated input tokens.	required
`model_capability`	`int`	Assessed model capability (1, 2, or 3).	`1`

Returns:

Type	Description
`SessionConfig`	SessionConfig with all parameters set.

`SessionConfig` `dataclass` ¶

Auto-configured session parameters.

`SourceGroundingEngine` ¶

Store and retrieve verbatim source passages for high-confidence facts.

Parameters:

Name	Type	Description	Default
`count_tokens`	`Callable[[str], int] \| None`	Callable that returns the token count for a piece of text. Defaults to a rough character-based estimate when None.	`None`

`passage_count` `property` ¶

Return the number of stored passages.

`store_passage(passage, fact_confidence=0.0)` ¶

Store a passage if its linked fact has confidence ≥ threshold.

Parameters:

Name	Type	Description	Default
`passage`	`SourcePassage`	Passage to store.	required
`fact_confidence`	`float`	Confidence score of the fact linked to the passage.	`0.0`

Returns:

Type	Description
`bool`	True if the passage was stored, False if it was below the threshold.

`get_passages_for_fact(fact_id)` ¶

Retrieve all source passages linked to a fact.

Parameters:

Name	Type	Description	Default
`fact_id`	`str`	Identifier of the fact.	required

Returns:

Type	Description
`list[SourcePassage]`	List of passages linked to the fact that remain in storage.

`build_source_grounded_envelope(scored_facts, budget_tokens, quality_tier='B')` ¶

Build an envelope with source passages allocated by tier.

Parameters:

Name	Type	Description	Default
`scored_facts`	`list[dict[str, Any]]`	Sorted list of fact dictionaries such as `{"id": ..., "text": ..., "score": ...}`.	required
`budget_tokens`	`int`	Total envelope budget in tokens.	required
`quality_tier`	`str`	Quality tier (S/A/B/C/D) controlling the fact/source budget split.	`'B'`

Returns:

Type	Description
`tuple[list[dict[str, Any]], list[SourcePassage]]`	A tuple of `(packed_facts, included_passages)`.

`format_envelope_section(fact, passages)` ¶

Format a fact with its source passages for envelope inclusion.

Format::

- {fact text} - Window N
  ↳ [SOURCE: Window N, tokens X-Y]
    "{verbatim original text}"

Parameters:

Name	Type	Description	Default
`fact`	`dict[str, Any]`	Fact dictionary containing at least `text` and `window`.	required
`passages`	`list[SourcePassage]`	Passages linked to the fact.	required

Returns:

Type	Description
`str`	Formatted envelope section string.

`to_dict()` ¶

Serialize the engine for persistence.

Returns:

Type	Description
`dict[str, Any]`	Dictionary with all passages and fact-to-passage mappings.

`from_dict(data, count_tokens=None)` `classmethod` ¶

Restore the engine from serialized state.

Parameters:

Name	Type	Description	Default
`data`	`dict[str, Any]`	Serialized state produced by :meth:`to_dict`.	required
`count_tokens`	`Callable[[str], int] \| None`	Optional token-counting callable; defaults to the engine's fallback estimator if None.	`None`

Returns:

Type	Description
`SourceGroundingEngine`	Reconstructed `SourceGroundingEngine` instance.

`SourcePassage` `dataclass` ¶

Verbatim passage from the original input linked to facts.

Attributes:

Name	Type	Description
`passage_id`	`str`	Unique identifier for the passage.
`text`	`str`	Verbatim source text.
`source_window`	`int`	Window number from which the passage was extracted.
`token_offset_start`	`int`	Start token offset within the source window.
`token_offset_end`	`int`	End token offset within the source window.
`linked_fact_ids`	`list[str]`	Identifiers of facts this passage grounds.
`token_count`	`int`	Token count of `text`.
`relevance_score`	`float`	Relevance score for retrieval ranking.

`to_dict()` ¶

Serialize the passage to a dictionary.

Returns:

Type	Description
`dict[str, Any]`	Dictionary with all passage fields.

`from_dict(data)` `classmethod` ¶

Restore a passage from a serialized dictionary.

Parameters:

Name	Type	Description	Default
`data`	`dict[str, Any]`	Serialized passage produced by :meth:`to_dict`.	required

Returns:

Type	Description
`SourcePassage`	Reconstructed `SourcePassage` instance.

`advanced.auto_ingest`¶

`crp.advanced.auto_ingest` ¶

Auto-ingest - oversized input handling with structure-aware chunking (§4.6).

Triggers when system_tokens + task_tokens > context_window - gen_reserve. Zero LLM cost by default: uses graduated extraction (stages 1-5) per chunk, then reconciles boundary duplicates/complements via embedding similarity.

`ProtectedSpan` `dataclass` ¶

Region that must not be split mid-structure.

`Chunk` `dataclass` ¶

One chunk of the oversized input.

`IngestResult` `dataclass` ¶

Summary returned by auto_ingest().

`IngestFact` `dataclass` ¶

Lightweight fact from per-chunk extraction.

`detect_protected_structures(text)` ¶

Find code blocks, tables, JSON blocks, numbered lists.

`merge_overlapping_spans(spans)` ¶

Merge overlapping/adjacent protected spans.

`split_at_boundaries(text, chunk_size_chars, overlap_chars, protected_spans)` ¶

Split text into chunks respecting protected structures.

`reconcile_chunk_boundaries(per_chunk_facts, embedding_fn=None)` ¶

Deduplicate/merge facts at chunk boundaries.

cosine_similarity > 0.95 → duplicate → skip
cosine_similarity > 0.75 AND token_overlap > 0.3 → complement → merge
Otherwise → new fact → keep

`merge_fact_texts(a, b)` ¶

Merge two complementary fact texts, keeping unique content.

`auto_ingest(system_prompt, task_input, task_intent_text, context_window, count_tokens, extract_fn=None, embedding_fn=None, store_raw_fn=None, session_id='')` ¶

Handle oversized inputs with structure-aware chunking.

Parameters:

Name	Type	Description	Default
`system_prompt`	`str`	The system prompt (not modified).	required
`task_input`	`str`	Raw oversized input text.	required
`task_intent_text`	`str`	Short description of task intent.	required
`context_window`	`int`	Total context window in tokens.	required
`count_tokens`	`Callable[[str], int]`	Token counting function.	required
`extract_fn`	`Callable[[str, str], list[IngestFact]] \| None`	Per-chunk fact extractor (stages 1-5). If None, returns dummy facts.	`None`
`embedding_fn`	`Callable[[str], list[float]] \| None`	Optional embedding function for reconciliation.	`None`
`store_raw_fn`	`Callable[[str, str], None] \| None`	Optional function to store raw input in cold storage.	`None`
`session_id`	`str`	Current session ID.	`''`

Returns:

Type	Description
`tuple[list[IngestFact], IngestResult]`	(reconciled_facts, ingest_result)

`advanced.cqs`¶

`crp.advanced.cqs` ¶

CQS - Context Quality Signaling, detect LLM context hunger (§12; CRP-SPEC-019).

Three signal types: hedging, reference_miss, repetition. Preserves Model Ignorance (Axiom 4): signals are detected structurally from generation output, never by injecting meta-protocol into the LLM.

Relevant specifications

CRP specification §12: Context Quality Signaling
CRP-SPEC-019: Cognitive Quality Recognition (CQR)

`ContextHungerSignal` `dataclass` ¶

Single context hunger signal detected from LLM output.

Attributes:

Name	Type	Description
`signal_type`	`str`	Signal category ("hedging", "reference_miss", or "repetition").
`strength`	`float`	Normalised signal strength in the range [0.0, 1.0].
`topic`	`str`	Short text snippet identifying the affected topic.
`window_id`	`str`	Identifier of the window that produced the generation.
`token_offset`	`int`	Token offset where the signal was observed.
`details`	`dict[str, Any]`	Additional structured details about the signal.

`CQSResponse` `dataclass` ¶

Response after CQS processing.

Attributes:

Name	Type	Description
`action`	`str`	Recommended action ("abandon_and_redispatch", "enrich_next", or "none").
`signals`	`list[ContextHungerSignal]`	Detected hunger signals that led to the action.
`enrichment_budget`	`int`	Suggested token budget for the next enrichment pass.
`enrichment_topics`	`list[str]`	Topics to target during enrichment.

`CQSDetector` ¶

Detect implicit context hunger from LLM generation output.

`detect_context_hunger(generation_text, window_id='', tokens_generated=None)` ¶

Scan generation output for context hunger signals.

Parameters:

Name	Type	Description	Default
`generation_text`	`str`	Raw text produced by the LLM.	required
`window_id`	`str`	Identifier of the originating window.	`''`
`tokens_generated`	`int \| None`	Number of tokens generated, if known.	`None`

Returns:

Type	Description
`list[ContextHungerSignal]`	List of detected signals (may be empty).

`respond_to_context_hunger(signals, tokens_generated=0)` ¶

Determine action based on detected signals.

§12.4: If max(strength) >= 0.8 AND tokens < 500 → abandon + redispatch. Otherwise → enrich next window.

Parameters:

Name	Type	Description	Default
`signals`	`list[ContextHungerSignal]`	Signals returned by :meth:`detect_context_hunger`.	required
`tokens_generated`	`int`	Number of tokens already generated for the current response.	`0`

Returns:

Type	Description
`CQSResponse`	A `CQSResponse` describing the recommended action, budget, and
`CQSResponse`	enrichment topics.

`advanced.cross_window`¶

`crp.advanced.cross_window` ¶

Cross-window validation - 3-tier consistency checks (§13; CRP-SPEC-019).

Tier 1: Extraction-based (always, zero LLM cost) Tier 2: LLM-targeted (2B+ models) Tier 3: Full LLM review (7B+ models)

Relevant specifications

CRP specification §13: Cross-window validation
CRP-SPEC-019: Cognitive Quality Recognition (CQR)

`ConsistencyIssue` `dataclass` ¶

Single consistency issue found during validation.

Attributes:

Name	Type	Description
`issue_type`	`str`	Category of issue, such as `numerical_contradiction`, `semantic_contradiction`, `undefined_reference`, or `structural_gap`.
`description`	`str`	Human-readable explanation of the issue.
`severity`	`str`	Severity label ("low", "medium", or "high").
`windows`	`list[int] \| None`	Window numbers involved in the issue, if known.
`confirmed`	`bool`	Whether the issue has been confirmed by a higher tier.
`facts`	`list[Any] \| None`	Indices or identifiers of facts involved in the issue.

`ValidationResult` `dataclass` ¶

Output of a validation tier.

Attributes:

Name	Type	Description
`tier`	`int`	Tier number (1, 2, or 3).
`issues`	`list[ConsistencyIssue]`	Issues produced by this tier.
`timestamp`	`float`	Unix timestamp of the validation run.
`window_range`	`tuple[int, int]`	Inclusive window range that was validated.

`has_issues` `property` ¶

Return True when the result contains at least one issue.

`high_severity_count` `property` ¶

Return the number of high-severity issues.

`ReviewCycleConfig` `dataclass` ¶

Configuration for review cycles.

Attributes:

Name	Type	Description
`enabled`	`bool`	Master switch for review cycles.
`tier_1_interval`	`int`	Window interval between Tier 1 validations.
`tier_2_enabled`	`bool`	Enable Tier 2 LLM-targeted validation.
`tier_2_interval`	`int`	Window interval between Tier 2 validations.
`tier_3_enabled`	`bool`	Enable Tier 3 full LLM review.
`tier_3_interval`	`int`	Window interval between Tier 3 validations.
`tier_3_min_model_capability`	`int`	Minimum model capability for Tier 3.
`correction_mode`	`str`	Correction behaviour ("flag" or "correct").
`max_correction_windows`	`int`	Maximum windows to attempt correcting.

`CrossWindowValidator` ¶

Three-tier cross-window consistency validation.

Parameters:

Name	Type	Description	Default
`dispatch_fn`	`Callable[[str, str], tuple[str, Any]] \| None`	Callable accepting `(prompt, context)` and returning `(output, metadata)`; used for LLM-based tiers. May be None.	`None`
`embedding_fn`	`Callable[[str], list[float]] \| None`	Callable that returns an embedding vector for a string; used for semantic contradiction detection. May be None.	`None`
`config`	`ReviewCycleConfig \| None`	`ReviewCycleConfig` overrides; defaults are used if None.	`None`

`extraction_based_validation(facts, window_outputs=None, planned_sections=None)` ¶

Zero-cost structural validation. Always runs.

Parameters:

Name	Type	Description	Default
`facts`	`list[dict[str, Any]]`	List of fact dictionaries to validate.	required
`window_outputs`	`list[str] \| None`	Optional aggregated window outputs, used for structural completeness checks.	`None`
`planned_sections`	`list[str] \| None`	Optional list of planned sections to verify.	`None`

Returns:

Type	Description
`ValidationResult`	A Tier 1 `ValidationResult`.

`targeted_llm_validation(tier1_issues)` ¶

Targeted LLM validation for Tier 1 issues.

Parameters:

Name	Type	Description	Default
`tier1_issues`	`list[ConsistencyIssue]`	Issues produced by Tier 1 validation.	required

Returns:

Type	Description
`ValidationResult`	A Tier 2 `ValidationResult` with confirmed/updated issue flags.

`full_llm_review(accumulated_output, task_intent, top_facts=None)` ¶

Dedicated review window with top facts + document map.

Parameters:

Name	Type	Description	Default
`accumulated_output`	`str`	Full generated output to review.	required
`task_intent`	`str`	Original task description.	required
`top_facts`	`list[dict[str, Any]] \| None`	Optional top facts to ground the review.	`None`

Returns:

Type	Description
`ValidationResult`	A Tier 3 `ValidationResult` parsed from the model's review.

`assess_review_capability()` ¶

Probe model to determine max validation tier (1, 2, or 3).

Returns:

Type	Description
`int`	Highest validation tier the configured model can support.

`apply_corrections(issues, task_intent='', blockers=None)` ¶

Apply corrections based on config mode.

Parameters:

Name	Type	Description	Default
`issues`	`list[ConsistencyIssue]`	Issues to process.	required
`task_intent`	`str`	Original task description (unused by current logic).	`''`
`blockers`	`list[str] \| None`	Optional list to populate with flagged issue summaries when `correction_mode` is "flag".	`None`

Returns:

Type	Description
`list[str]`	List of correction actions taken.

`should_run_tier(window_index, tier)` ¶

Check if a validation tier should run at this window index.

Parameters:

Name	Type	Description	Default
`window_index`	`int`	Current window number.	required
`tier`	`int`	Tier to evaluate (1, 2, or 3).	required

Returns:

Type	Description
`bool`	True when the tier is enabled and the window index aligns with its
`bool`	configured interval.

`advanced.curator`¶

`crp.advanced.curator` ¶

LLM context curation - progressive understanding synthesis (§18).

Periodically dispatches curation windows to build an evolving synthesis of findings, relationships, and gaps. Injected into envelopes as Section 1.5 between CRITICAL STATE and DISCOVERIES.

`LLMSynthesis` `dataclass` ¶

Curated synthesis from LLM review of accumulated facts.

`to_dict()` ¶

Serialize this synthesis to a JSON-ready dict.

`from_dict(data)` `classmethod` ¶

Create a new instance from a dictionary.

Parameters:

Name	Type	Description	Default
`data`	`dict[str, Any]`	The data value.	required

Returns:

Type	Description
`LLMSynthesis`	`LLMSynthesis`.

`CurationConfig` `dataclass` ¶

Configuration for LLM curation.

`LLMContextCurator` ¶

LLM-driven context curation with progressive understanding.

`current_synthesis` `property` ¶

Return the current synthesis.

`evolution_count` `property` ¶

Return the current evolution count.

`should_curate(window_index)` ¶

Check if curation should run at this window.

`curate(window_index, top_facts, recent_output_summary='')` ¶

Run curation (initial or progressive).

Returns new synthesis or None if dispatch unavailable.

`format_for_envelope()` ¶

Format current synthesis for envelope injection (Section 1.5).

`to_dict()` ¶

Serialize the curator state, history and configuration to a dict.

`advanced.feedback`¶

`crp.advanced.feedback` ¶

Human-in-the-loop feedback - fact override, confidence adjustment (§18, MAY).

`FeedbackEntry` `dataclass` ¶

Single human feedback action.

`FeedbackLoop` ¶

Human-in-the-loop corrections for facts in warm state.

`entry_count` `property` ¶

Return the current entry count.

`override_fact(fact_id, corrected_text, reason='')` ¶

Replace fact text with human-provided correction.

`boost_confidence(fact_id, delta=0.1, reason='')` ¶

Increase fact confidence based on human validation.

`penalize_confidence(fact_id, delta=-0.2, reason='')` ¶

Decrease fact confidence based on human rejection.

`reject_fact(fact_id, reason='')` ¶

Mark fact as rejected (confidence → 0).

`get_adjusted_confidence(fact_id, base_confidence)` ¶

Get confidence after applying all feedback adjustments.

`get_entries_for_fact(fact_id)` ¶

Return all feedback entries associated with fact_id.

`to_dict()` ¶

Serialize all feedback entries and cumulative adjustments to a dict.

`advanced.hierarchical`¶

`crp.advanced.hierarchical` ¶

Hierarchical processing - map-reduce-validate for Tier C/D inputs (§4.5, §11).

Splits massive inputs into segments, processes each independently, reduces iteratively, and validates cross-window consistency.

`HierarchicalPlan` `dataclass` ¶

Plan for hierarchical processing.

`HierarchicalConfig` `dataclass` ¶

Configuration for hierarchical processing.

`SegmentResult` `dataclass` ¶

Output of processing one segment.

`HierarchicalProcessor` ¶

Map-reduce-validate pattern for oversized inputs.

`plan(total_tokens, config=None)` ¶

Create a hierarchical processing plan.

`map_phase(segments, task_intent)` ¶

MAP: Process each segment independently.

`reduce_phase(syntheses, task_intent, fan_in=DEFAULT_FAN_IN)` ¶

REDUCE: Iteratively merge syntheses until ≤ fan_in remain.

`hierarchical_dispatch(task_intent, large_input, config=None)` ¶

Full map-reduce-validate pipeline for oversized input.

Returns (final_syntheses, plan).

`chain_degradation(levels, per_level=0.03)` ¶

Compute effective degradation after N hierarchy levels.

d_chain(L) = 1 - (1 - per_level)^L

`effective_context(context_window, levels, per_level=0.03)` ¶

Effective context capacity after hierarchical degradation.

EffCtx_hier(N) = C × (1 - d_chain(⌈log_k(N)⌉))

`advanced.meta_learning`¶

`crp.advanced.meta_learning` ¶

Meta-learning - ORC, ICML, Reasoning Template Library (§19; CRP-SPEC-019).

Three mechanisms

Orchestrated Reasoning Chains (ORC): decompose complex reasoning into micro-steps
In-Context Meta-Learning (ICML): reasoning scaffolds + few-shot examples
Reasoning Template Library (RTL): store/retrieve successful reasoning traces

Relevant specifications

CRP-SPEC-019: Cognitive Quality Recognition (CQR)
CRP specification §19: Meta-learning / amplified reasoning

`ReasoningStep` `dataclass` ¶

Single step in a reasoning chain.

Attributes:

Name	Type	Description
`step_description`	`str`	Human-readable description of what this step does.
`system_prompt_template`	`str`	Prompt template used to execute the step.
`expected_output_format`	`str`	Description of the output shape expected from the step.
`scaffold_level`	`int`	Amount of scaffolding to apply (0–3).

`ReasoningTrace` `dataclass` ¶

Complete reasoning trace for RTL storage.

Attributes:

Name	Type	Description
`trace_id`	`str`	Unique identifier for the trace.
`task_type`	`str`	Category of task this trace addresses.
`task_summary`	`str`	Short summary of the original task.
`steps`	`list[ReasoningStep]`	Ordered reasoning steps that produced the result.
`model_class`	`str`	Model capability class (e.g., "0.5B-1B", "2B-7B", "7B+").
`quality_score`	`float`	Quality score assigned to the trace.
`created_at`	`float`	Unix timestamp when the trace was created.
`usage_count`	`int`	Number of times the trace has been retrieved.

`to_dict()` ¶

Serialize the trace to a JSON-friendly dictionary.

Returns:

Type	Description
`dict[str, Any]`	Dictionary representation of the trace, including nested steps.

`from_dict(data)` `classmethod` ¶

Restore a ReasoningTrace from a dictionary.

Parameters:

Name	Type	Description	Default
`data`	`dict[str, Any]`	Serialized trace data produced by :meth:`to_dict`.	required

Returns:

Type	Description
`ReasoningTrace`	Reconstructed `ReasoningTrace` instance.

`ORCResult` `dataclass` ¶

Result of orchestrated reasoning chain.

Attributes:

Name	Type	Description
`steps_completed`	`int`	Number of steps that produced output.
`steps_total`	`int`	Total number of steps in the chain.
`final_output`	`str`	Synthesised final response.
`step_outputs`	`list[str]`	Raw output from each executed step.
`quality_score`	`float`	Estimated quality score for the chain.
`trace`	`ReasoningTrace \| None`	Optional `ReasoningTrace` captured for the RTL.

`MetaLearningConfig` `dataclass` ¶

Configuration for meta-learning features.

Attributes:

Name	Type	Description
`enabled`	`bool`	Master switch for meta-learning.
`orc_enabled`	`bool`	Enable Orchestrated Reasoning Chains.
`orc_max_steps`	`int`	Maximum number of reasoning steps allowed in ORC.
`orc_min_model_capability`	`int`	Minimum model capability required for ORC.
`icml_enabled`	`bool`	Enable In-Context Meta-Learning.
`icml_max_examples`	`int`	Maximum few-shot examples to inject.
`rtl_enabled`	`bool`	Enable Reasoning Template Library storage/retrieval.
`rtl_min_quality_for_storage`	`float`	Minimum quality score for storing a trace.
`scaffold_level`	`str`	Default scaffolding level ("auto", "none", "light", "heavy").
`curation_interval`	`int`	Number of windows between RTL curation passes.

`MetaLearningEngine` ¶

ORC + ICML + RTL meta-learning capabilities.

Parameters:

Name	Type	Description	Default
`dispatch_fn`	`Callable[[str, str], tuple[str, Any]] \| None`	Callable accepting `(prompt, context)` and returning `(output, metadata)`. Used to execute reasoning steps and probe the model. May be None for offline planning.	`None`
`model_capability`	`int`	Integer capability level of the active model.	`1`
`config`	`MetaLearningConfig \| None`	`MetaLearningConfig` overrides; defaults are used if None.	`None`

`trace_count` `property` ¶

Return the number of traces currently stored in the RTL.

`should_use_orc(task_complexity=3, resource_pressure='NONE', probe_quality=0.0)` ¶

Gate check for ORC activation.

Gate 1: resource_pressure >= HIGH → False Gate 2: model_capability >= task_complexity → False Gate 3: probe_quality >= 0.7 → False (ORC unnecessary)

Parameters:

Name	Type	Description	Default
`task_complexity`	`int`	Estimated complexity of the task on a 1–5 scale.	`3`
`resource_pressure`	`str`	Resource pressure label ("NONE", "LOW", "MODERATE", "HIGH", "CRITICAL").	`'NONE'`
`probe_quality`	`float`	Quality score from a zero-shot probe; high values indicate ORC is unnecessary.	`0.0`

Returns:

Type	Description
`bool`	True when ORC should be used to break the task into steps.

`orchestrated_reasoning(task_intent, task_complexity=3, resource_pressure='NONE')` ¶

Decompose and execute an orchestrated reasoning chain.

Parameters:

Name	Type	Description	Default
`task_intent`	`str`	Natural-language description of the task to solve.	required
`task_complexity`	`int`	Estimated complexity of the task.	`3`
`resource_pressure`	`str`	Resource pressure label; reduces the number of allowed steps under higher pressure.	`'NONE'`

Returns:

Type	Description
`ORCResult`	An `ORCResult` containing step outputs and the synthesised final
`ORCResult`	answer.

`build_reasoning_scaffold(task_intent)` ¶

Build a reasoning scaffold adapted to model capability.

Capability ≤ 1 (0.5B-1B): Full step-by-step template Capability ≤ 2 (2B-7B): Light approach Capability > 2: No scaffolding

Parameters:

Name	Type	Description	Default
`task_intent`	`str`	Task description to scaffold.	required

Returns:

Type	Description
`str`	Scaffold string to prepend to the prompt, or an empty string when
`str`	scaffolding is disabled.

`build_metacognitive_envelope(task_intent, base_envelope='', few_shot_traces=None)` ¶

Build an envelope with reasoning scaffold + few-shot examples.

Parameters:

Name	Type	Description	Default
`task_intent`	`str`	Task description.	required
`base_envelope`	`str`	Existing envelope text to preserve, if any.	`''`
`few_shot_traces`	`list[ReasoningTrace] \| None`	Optional explicit traces to include as examples. When omitted, the RTL is queried for matching traces.	`None`

Returns:

Type	Description
`str`	Combined envelope string containing the base envelope, scaffold,
`str`	and any few-shot examples.

`store_trace(trace)` ¶

Store a reasoning trace if quality meets the configured threshold.

Parameters:

Name	Type	Description	Default
`trace`	`ReasoningTrace`	`ReasoningTrace` to store.	required

Returns:

Type	Description
`bool`	True when the trace was stored, False when RTL is disabled or the
`bool`	trace quality is too low.

`to_dict()` ¶

Serialize the engine and its trace library.

Returns:

Type	Description
`dict[str, Any]`	Dictionary with the trace library and active configuration.

`advanced.parallel`¶

`crp.advanced.parallel` ¶

Parallel fan-out - N independent windows dispatched concurrently (§4.4).

`FanOutTask` `dataclass` ¶

One independent task for parallel dispatch.

`FanOutResult` `dataclass` ¶

Result of one parallel dispatch.

`ParallelFanOut` ¶

Dispatch N independent windows and merge results.

Algorithm (§4.4): 1. Identify N independent tasks 2. Construct independent envelopes from warm_state 3. Dispatch all N windows (sequential fallback if no async) 4. Collect all N outputs 5. Extract facts from all N outputs 6. Merge facts into warm_state 7. Update DAG with fan-out edges 8. Continue with next dependent task

`fan_out(tasks)` ¶

Dispatch tasks (sequentially - async version would override).

Returns results in same order as tasks.

`merge_results(results, existing_facts=None)` ¶

Merge fan-out results. Successful results first, failures last.

`advanced.review_cycle`¶

`crp.advanced.review_cycle` ¶

Review cycle management - active LLM review patterns (§14).

Three interaction patterns

Pre-generation planning (predict chain > 5 windows)
Checkpoint review (periodic, Tier 3 models only)
Post-generation self-assessment (quality scoring + targeted re-gen)

`ReviewGuidance` `dataclass` ¶

Output from a checkpoint review.

`AssessmentResult` `dataclass` ¶

Output from post-generation self-assessment.

`PlannedSection` `dataclass` ¶

One section in the generation plan.

`DocumentPlan` `dataclass` ¶

Full generation plan from pre-generation planning.

`ReviewCycleManager` ¶

Active LLM review cycles - planning, checkpoint, assessment.

`pre_generation_plan(task_intent, predicted_chain_length=0)` ¶

Generate document plan when chain > 5 windows.

Returns None if chain is short or no dispatch_fn.

`checkpoint_review(window_index, review_interval=20, task_intent='', top_facts=None, gap_summary='')` ¶

Periodic review at checkpoint windows.

Gate: model_capability < 3 → None Gate: window_index not at interval → None

`post_generation_assessment(accumulated_output, task_intent)` ¶

Score output quality and flag issues.

Weak model → basic heuristic scoring. Strong model → full LLM self-assessment.

`targeted_regeneration(issues, task_intent)` ¶

Re-generate targeted fixes for each issue (capped at max_corrections).

`advanced.scale_mode`¶

`crp.advanced.scale_mode` ¶

Scale-mode selector - auto-configure session by quality tier (§8.3, §15).

Classifies input into quality tiers S/A/B/C/D based on token ratio, then configures processing mode, validation tiers, review cycles, etc.

`QualityTier` ¶

Bases: IntEnum

Quality tiers - S (single window) through D (>1000 windows at 128K ctx).

`SessionConfig` `dataclass` ¶

Auto-configured session parameters.

`ScaleModeSelector` ¶

Auto-configure session based on quality tier and model capability.

`configure_session(estimated_tokens, model_capability=1)` ¶

Auto-configure session based on input size and model capability.

Parameters:

Name	Type	Description	Default
`estimated_tokens`	`int`	Total estimated input tokens.	required
`model_capability`	`int`	Assessed model capability (1, 2, or 3).	`1`

Returns:

Type	Description
`SessionConfig`	SessionConfig with all parameters set.

`classify_quality_tier(estimated_tokens, context_window)` ¶

Classify input into quality tier based on token-to-context ratio.

`select_processing_mode(estimated_tokens, context_window)` ¶

Select processing mode based on windows needed.

`advanced.source_grounding`¶

`crp.advanced.source_grounding` ¶

Source grounding - store/retrieve verbatim source passages (§17).

Stores passages for facts with confidence ≥ 0.8. Integrates passages into envelopes with tier-based budget allocation.

Relevant specifications

CRP specification §17: Source grounding
CRP-SPEC-024: Coverage Differential Retrieval (CDR)
CRP-SPEC-025: Context Differential Graph Retrieval (CDGR)

`SourcePassage` `dataclass` ¶

Verbatim passage from the original input linked to facts.

Attributes:

Name	Type	Description
`passage_id`	`str`	Unique identifier for the passage.
`text`	`str`	Verbatim source text.
`source_window`	`int`	Window number from which the passage was extracted.
`token_offset_start`	`int`	Start token offset within the source window.
`token_offset_end`	`int`	End token offset within the source window.
`linked_fact_ids`	`list[str]`	Identifiers of facts this passage grounds.
`token_count`	`int`	Token count of `text`.
`relevance_score`	`float`	Relevance score for retrieval ranking.

`to_dict()` ¶

Serialize the passage to a dictionary.

Returns:

Type	Description
`dict[str, Any]`	Dictionary with all passage fields.

`from_dict(data)` `classmethod` ¶

Restore a passage from a serialized dictionary.

Parameters:

Name	Type	Description	Default
`data`	`dict[str, Any]`	Serialized passage produced by :meth:`to_dict`.	required

Returns:

Type	Description
`SourcePassage`	Reconstructed `SourcePassage` instance.

`SourceGroundingEngine` ¶

Store and retrieve verbatim source passages for high-confidence facts.

Parameters:

Name	Type	Description	Default
`count_tokens`	`Callable[[str], int] \| None`	Callable that returns the token count for a piece of text. Defaults to a rough character-based estimate when None.	`None`

`passage_count` `property` ¶

Return the number of stored passages.

`store_passage(passage, fact_confidence=0.0)` ¶

Store a passage if its linked fact has confidence ≥ threshold.

Parameters:

Name	Type	Description	Default
`passage`	`SourcePassage`	Passage to store.	required
`fact_confidence`	`float`	Confidence score of the fact linked to the passage.	`0.0`

Returns:

Type	Description
`bool`	True if the passage was stored, False if it was below the threshold.

`get_passages_for_fact(fact_id)` ¶

Retrieve all source passages linked to a fact.

Parameters:

Name	Type	Description	Default
`fact_id`	`str`	Identifier of the fact.	required

Returns:

Type	Description
`list[SourcePassage]`	List of passages linked to the fact that remain in storage.

`build_source_grounded_envelope(scored_facts, budget_tokens, quality_tier='B')` ¶

Build an envelope with source passages allocated by tier.

Parameters:

Name	Type	Description	Default
`scored_facts`	`list[dict[str, Any]]`	Sorted list of fact dictionaries such as `{"id": ..., "text": ..., "score": ...}`.	required
`budget_tokens`	`int`	Total envelope budget in tokens.	required
`quality_tier`	`str`	Quality tier (S/A/B/C/D) controlling the fact/source budget split.	`'B'`

Returns:

Type	Description
`tuple[list[dict[str, Any]], list[SourcePassage]]`	A tuple of `(packed_facts, included_passages)`.

`format_envelope_section(fact, passages)` ¶

Format a fact with its source passages for envelope inclusion.

Format::

- {fact text} - Window N
  ↳ [SOURCE: Window N, tokens X-Y]
    "{verbatim original text}"

Parameters:

Name	Type	Description	Default
`fact`	`dict[str, Any]`	Fact dictionary containing at least `text` and `window`.	required
`passages`	`list[SourcePassage]`	Passages linked to the fact.	required

Returns:

Type	Description
`str`	Formatted envelope section string.

`to_dict()` ¶

Serialize the engine for persistence.

Returns:

Type	Description
`dict[str, Any]`	Dictionary with all passages and fact-to-passage mappings.

`from_dict(data, count_tokens=None)` `classmethod` ¶

Restore the engine from serialized state.

Parameters:

Name	Type	Description	Default
`data`	`dict[str, Any]`	Serialized state produced by :meth:`to_dict`.	required
`count_tokens`	`Callable[[str], int] \| None`	Optional token-counting callable; defaults to the engine's fallback estimator if None.	`None`

Returns:

Type	Description
`SourceGroundingEngine`	Reconstructed `SourceGroundingEngine` instance.

crp.advanced¶

advanced¶

crp.advanced ¶

IngestFact dataclass ¶

IngestResult dataclass ¶

ContextHungerSignal dataclass ¶

CQSDetector ¶

detect_context_hunger(generation_text, window_id='', tokens_generated=None) ¶

respond_to_context_hunger(signals, tokens_generated=0) ¶

CQSResponse dataclass ¶

ConsistencyIssue dataclass ¶

CrossWindowValidator ¶

extraction_based_validation(facts, window_outputs=None, planned_sections=None) ¶

targeted_llm_validation(tier1_issues) ¶

full_llm_review(accumulated_output, task_intent, top_facts=None) ¶

assess_review_capability() ¶

apply_corrections(issues, task_intent='', blockers=None) ¶

should_run_tier(window_index, tier) ¶

ValidationResult dataclass ¶

has_issues property ¶

high_severity_count property ¶

CurationConfig dataclass ¶

LLMContextCurator ¶

current_synthesis property ¶

evolution_count property ¶

should_curate(window_index) ¶

curate(window_index, top_facts, recent_output_summary='') ¶

format_for_envelope() ¶

to_dict() ¶

LLMSynthesis dataclass ¶

to_dict() ¶

from_dict(data) classmethod ¶

FeedbackEntry dataclass ¶

FeedbackLoop ¶

entry_count property ¶

override_fact(fact_id, corrected_text, reason='') ¶

boost_confidence(fact_id, delta=0.1, reason='') ¶

penalize_confidence(fact_id, delta=-0.2, reason='') ¶

reject_fact(fact_id, reason='') ¶

get_adjusted_confidence(fact_id, base_confidence) ¶

get_entries_for_fact(fact_id) ¶

to_dict() ¶

HierarchicalPlan dataclass ¶

HierarchicalProcessor ¶

plan(total_tokens, config=None) ¶

map_phase(segments, task_intent) ¶

reduce_phase(syntheses, task_intent, fan_in=DEFAULT_FAN_IN) ¶

hierarchical_dispatch(task_intent, large_input, config=None) ¶

MetaLearningEngine ¶

trace_count property ¶

should_use_orc(task_complexity=3, resource_pressure='NONE', probe_quality=0.0) ¶

orchestrated_reasoning(task_intent, task_complexity=3, resource_pressure='NONE') ¶

build_reasoning_scaffold(task_intent) ¶

build_metacognitive_envelope(task_intent, base_envelope='', few_shot_traces=None) ¶

store_trace(trace) ¶

to_dict() ¶

ORCResult dataclass ¶

ReasoningTrace dataclass ¶

to_dict() ¶

from_dict(data) classmethod ¶

FanOutResult dataclass ¶

FanOutTask dataclass ¶

ParallelFanOut ¶

fan_out(tasks) ¶

merge_results(results, existing_facts=None) ¶

AssessmentResult dataclass ¶

ReviewCycleManager ¶

pre_generation_plan(task_intent, predicted_chain_length=0) ¶

checkpoint_review(window_index, review_interval=20, task_intent='', top_facts=None, gap_summary='') ¶

post_generation_assessment(accumulated_output, task_intent) ¶

targeted_regeneration(issues, task_intent) ¶

ReviewGuidance dataclass ¶

QualityTier ¶

ScaleModeSelector ¶

configure_session(estimated_tokens, model_capability=1) ¶

SessionConfig dataclass ¶

SourceGroundingEngine ¶

passage_count property ¶

store_passage(passage, fact_confidence=0.0) ¶

get_passages_for_fact(fact_id) ¶

`advanced`¶

`crp.advanced` ¶

`IngestFact` `dataclass` ¶

`IngestResult` `dataclass` ¶

`ContextHungerSignal` `dataclass` ¶

`CQSDetector` ¶

`detect_context_hunger(generation_text, window_id='', tokens_generated=None)` ¶

`respond_to_context_hunger(signals, tokens_generated=0)` ¶

`CQSResponse` `dataclass` ¶

`ConsistencyIssue` `dataclass` ¶

`CrossWindowValidator` ¶

`extraction_based_validation(facts, window_outputs=None, planned_sections=None)` ¶

`targeted_llm_validation(tier1_issues)` ¶

`full_llm_review(accumulated_output, task_intent, top_facts=None)` ¶

`assess_review_capability()` ¶

`apply_corrections(issues, task_intent='', blockers=None)` ¶

`should_run_tier(window_index, tier)` ¶

`ValidationResult` `dataclass` ¶

`has_issues` `property` ¶

`high_severity_count` `property` ¶

`CurationConfig` `dataclass` ¶

`LLMContextCurator` ¶

`current_synthesis` `property` ¶

`evolution_count` `property` ¶

`should_curate(window_index)` ¶

`curate(window_index, top_facts, recent_output_summary='')` ¶

`format_for_envelope()` ¶

`to_dict()` ¶

`LLMSynthesis` `dataclass` ¶

`to_dict()` ¶

`from_dict(data)` `classmethod` ¶

`FeedbackEntry` `dataclass` ¶

`FeedbackLoop` ¶

`entry_count` `property` ¶

`override_fact(fact_id, corrected_text, reason='')` ¶

`boost_confidence(fact_id, delta=0.1, reason='')` ¶

`penalize_confidence(fact_id, delta=-0.2, reason='')` ¶

`reject_fact(fact_id, reason='')` ¶

`get_adjusted_confidence(fact_id, base_confidence)` ¶

`get_entries_for_fact(fact_id)` ¶

`to_dict()` ¶

`HierarchicalPlan` `dataclass` ¶

`HierarchicalProcessor` ¶

`plan(total_tokens, config=None)` ¶

`map_phase(segments, task_intent)` ¶

`reduce_phase(syntheses, task_intent, fan_in=DEFAULT_FAN_IN)` ¶

`hierarchical_dispatch(task_intent, large_input, config=None)` ¶

`MetaLearningEngine` ¶

`trace_count` `property` ¶

`should_use_orc(task_complexity=3, resource_pressure='NONE', probe_quality=0.0)` ¶

`orchestrated_reasoning(task_intent, task_complexity=3, resource_pressure='NONE')` ¶

`build_reasoning_scaffold(task_intent)` ¶

`build_metacognitive_envelope(task_intent, base_envelope='', few_shot_traces=None)` ¶

`store_trace(trace)` ¶

`to_dict()` ¶

`ORCResult` `dataclass` ¶

`ReasoningTrace` `dataclass` ¶

`to_dict()` ¶

`from_dict(data)` `classmethod` ¶

`FanOutResult` `dataclass` ¶

`FanOutTask` `dataclass` ¶

`ParallelFanOut` ¶

`fan_out(tasks)` ¶

`merge_results(results, existing_facts=None)` ¶

`AssessmentResult` `dataclass` ¶

`ReviewCycleManager` ¶

`pre_generation_plan(task_intent, predicted_chain_length=0)` ¶

`checkpoint_review(window_index, review_interval=20, task_intent='', top_facts=None, gap_summary='')` ¶

`post_generation_assessment(accumulated_output, task_intent)` ¶

`targeted_regeneration(issues, task_intent)` ¶

`ReviewGuidance` `dataclass` ¶

`QualityTier` ¶

`ScaleModeSelector` ¶

`configure_session(estimated_tokens, model_capability=1)` ¶

`SessionConfig` `dataclass` ¶

`SourceGroundingEngine` ¶

`passage_count` `property` ¶

`store_passage(passage, fact_confidence=0.0)` ¶

`get_passages_for_fact(fact_id)` ¶

`build_source_grounded_envelope(scored_facts, budget_tokens, quality_tier='B')` ¶