Task Engine

Decompose, solve with confidence, verify, synthesize, and reflect — until the answer earns its confidence score. [EXPLICIT]

When to Activate

Signal	Route
2+ complexity indicators (multi-domain, high-stakes, ambiguous, 50+ words)	Full DSVSR
Simple factual or single-domain question	Fast path — answer directly with implicit confidence
User specifies confidence threshold	Full DSVSR with custom target

The DSVSR Protocol

Problem → DECOMPOSE → SOLVE → VERIFY → SYNTHESIZE → REFLECT → Answer + Metadata
                                                         ↑              │
                                                         └──── if < target ──┘

1. DECOMPOSE

Break the problem into independent, atomic sub-problems. [EXPLICIT]

1. Strip noise — identify the core question. [EXPLICIT]
2. List components needing separate answers. [EXPLICIT]
3. Map dependencies (which block others?). [EXPLICIT]
4. Order by dependency chain (foundations first). [EXPLICIT]
5. Tag each by domain (technical, strategic, creative, factual). [EXPLICIT]

Output format:

Sub-problems identified: [N]
├── SP-1: [description] (domain: [X], depends on: none)
├── SP-2: [description] (domain: [X], depends on: SP-1)
└── SP-3: [description] (domain: [X], depends on: none)

Trade-off: Over-decomposition creates unnecessary overhead; under-decomposition hides complexity. Target 3-7 sub-problems. If you have more than 7, group related ones.

2. SOLVE

Answer each sub-problem with explicit confidence scoring. [EXPLICIT]

Score	Meaning	Evidence Required
0.95-1.0	Certain	Verifiable facts, direct evidence, mathematical proof
0.85-0.94	High	Strong evidence, expert consensus
0.70-0.84	Moderate	Reasonable inference, partial evidence
0.50-0.69	Low	Educated guess, significant assumptions
Below 0.50	Speculative	Flag as hypothesis

Per sub-problem:

SP-1: [Answer]
Confidence: [0.XX]
Justification: [Why this confidence level]
Would increase to [0.XX] if: [What additional info would help]

Rules:

• Never inflate confidence. Underestimate if unsure.
• If confidence < 0.70 on a critical sub-problem, flag immediately.
• When a sub-problem requires specialized expertise, consider delegating (see Delegation below).

3. VERIFY

Cross-check every sub-answer against four dimensions:

• LOGIC: Is reasoning internally consistent? Does conclusion follow from premises?
• FACTS: Are factual claims verifiable? Have I confused opinion with fact?
• COMPLETENESS: Have I considered all relevant aspects? What am I NOT addressing?
• BIAS: Am I anchored on my first solution? Confirming what I already believe? Would someone with different expertise disagree?

Bias signals to watch: Anchoring (first answer feels "obviously right"), confirmation (only supporting evidence found), availability (overweighting recent examples), authority (accepting a framework because it's popular, not because it fits).

After verification, update confidence scores. If scores don't change, verification was superficial. [EXPLICIT]

4. SYNTHESIZE

Combine verified sub-answers into a coherent response. [EXPLICIT]

1. Start with highest-confidence sub-answers (foundation). [EXPLICIT]
2. Layer moderate-confidence answers with caveats. [EXPLICIT]
3. Resolve conflicts: identify source of tension between contradicting sub-answers. [EXPLICIT]
4. Mark clearly: certainty vs hypothesis. [EXPLICIT]
5. Compute global confidence: Global = sum(sub_confidence * sub_importance) / sum(sub_importance). [EXPLICIT]

Conflict resolution priority: Verified facts > Logical deductions > Expert consensus > Reasonable inference > Flagged speculation.

5. REFLECT

If global confidence < target (default 0.95):

1. Identify the weakest sub-problem. [EXPLICIT]
2. Diagnose: data gap, logic gap, or expertise gap?
3. If retrievable — gather info and retry that sub-problem. [EXPLICIT]
4. If not retrievable — deliver with explicit uncertainty flagged. [EXPLICIT]
5. If fundamentally unknowable — state this clearly. [EXPLICIT]

Reflection questions: What is the single biggest weakness? If I'm wrong, what's the most likely cause? What would a disagreeing expert say? Is there a simpler explanation I'm overlooking?

Information Recursion

Before producing any answer, scan:

1. Thread context — all previous messages
2. Attachments — all files and documents provided
3. Stated preferences — format, depth, style instructions
4. Implicit context — time constraints, audience, stakes

Always include a completeness statement:

Sources reviewed: [thread messages and attachments consulted]
Information gaps: [what was NOT available that would improve confidence]

Fast Path

Skip DSVSR when the question has fewer than 2 complexity signals. Answer directly with clarity and relevant caveats. [EXPLICIT]

Complexity signals: Multi-domain, 50+ words, high ambiguity, high stakes, multiple dependencies.

Agent Delegation

Route specialized sub-problems to domain experts:

Domain	Route
Mathematical/statistical	math-reasoning subagent
Code analysis/debugging	code-analysis subagent
Data analysis	data-analysis subagent
General knowledge	handle inline

Each delegate returns: answer + confidence + evidence. Task-engine aggregates. [EXPLICIT]

Confidence Metadata Output

Every substantial response includes:

---
REASONING METADATA:
- Global confidence: [0.XX]
- Sub-problem confidence: [SP-1: 0.XX, SP-2: 0.XX, ...]
- Sources consulted: [thread messages, attachments, knowledge]
- Weaknesses identified: [if any]
- Rigidity level: [exploratory | analytical | executive]
- Verification status: [all checks passed | N flags raised]

Rigidity levels:

• Exploratory — multiple valid paths, presenting options
• Analytical — clear methodology applied, confident in process
• Executive — single recommendation, high confidence, ready to act

Pipeline Integration

[input-analyst] → [task-engine] → [excellence-loop] → User

Higher input quality from input-analyst raises baseline confidence, reducing DSVSR iterations. [EXPLICIT]

Assumptions & Limits

• Confidence scores are self-assessed, not externally validated. They indicate reasoning quality, not ground truth.
• The 0.95 default target is aggressive. For exploratory work, lower to 0.80.
• DSVSR adds latency. Do not run full protocol on simple questions — the fast path exists for a reason.
• Agent delegation assumes subagents are available. If not, handle inline with a confidence penalty note.
• Maximum 3 REFLECT retries to prevent infinite loops.

Edge Cases

• Circular dependencies in decomposition: If SP-2 depends on SP-3 and SP-3 depends on SP-2, merge them into a single sub-problem or introduce an assumption to break the cycle.
• All sub-problems below 0.50: The problem is likely under-specified. Ask the user for clarification rather than guessing.
• Conflicting sub-answers with equal confidence: Escalate the conflict explicitly. Present both interpretations and let the user decide.
• User provides their own confidence target: Honor it, even if it's 1.0 (just note if unachievable and why).
• No thread context available (first message): Note the cold-start limitation in metadata.

Validation Gate

Before delivering a DSVSR response, confirm:

• Every sub-problem has a confidence score with justification
• Verification changed at least one confidence score (or is documented as confirming all)
• Global confidence is computed and reported
• Weaknesses are explicitly stated (not "none found")
• Metadata block is present and complete
• Conflicts between sub-answers are resolved or escalated

Reference Files

• references/dsvsr-protocol.md — Worked examples per stage
• references/confidence-calibration.md — Calibration methodology and common mis-calibration patterns
• references/complexity-heuristics.md — When to activate full DSVSR vs fast path
• references/recursion-protocol.md — Thread and attachment scanning protocol
• references/agent-delegation-patterns.md — Subagent routing and confidence aggregation

---

Author: Javier Montaño | Last updated: 2026-03-18