feat(frameworks): add dimension combinations, best practices, and compliance evaluation

Each framework definition now supports three optional standard blocks:

- `dimension_combinations` — named synergies, prerequisites, and tensions between
  dimensions, each with guidance an agent can inject to steer collaboration
- `best_practices` — actionable per-dimension recommendations with antipatterns
  and observable signals for detecting when a practice is already followed
- `evaluation_criteria` — machine-checkable keyword-signal criteria with weights
  and pass thresholds, enabling scored compliance assessment

Engine:
- Add `evaluateCompliance(text, framework)` — scans collaboration text against
  a framework's evaluation_criteria and returns a 0-100 weighted score,
  per-criterion pass/fail, and matched keywords

MCP:
- Add `evaluate_compliance` tool — agents can call this mid-session or at end
  to check framework adherence and identify which dimensions need attention
- Update `get_framework_detail` description to surface the new blocks

4D Framework YAML (v1.0.0 -> v1.1.0):
- 4 dimension_combinations (del+des synergy, des+dis prerequisite,
  dis+dil synergy, del+dil tension)
- 8 best_practices across all four dimensions plus a cycle-close practice
- 4 evaluation_criteria (one per dimension, weights sum to 1.0)

Tests: 10 new test cases covering schema validation, combination/criterion
constraints, and all evaluateCompliance edge cases.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: faical-yannick-congo

frameworks/4d-framework.yaml

@@ -1,6 +1,6 @@
 id: 4d-framework
 name: AI Fluency 4D Framework
-version: 1.0.0
+version: 1.1.0
 contributor: Dakan & Feller
 description: >
   Four dimensions of good human-AI collaboration. Originally described
@@ -29,3 +29,264 @@ tags:
   - ai-fluency
   - 4d
 reference: https://fluently.ctrl6.com
+
+# ── How dimensions interact ────────────────────────────────────────────────────
+# Synergies, prerequisites, and tensions between dimensions.
+# Agents can use these to diagnose imbalances and steer the collaboration
+# back on track when a dimension is weak or misaligned.
+
+dimension_combinations:
+  - id: del-des-synergy
+    dimensions: [delegation, description]
+    type: synergy
+    label: Autonomy drives context depth
+    description: >
+      The more autonomy delegated to AI, the richer the description must be to
+      constrain the solution space. Low-delegation tasks tolerate loose framing;
+      high-delegation tasks require explicit constraints and success criteria to
+      prevent the AI from filling gaps with unsafe assumptions.
+    guidance: >
+      When delegation is set to "automated", verify that the description covers
+      scope boundaries, constraints, edge cases, and success criteria before
+      handing off. If the description is thin, reduce delegation to "augmented"
+      until the context is complete.
+
+  - id: des-dis-prerequisite
+    dimensions: [description, discernment]
+    type: prerequisite
+    label: Context enables evaluation
+    description: >
+      A well-formed description makes discernment possible: output can only be
+      evaluated against criteria that were stated upfront. Without a clear
+      description, discernment becomes arbitrary and inconsistent.
+    guidance: >
+      Embed explicit acceptance criteria or evaluation questions directly in the
+      description. If you cannot articulate what good output looks like before
+      the AI responds, the description is not complete enough to proceed.
+
+  - id: dis-dil-synergy
+    dimensions: [discernment, diligence]
+    type: synergy
+    label: Evaluation quality anchors accountability
+    description: >
+      Discernment feeds diligence directly: shallow evaluation criteria make
+      accountability meaningless because sign-off happens without genuine review,
+      creating the appearance of oversight without its substance.
+    guidance: >
+      Define your discernment checklist before committing to a diligence workflow.
+      Accountability assignments without evaluation criteria are ceremonial, not real.
+
+  - id: del-dil-tension
+    dimensions: [delegation, diligence]
+    type: tension
+    label: Autonomy vs. accountability
+    description: >
+      High AI autonomy reduces the human review surface, which naturally erodes
+      diligence unless explicit checkpoints are maintained. The more the AI acts
+      independently, the more deliberate the accountability structure must be.
+    guidance: >
+      When delegation is "automated", introduce mandatory diligence gates such as
+      output spot-checks, audit logs, or periodic human-in-the-loop reviews.
+      A purely automated pipeline with no diligence gate is a liability, not
+      an efficiency gain.
+
+# ── Actionable recommendations ────────────────────────────────────────────────
+# Best practices an agent can surface to steer a collaboration toward better
+# outcomes. Each includes a signal so the agent can detect whether the
+# practice is already being followed before injecting guidance.
+
+best_practices:
+  - id: bp-del-first
+    title: Establish autonomy level before framing the task
+    dimension: delegation
+    description: >
+      Decide whether the AI acts autonomously, augments human work, or operates
+      under close supervision before writing the task prompt. This choice shapes
+      every downstream decision about context depth, review rigor, and accountability.
+    antipattern: >
+      Jumping directly to task framing without agreeing on who makes the final
+      decision, leading to implicit and often misaligned autonomy expectations
+      that only surface when the output is already wrong.
+    signal: Opening prompt contains an explicit autonomy level or role assignment for the AI.
+
+  - id: bp-del-risk-match
+    title: Match delegation level to task risk
+    dimension: delegation
+    description: >
+      Reserve "automated" delegation for tasks where mistakes are cheap to catch
+      and correct. Use "augmented" when errors have downstream consequences, and
+      "supervised" when the domain is high-stakes or the AI's reliability is unknown.
+    antipattern: >
+      Treating all tasks as equally safe to automate because a previous task went
+      well. Risk calibration must be per-task, not per-tool.
+    signal: Delegation level is justified by a stated risk assessment or consequence estimate.
+
+  - id: bp-des-constraints
+    title: Front-load constraints and success criteria
+    dimension: description
+    description: >
+      State non-negotiable constraints, scope limits, and what "done" looks like
+      at the start of the description rather than correcting the AI after the fact.
+      Constraints discovered late cost more to fix than constraints stated early.
+    antipattern: >
+      Iterating through multiple AI responses to discover implicit constraints
+      that should have been stated upfront, wasting cycles and eroding confidence
+      in the AI's reliability.
+    signal: Task description includes at least one explicit constraint or acceptance criterion.
+
+  - id: bp-des-context
+    title: Include why, not just what
+    dimension: description
+    description: >
+      Provide the motivation and broader context behind the task, not just the
+      immediate ask. AI models use context to make better judgment calls at the
+      edges; without it they default to generic behavior.
+    antipattern: >
+      Writing task descriptions as bare commands without context, leading to
+      technically correct but practically useless outputs that ignore real intent.
+    signal: Description includes a stated purpose, audience, or motivating context.
+
+  - id: bp-dis-criteria
+    title: Define evaluation criteria before reviewing output
+    dimension: discernment
+    description: >
+      Agree on what good output looks like — and what would disqualify it —
+      before asking the AI to generate it. This prevents post-hoc rationalisation
+      of mediocre outputs and keeps evaluation consistent across reviewers.
+    antipattern: >
+      Adjusting the criteria to match what the AI produced rather than what was
+      actually needed, or accepting the first plausible response without any check.
+    signal: Collaboration sequence includes an explicit review or validation step with stated criteria.
+
+  - id: bp-dis-pushback
+    title: Define when to push back or iterate
+    dimension: discernment
+    description: >
+      Agree upfront on the conditions under which you will reject an AI response
+      and re-prompt rather than accept a partial result. Without a pushback policy,
+      reviewers default to accepting whatever the AI produces.
+    antipattern: >
+      Treating AI output as a starting point for human editing rather than as a
+      proposal to be evaluated, blurring the line between AI work and human work.
+    signal: Collaboration includes a defined condition for rejection or re-prompting.
+
+  - id: bp-dil-sign-off
+    title: Assign a named human accountable for the final output
+    dimension: diligence
+    description: >
+      Every AI-assisted task should end with a named human who takes explicit
+      responsibility for the output before it is used, shared, or deployed.
+    antipattern: >
+      Deploying or sharing AI output without a clear sign-off process, making
+      accountability diffuse across the team and errors harder to trace.
+    signal: Collaboration includes an explicit approval step with a named role or person.
+
+  - id: bp-cycle-close
+    title: Close every cycle with a brief reflection
+    description: >
+      After completing a collaboration cycle, briefly review whether all four
+      dimensions were genuinely addressed. Imbalances indicate where the next
+      cycle should invest effort.
+    antipattern: >
+      Treating each AI session as independent without reviewing overall collaboration
+      quality, leading to systematic blind spots that compound across projects.
+    signal: Cycle ends with a retrospective, quality review, or explicit dimension balance check.
+
+# ── Machine-checkable compliance criteria ─────────────────────────────────────
+# These criteria let the scorer and agents evaluate a collaboration text for
+# framework adherence using keyword signals.
+# Weights sum to 1.0. pass_threshold = minimum matching signals required.
+
+evaluation_criteria:
+  - id: eval-del-explicit
+    dimension: delegation
+    label: Autonomy level is explicit
+    description: >
+      The collaboration includes a clear statement of the AI's autonomy level
+      (automated, augmented, or supervised) or an equivalent role assignment.
+    weight: 0.25
+    signals:
+      present:
+        - automated
+        - augmented
+        - supervised
+        - decides
+        - approves
+        - oversight
+        - autonomous
+        - human-in-the-loop
+        - review before
+        - sign off before
+    pass_threshold: 1
+
+  - id: eval-des-rich
+    dimension: description
+    label: Task description includes constraints or success criteria
+    description: >
+      The task framing contains at least two explicit constraints, scope
+      boundaries, goals, or acceptance conditions — not just a bare command.
+    weight: 0.25
+    signals:
+      present:
+        - constraint
+        - requirement
+        - must not
+        - must be
+        - should not
+        - should be
+        - criteria
+        - success
+        - goal
+        - objective
+        - scope
+        - out of scope
+        - acceptable
+        - definition of done
+        - expected output
+    pass_threshold: 2
+
+  - id: eval-dis-review
+    dimension: discernment
+    label: Output review or validation step is present
+    description: >
+      The collaboration includes at least one step where AI output is evaluated
+      against defined criteria before being accepted or acted upon.
+    weight: 0.25
+    signals:
+      present:
+        - review
+        - verify
+        - validate
+        - check
+        - evaluate
+        - confirm
+        - test
+        - assess
+        - cross-check
+        - compare
+        - quality
+        - accuracy
+    pass_threshold: 1
+
+  - id: eval-dil-accountability
+    dimension: diligence
+    label: Accountability is assigned
+    description: >
+      The collaboration includes an explicit sign-off, ownership assignment, or
+      accountability statement for the final output.
+    weight: 0.25
+    signals:
+      present:
+        - approve
+        - sign off
+        - accountable
+        - responsible
+        - owner
+        - lead
+        - final review
+        - authorized
+        - reviewed by
+        - approved by
+        - sign-off
+        - takes ownership
+    pass_threshold: 1