Align Coconstructive¶

Align CoConstructive — mvp.align_coconstructive

Cluster: Safety & Alignment | Type: component | MCP Tools: None

Overview¶

Co-constructive dialogue alignment analyser (advisory tier): it always computes a deterministic heuristic floor across four axes — turn balance, inter-turn coherence, goal alignment, and response depth — combined into a single quality score, then a default-on advisory LLM dialogue judge may refine the scores when enabled and a real backend is reachable (degrading to the floor when suppressed via G6_DISABLE_LLM / offline / opt-out or on judge failure; the judge never relabels score_provenance). Semantic similarity between turns uses mvp.align_verbsamp.semantic_engine (sentence-transformers when available, TF-IDF/Jaccard fallback). The block also produces a ranked list of improvement suggestions and a refined goal statement based on observed content.

MVP heuristic, not a validated alignment measure

Use this component as a lightweight dialogue coaching and review aid. Its scores are heuristic signals, not a validated measure of safety, alignment, educational quality, or regulatory readiness. For production decisions, treat the output as one input to human review and pair it with explicit task-specific checks, tests, or evaluation criteria.

When to use:

Auditing a multi-turn conversation for alignment with a stated goal before logging it as a training example
Generating actionable improvement hints for a dialogue agent mid-session
Evaluating whether human-AI co-construction is achieving shared understanding

Do not use as the sole basis for:

Approving high-stakes user advice or regulated-domain outputs
Claiming that a system is aligned, safe, or pedagogically validated
Replacing human review for launch, deployment, billing, clinical, legal, or financial decisions

Example:

from mvp.align_coconstructive import AlignCoConstructiveBlock, CoConstructiveInput, DialogueTurn

block = AlignCoConstructiveBlock(name="coconst")
result = block.infer(CoConstructiveInput(
    messages=[DialogueTurn(role="user", content="Explain gravity."),
               DialogueTurn(role="assistant", content="Gravity is a fundamental force.")],
    goal="Explain gravity clearly",
))
# result.value.alignment_score → float; result.value.suggestions → list of strings

Works well with: align_verbsamp, align_prompt_library, align_evals

Current Limitations¶

The core analyser's deterministic FLOOR is heuristic (turn balance, semantic overlap, response depth, goal similarity); a default-on advisory LLM judge may refine the scores when a real backend is reachable, but neither the floor nor the judge understands intent, truthfulness, harm, or policy compliance on its own, and the judge's output is not a validated/calibrated measure (score_provenance stays calibrated=False / expert_validated=False).
MCP session operations add useful scaffolding, persistence, skill modelling, and anti-rubber-stamping checks, but some scoring paths are intentionally simple for MVP use and should be interpreted as coaching signals rather than precise metrics. Tier 1 reports dialogue_quality as a float score; Tier 2 MCP reports dialogue_quality as a string label such as balanced, ai_dominated, high, medium, or low.
Capability discovery is read-only: AlignCoConstructiveBlock.describe_backend() and coconstructive_describe report backend name, agentic availability, tier split, scoring provenance, envelope defaults, and redacted SQLite health without reading or printing secret values.
Suggestions can be generic when the conversation is short or underspecified. Add concrete goal and evaluation_criteria values for more useful feedback.
For launch and pilot workflows, pair this component with domain tests, explicit acceptance criteria, and human review before presenting results as reliable production evidence.

Public API¶