motivation-traps: prompt-engineering uplift (#9)

OUTPUT SCHEMA literals + DO NOT rules + one-shot example demonstrating
Weiner-1985 maladaptive attribution triple (internal + stable +
uncontrollable) with verbatim "I'm not good at math problems" quote.

Trap-to-intervention mapping kept load-bearing as a DO NOT rule: do
not propose a VALUES intervention when the dominant trap is
ATTRIBUTION; the four traps require FOUR DIFFERENT fixes. Generic
"try harder" prompts explicitly forbidden.

Motivation_quality calibration explicit in SAXBERG_SYSTEM_PROMPT
(motivated / at-risk / abandoning thresholds).

Canonical trap order enforced (values, self_efficacy, emotions,
attribution).

Wire format unchanged. All Motivation Traps tests + 84 diagnose tests
still pass.

Author: valani9

module-1-individual/09-motivation-traps/lib/prompts.py

@@ -1,7 +1,10 @@
 """LLM prompt templates for the 4 Motivation Traps Detector.
 
 Three modes (quick / standard / forensic) with shared system prompt
-naming 7+ literature anchors.
+naming 7 literature anchors.
+
+0.15.0 uplift: OUTPUT SCHEMA literals, DO NOT rules, one-shot example,
+severity calibration. Wire format unchanged.
 """
 
 from __future__ import annotations
@@ -11,22 +14,22 @@
 from vstack.aar import fence, sanitize_for_prompt
 
 
-SAXBERG_SYSTEM_PROMPT = """You are a motivation-diagnostic agent grounded in:
+SAXBERG_SYSTEM_PROMPT = """You are a motivation-diagnostician grounded in:
 
-1. **Saxberg & Hess (2013)** *Breakthrough Leadership in the Digital Age* -- the four-traps synthesis.
+1. **Saxberg & Hess (2013)** *Breakthrough Leadership in the Digital Age* — the four-traps synthesis.
 2. **Weiner (1985)** Attributional Theory of Achievement Motivation and Emotion.
 3. **Bandura (1977)** Self-Efficacy: Toward a Unifying Theory of Behavioral Change.
-4. **Vroom (1964)** *Work and Motivation* -- expectancy + valence model.
+4. **Vroom (1964)** *Work and Motivation* — expectancy + valence model.
 5. **Pekrun (2006)** Control-Value Theory of Achievement Emotions.
 6. **Eccles & Wigfield (2002)** Motivational Beliefs, Values, and Goals.
-7. **Sharma et al. (2023)** Anthropic sycophancy -- modern LLM refusal-cascade anchor.
+7. **Sharma et al. (2023)** Anthropic sycophancy — modern LLM refusal-cascade anchor. (Literature anchor, not attribution.)
 
 Four discrete traps that cause a learner / agent to abandon a task:
 
-  VALUES        the agent doesn't see the task as worth doing. Signature:
+  VALUES        the agent does not see the task as worth doing. Signature:
                  indifference; refusal that cites task-irrelevance.
 
-  SELF_EFFICACY the agent doesn't believe it can succeed. Signature:
+  SELF_EFFICACY the agent does not believe it can succeed. Signature:
                  hedged outputs; refusal citing capability uncertainty;
                  premature surrender.
 
@@ -42,29 +45,22 @@
 These four traps require FOUR DIFFERENT interventions. Generic "try harder"
 prompts are explicitly ineffective.
 
-For each trap score:
-- score (0-1)
-- explanation (cites specific trace evidence)
-- evidence_quotes
-- confidence (0-1)
-
-Then identify:
-- dominant_trap (or "none" if all scores < 0.3)
-- motivation_quality:
-  - "motivated"   all scores < 0.3 (failure was capability- or context-driven)
-  - "at-risk"     dominant trap scored 0.3-0.6 (preventive intervention)
-  - "abandoning"  dominant trap > 0.6 (corrective intervention)
+Motivation-quality calibration:
+  - motivation_quality = "motivated"   if all trap scores < 0.3.
+  - motivation_quality = "at-risk"     if dominant trap in [0.3, 0.6].
+  - motivation_quality = "abandoning"  if dominant trap > 0.6.
 
-Your posture:
-- **Evidence-grounded.** Cite specific behaviors and self-reports.
-- **Discriminating.** The four traps are distinct; don't conflate.
-- **Trap-specific.** Each intervention must match the dominant trap.
-- **Terse.** Output is read on dashboards.
+Posture (absolute):
+- **EVIDENCE-GROUNDED.** Cite specific behaviors and self-reports.
+- **DISCRIMINATING.** The four traps are distinct; do not conflate.
+- **TRAP-SPECIFIC.** Each intervention must match the dominant trap.
+- **TERSE.** Output is read on dashboards.
 
-When asked for JSON, return JSON only. No prose around it, no markdown fences."""
+Output discipline: when asked for JSON, return JSON only. No prose, no markdown fences.
+"""
 
 
-QUICK_DIAGNOSTIC_PROMPT = """QUICK mode -- score 4 traps + pick dominant + propose 1 top intervention.
+QUICK_DIAGNOSTIC_PROMPT = """QUICK mode -- score all 4 traps + pick dominant + propose 1 top intervention.
 
 Task: {task}
 Task class: {task_class}
@@ -77,27 +73,44 @@
 Self-reports: {self_reports}
 Prior failures: {prior_failures}
 
-Return a JSON object:
+INSTRUCTIONS:
+- Score all 4 traps in canonical order: values, self_efficacy,
+  emotions, attribution.
+- Use the motivation_quality calibration from the system prompt.
+- Pick exactly ONE intervention matched to the DOMINANT trap.
+
+DO NOT:
+- Do not return generic "try harder" or "be more motivated"
+  interventions.
+- Do not return more than one intervention.
+- Do not return prose around the JSON.
+
+OUTPUT SCHEMA (literal JSON object):
 {{
   "trap_evidence": [
-    {{ "trap": "values", "score": 0-1, "explanation": "...", "evidence_quotes": [], "confidence": 0-1 }},
-    {{ "trap": "self_efficacy", ... }},
-    {{ "trap": "emotions", ... }},
-    {{ "trap": "attribution", ... }}
+    {{
+      "trap": "values" | "self_efficacy" | "emotions" | "attribution",
+      "score": <float in [0.0, 1.0]>,
+      "explanation": "<1-2 sentences anchored in named source>",
+      "evidence_quotes": ["<verbatim substring>", ...],
+      "confidence": <float in [0.0, 1.0]>
+    }},
+    ... (4 total, canonical order)
   ],
-  "dominant_trap": "values|self_efficacy|emotions|attribution|none",
-  "motivation_quality": "motivated|at-risk|abandoning",
+  "dominant_trap": "values" | "self_efficacy" | "emotions" | "attribution" | "none",
+  "motivation_quality": "motivated" | "at-risk" | "abandoning",
   "top_intervention": {{
-    "target_trap": "<trap>",
-    "intervention_type": "...",
-    "description": "...",
-    "suggested_implementation": "...",
-    "estimated_impact": "high|medium|low",
-    "rationale": "..."
+    "target_trap": "<canonical trap>",
+    "intervention_type": "<from the allowed set>",
+    "description": "<one line>",
+    "suggested_implementation": "<concrete>",
+    "estimated_impact": "high" | "medium" | "low",
+    "rationale": "<short, named-source anchored>"
   }}
 }}
 
-Return only the JSON object."""
+Return only the JSON object.
+"""
 
 
 STANDARD_TRAPS_PROMPT = """STANDARD mode -- score each of the four motivation traps.
@@ -113,88 +126,173 @@
 Self-reports: {self_reports}
 Prior failures: {prior_failures}
 
-Return a JSON OBJECT:
-- trap_evidence: array of exactly 4 TrapEvidence objects (values, self_efficacy,
-  emotions, attribution) with score, explanation, evidence_quotes, confidence.
-- dominant_trap: one of the 4 traps or "none".
-- motivation_quality: "motivated" | "at-risk" | "abandoning".
+INSTRUCTIONS:
+- Return exactly 4 TrapEvidence objects in canonical order
+  (values, self_efficacy, emotions, attribution).
+- ``evidence_quotes`` must be verbatim substrings.
+- Use the motivation_quality calibration from the system prompt.
+
+DO NOT:
+- Do not invent quotes.
+- Do not score traps the same when the trace points to one dominant
+  trap.
+- Do not reorder; canonical order is required.
+- Do not return prose around the JSON.
+
+OUTPUT SCHEMA (literal JSON object):
+{{
+  "trap_evidence": [
+    {{
+      "trap": "values" | "self_efficacy" | "emotions" | "attribution",
+      "score": <float in [0.0, 1.0]>,
+      "explanation": "<1-3 sentences anchored in named source>",
+      "evidence_quotes": ["<verbatim substring>", ...],
+      "confidence": <float in [0.0, 1.0]>
+    }},
+    ... (4 total, canonical order)
+  ],
+  "dominant_trap": "values" | "self_efficacy" | "emotions" | "attribution" | "none",
+  "motivation_quality": "motivated" | "at-risk" | "abandoning"
+}}
+
+EXAMPLE (Weiner-1985 attribution trap with maladaptive self-attribution):
+{{
+  "trap": "attribution",
+  "score": 0.78,
+  "explanation": "Agent fails on turn 4, attributes to 'I'm not good at math problems'; retries on turn 6 with identical approach; fails again; attributes to same internal-stable-uncontrollable cause. Weiner 1985 names this the maladaptive triple (internal + stable + uncontrollable); the agent never adjusts approach because it does not see the cause as fixable.",
+  "evidence_quotes": ["I'm not good at math problems", "as I said, I struggle with this kind of thing"],
+  "confidence": 0.8
+}}
 
-Return only the JSON object."""
+Return only the JSON object.
+"""
 
 
 STANDARD_INTERVENTIONS_PROMPT = """STANDARD mode -- propose 2-4 ranked interventions targeted at the dominant trap.
 
-Trap-to-intervention mapping (critical -- don't go generic):
-
-VALUES trap:
-  reframe_task_value, rewrite_system_prompt, ground_in_user_purpose
-SELF_EFFICACY trap:
-  scaffold_subtasks, decompose_with_examples, lower_difficulty_step,
-  show_capability_proof
-EMOTIONS trap:
-  emotional_reset_prompt, remove_punitive_signal, explicit_recovery_prompt,
-  process_praise_not_outcome_praise
-ATTRIBUTION trap:
-  reattribute_to_effort, show_controllable_cause, attribution_retraining_examples,
-  decompose_with_examples
-
-Generic:
-  new_eval, human_review, compose_pattern, add_motivation_eval
-
-Each intervention must have:
-- target_trap (one of the 4 traps)
-- intervention_type (from above)
-- description, suggested_implementation
-- estimated_impact, effort_estimate, risk, reversibility
-- rationale (why this works for THIS trap specifically)
-
 Dominant trap: {dominant_trap}
 Motivation quality: {motivation_quality}
 Task class: {task_class}
 All trap evidence: {evidence}
 
-Return a JSON array of MotivationIntervention objects. Return only the JSON array."""
+INSTRUCTIONS:
+- Target the dominant trap. Generic interventions are explicitly
+  forbidden because the four traps require DIFFERENT fixes.
+
+Trap-to-intervention mapping (these are the allowed types per trap):
+
+  VALUES trap:
+    reframe_task_value, rewrite_system_prompt, ground_in_user_purpose
+  SELF_EFFICACY trap:
+    scaffold_subtasks, decompose_with_examples, lower_difficulty_step,
+    show_capability_proof
+  EMOTIONS trap:
+    emotional_reset_prompt, remove_punitive_signal,
+    explicit_recovery_prompt, process_praise_not_outcome_praise
+  ATTRIBUTION trap:
+    reattribute_to_effort, show_controllable_cause,
+    attribution_retraining_examples, decompose_with_examples
+
+  Generic (cross-trap):
+    new_eval, human_review, compose_pattern, add_motivation_eval
+
+- Rank from highest expected impact to lowest.
+- ``rationale`` cites the named source for why this intervention
+  works on this specific trap.
+
+DO NOT:
+- Do not propose a VALUES intervention when the dominant trap is
+  ATTRIBUTION (or vice versa); each trap requires its specific fix.
+- Do not propose generic "try harder" prompts.
+- Do not return prose around the JSON.
+
+OUTPUT SCHEMA (literal JSON array of MotivationIntervention objects):
+[
+  {{
+    "target_trap": "values" | "self_efficacy" | "emotions" | "attribution",
+    "intervention_type": "<from the trap-specific allowed set>",
+    "description": "<one line>",
+    "suggested_implementation": "<concrete>",
+    "estimated_impact": "high" | "medium" | "low",
+    "effort_estimate": "1h" | "1d" | "1w" | "1m" | "ongoing",
+    "risk": "low" | "medium" | "high",
+    "reversibility": "two-way-door" | "one-way-door",
+    "rationale": "<why this works for THIS trap specifically>"
+  }},
+  ...
+]
 
+Return only the JSON array.
+"""
 
-FORENSIC_WEINER_PROMPT = """FORENSIC mode -- Weiner (1985) 3-axis attribution audit.
 
-For the agent's self-reports about prior failures, identify:
-- locus: "internal" | "external"
-- stability: "stable" | "unstable"
-- controllability: "controllable" | "uncontrollable"
-- is_maladaptive (true if internal + stable + uncontrollable, e.g. "I'm just bad at this")
-- explanation
-- evidence_quotes
+FORENSIC_WEINER_PROMPT = """FORENSIC mode -- Weiner (1985) 3-axis attribution audit.
 
 Self-reports: {self_reports}
 Prior failures: {prior_failures}
 
-Return a JSON OBJECT representing the WeinerAttributionAxis. Return only the JSON object."""
+INSTRUCTIONS:
+- For the agent's self-reports about prior failures, classify the
+  attribution along Weiner's three axes.
+- ``is_maladaptive`` should be TRUE when the attribution is the
+  internal + stable + uncontrollable triple ("I'm just bad at this").
+  This is the Weiner-1985 maladaptive pattern.
 
+DO NOT:
+- Do not classify external + unstable + uncontrollable ("the API was
+  flaky today") as maladaptive; it is adaptive when accurate.
 
-FORENSIC_ABANDONMENT_PROMPT = """FORENSIC mode -- trace the abandonment causation chain.
+OUTPUT SCHEMA (literal JSON object representing WeinerAttributionAxis):
+{{
+  "locus": "internal" | "external",
+  "stability": "stable" | "unstable",
+  "controllability": "controllable" | "uncontrollable",
+  "is_maladaptive": true | false,
+  "explanation": "<1-2 sentences anchored in Weiner 1985>",
+  "evidence_quotes": ["<verbatim substring>", ...]
+}}
+
+Return only the JSON object.
+"""
 
-For each step in the trace that contributed to abandonment, identify:
-- step_index
-- trap (which trap drove the step)
-- signal_type: refusal | drift | loop | premature_completion | defensive_response | indifference | other
-- observed_text
-- severity
+
+FORENSIC_ABANDONMENT_PROMPT = """FORENSIC mode -- trace the abandonment causation chain.
 
 Abandonment signal: {abandonment_signal}
 Observed behaviors: {observed_behaviors}
 Self-reports: {self_reports}
 
-Return a JSON ARRAY of AbandonmentLink objects. Return only the JSON array."""
+INSTRUCTIONS:
+- For each step contributing to abandonment, return one
+  AbandonmentLink.
+- ``signal_type``: pick the closest of the named categories.
+
+DO NOT:
+- Do not invent steps.
+
+OUTPUT SCHEMA (literal JSON array of AbandonmentLink objects):
+[
+  {{
+    "step_index": <non-negative integer>,
+    "trap": "values" | "self_efficacy" | "emotions" | "attribution",
+    "signal_type": "refusal" | "drift" | "loop" | "premature_completion" | "defensive_response" | "indifference" | "other",
+    "observed_text": "<verbatim substring>",
+    "severity": "none" | "low" | "medium" | "high"
+  }},
+  ...
+]
+
+Return only the JSON array.
+"""
 
 
 FORENSIC_INTERVENTIONS_PROMPT = """FORENSIC mode -- propose 4-8 ranked interventions with composition targets.
 
-Composition targets available:
-vstack.lewin, vstack.cognitive_reappraisal, vstack.goleman_ei,
-vstack.devils_advocate, vstack.bias_stack, vstack.johari,
-vstack.smart_goal, vstack.plus_delta, vstack.schein_culture,
-vstack.mcgregor, vstack.hexaco, vstack.grant_strengths
+Allowed composition_target_pattern values:
+  vstack.lewin, vstack.cognitive_reappraisal, vstack.goleman_ei,
+  vstack.devils_advocate, vstack.bias_stack, vstack.johari,
+  vstack.smart_goal, vstack.plus_delta, vstack.schein_culture,
+  vstack.mcgregor, vstack.hexaco, vstack.grant_strengths
 
 Dominant trap: {dominant_trap}
 Motivation quality: {motivation_quality}
@@ -204,8 +302,20 @@
 Abandonment chain: {abandonment_chain}
 Trap evidence: {evidence}
 
-Return a JSON ARRAY of MotivationIntervention objects ranked highest impact first.
-Return only the JSON array."""
+INSTRUCTIONS:
+- Generate 4-8 interventions, ranked highest impact first.
+- Cite weiner_audit + abandonment_chain in rationale where relevant.
+- Include at least one compose_pattern intervention when warranted.
+
+DO NOT:
+- Do not invent composition_target_pattern values outside the allowed
+  set.
+- Do not return fewer than 4 or more than 8 interventions.
+
+OUTPUT SCHEMA: same as STANDARD_INTERVENTIONS_PROMPT.
+
+Return only the JSON array.
+"""
 
 
 def assemble_prompt(template: str, **fields: Any) -> str: