eval::threshold_adapter::ThresholdAdapter¶

Public Functions¶

	Name
	init(self self, teacher_client teacher_client, Optional project_root[Path] =None, Optional] config[Dict[str, Any] =None)
ThresholdRecommendation	evaluate_thresholds(self self, str niche, Dict current_metrics[str, Any], Dict current_thresholds[str, Any], Optional]] prior_runs[List[Dict[str, Any] =None, Optional model_name[str] =None)
bool	enabled(self self)
float	confidence_minimum(self self)
float	max_adjustment_percent(self self)

Protected Functions¶

	Name
Dict[str, Any]	_load_config(self self, Path project_root)
str	_build_prompt(self self, str niche, Dict current_metrics[str, Any], Dict current_thresholds[str, Any], List] prior_runs[Dict[str, Any])
str	_call_llm(self self, str prompt, Optional model_name[str] =None)
Dict[str, Any]	_parse_response(self self, str content)
bool	_should_auto_apply(self self, ThresholdRecommendation recommendation, Dict current_thresholds[str, Any])
bool	_all_changes_within_cap(self self, Dict recs[str, Any], Dict current_thresholds[str, Any])
Dict[str, Any]	_compute_updated_thresholds(self self, ThresholdRecommendation recommendation, Dict current_thresholds[str, Any])

Protected Attributes¶

	Name
	_teacher
	_project_root
	_config
	_confidence_minimum
	_max_adjustment_percent
	_enabled

Detailed Description¶

class eval::threshold_adapter::ThresholdAdapter;

LLM-based adaptive threshold evaluator via TeacherClient.

Sends structured evaluation metrics to an LLM, parses the JSON
recommendation, and conditionally auto-applies threshold changes
when confidence and magnitude criteria are met.

Args:
    teacher_client: Phase 1 ``TeacherClient`` instance for LLM calls.
    project_root: Root of the gnus-poc project (for config loading).
    config: Optional dict of ``adaptive_thresholding`` settings.
        Loaded from ``config/pipeline.yaml`` if not provided.

Public Functions Documentation¶

function init¶

__init__(
    self self,
    teacher_client teacher_client,
    Optional project_root[Path] =None,
    Optional] config[Dict[str, Any] =None
)

function evaluate_thresholds¶

ThresholdRecommendation evaluate_thresholds(
    self self,
    str niche,
    Dict current_metrics[str, Any],
    Dict current_thresholds[str, Any],
    Optional]] prior_runs[List[Dict[str, Any] =None,
    Optional model_name[str] =None
)

Prompt the LLM for threshold recommendations and conditionally apply.

Args:
    niche: Specialist niche name.
    current_metrics: Dict with keys ``ppl``, ``bleu``, ``rouge_l``,
        ``latency`` matching the current evaluation run.
    current_thresholds: Dict with keys ``ppl_max``, ``bleu_min``,
        ``consecutive_failures``.
    prior_runs: Optional list of prior-run metric dicts.
    model_name: Optional teacher model override.

Returns:
    ``ThresholdRecommendation`` with the LLM's analysis and any
    auto-applied threshold updates.

function enabled¶

bool enabled(
    self self
)

function confidence_minimum¶

float confidence_minimum(
    self self
)

function max_adjustment_percent¶

float max_adjustment_percent(
    self self
)

Protected Functions Documentation¶

function _load_config¶

Dict[str, Any] _load_config(
    self self,
    Path project_root
)

Load ``adaptive_thresholding`` section from pipeline.yaml.```


### function _build_prompt

```python
str _build_prompt(
    self self,
    str niche,
    Dict current_metrics[str, Any],
    Dict current_thresholds[str, Any],
    List] prior_runs[Dict[str, Any]
)

Format the THRESHOLD_EVALUATION_PROMPT with runtime data.

Prior run metrics are embedded as a JSON code block to prevent
prompt injection (T-02-09 mitigation).

function _call_llm¶

str _call_llm(
    self self,
    str prompt,
    Optional model_name[str] =None
)

Send the prompt to TeacherClient and return the content string.```


### function _parse_response

```python
Dict[str, Any] _parse_response(
    self self,
    str content
)

Extract JSON from an LLM response string.

Attempts to find a JSON code block first, then falls back to
parsing the entire content as JSON.  Returns an empty dict on
parse failure to avoid crashing the pipeline.

function _should_auto_apply¶

bool _should_auto_apply(
    self self,
    ThresholdRecommendation recommendation,
    Dict current_thresholds[str, Any]
)

Check whether the recommendation meets auto-apply criteria.

Two conditions must BOTH be true (T-02-10):
1. ``confidence`` > ``confidence_minimum`` (default 0.9)
2. Every recommended threshold change is < ``max_adjustment_percent``
   (default 20%) of the current value.

function _all_changes_within_cap¶

bool _all_changes_within_cap(
    self self,
    Dict recs[str, Any],
    Dict current_thresholds[str, Any]
)

Return True if ALL recommended changes are within the adjustment cap.```


### function _compute_updated_thresholds

```python
Dict[str, Any] _compute_updated_thresholds(
    self self,
    ThresholdRecommendation recommendation,
    Dict current_thresholds[str, Any]
)

Build the updated thresholds dict from the recommendation.```


## Protected Attributes Documentation

### variable _teacher

```python
_teacher;

variable _project_root¶

_project_root;

variable _config¶

_config;

variable _confidence_minimum¶

_confidence_minimum;

variable _max_adjustment_percent¶

_max_adjustment_percent;

variable _enabled¶

_enabled;

Updated on 2026-10-09 at 03:59:43 +0000