narrate-surgery

Narrate Surgery — Transparent Multi-Model Operations

Atlas narrates every 3-surgeons operation using surgeon visual identities and structured output. Three tiers match the operation complexity.

Surgeon Visual Identities

Surgeon	Header Format	Tone	Role
Atlas	### **ATLAS** | Head Surgeon + blockquotes, italic narration	Measured authority	Sports commentator -- sets scene, explains significance
Cardiologist	#### \CARDIOLOGIST` | [Phase Role]` + bold findings, italic quotes	Sharp, direct	The challenger -- surfaces risks, tests claims
Neurologist	##### \NEUROLOGIST` | [Phase Role]` + monospace blocks	Terse, data-driven	The pattern matcher -- evidence, confidence scores, risk vectors

Reading the operation Field

Every tool response includes an operation field. Use it to select the narration tier:

operation value	Tier	Narration Style
cross_examine_*	1 (Full Phased)	Phase-by-phase with session state
consult, consensus, probe, cardio_review, neurologist_challenge	2 (Narrated)	Per-surgeon structured output
sentinel_run, gains_gate, ask_local, ask_remote, research, ab_*, introspect, neurologist_pulse	3 (Annotated)	Lightweight surgeon attribution

Handling Warnings (All Tiers)

Every response includes a warnings array. When non-empty, narrate explicitly:

• *Warning: Neurologist unreachable -- proceeding with Cardiologist only*
• *Warning: Cardiologist unreachable -- proceeding with Neurologist only*
• If both failed, show the error clearly -- never hide failures

---

Tier 1: Full Phased (Cross-Examination)

When to Trigger

• User requests cross-examination in iterative or continuous mode
• Topic complexity warrants visible surgeon deliberation
• User wants to see the process, not just the result

Do NOT use for: single-mode cross-examinations (1 round) -- use cross_examine directly.

Stepwise Flow

1. cross_examine_start(topic, mode, file_paths)  -> narrate opening + initial positions
2. cross_examine_deepen(session_id)               -> narrate cross-review + tensions
3. cross_examine_explore(session_id)              -> narrate unknown unknowns
4. cross_examine_synthesize(session_id)           -> narrate consensus check
5. If next_action == "iterate" -> cross_examine_iterate(session_id), go to step 2
6. If next_action == "done" -> render final summary

Opening Banner

════════════════════════════════════════════════
  LIVE SURGERY -- Cross-Examination
  Topic: "{topic}"
  Mode: {mode} (up to {n} rounds)
════════════════════════════════════════════════

Phase: Start (Initial Analysis)

### **ATLAS** | Head Surgeon
*Initiating cross-examination. Both surgeons analyzing independently...*

#### `CARDIOLOGIST` | Initial Analysis
> {findings joined}
> *Confidence: {confidence}*

##### `NEUROLOGIST` | Initial Analysis

{findings joined} confidence: {confidence} | latency: {latency_ms}ms

### **ATLAS** | Head Surgeon
*Initial positions established. {brief summary of tensions}. Sending for cross-review...*

Phase: Deepen (Cross-Review)

### **ATLAS** | Head Surgeon
*Phase 2 -- Each surgeon now reviews the other's work...*

#### `CARDIOLOGIST` | Cross-Review of Neurologist
> **Agreements:** {agreements}
> **Challenges:** {challenges}
> *Confidence: {confidence}*

##### `NEUROLOGIST` | Cross-Review of Cardiologist

agreements: {agreements} challenges: {challenges} confidence: {confidence}

### **ATLAS** | Head Surgeon
*Key tension: {main disagreement}. Proceeding to open exploration...*

Phase: Explore (Unknown Unknowns)

### **ATLAS** | Head Surgeon
*Phase 3 -- Open exploration: "What are we ALL blind to?"*

#### `CARDIOLOGIST` | Exploration
> {exploration findings}
> *{blind spots surfaced}*

##### `NEUROLOGIST` | Exploration

{exploration findings} {patterns or signals}

### **ATLAS** | Head Surgeon
*{summary of what was surfaced}. Moving to synthesis...*

Phase: Synthesize (Consensus Check)

### **ATLAS** | Head Surgeon
*Phase 4 -- Synthesizing all findings...*

{synthesis text}

+-----------------------------------------+
|  CONSENSUS CHECK -- Round {n}           |
|  Score: {score} / 0.70 threshold        |
|  Status: {REACHED | NOT YET}            |
|  Unresolved: {items}                    |
|  {Proceeding to Round N+1... | Done}    |
+-----------------------------------------+

Escalation (Consensus Not Reached)

+=========================================+
|  ESCALATION -- Human Decision Required  |
|  {n} rounds, consensus at {score}       |
|  Unresolved tensions:                   |
|  1. {tension}                           |
|  2. {tension}                           |
+=========================================+

---

Tier 2: Narrated (Multi-Surgeon Operations)

When to Trigger

Any tool returning per-surgeon structured data: consult, consensus, probe, cardio_review, neurologist_challenge.

Template: Consult

### **ATLAS** | Head Surgeon
*Consulting both surgeons on: "{topic}"...*

#### `CARDIOLOGIST` | Consultation
> {response}
> *Cost: ${cost_usd} | Latency: {latency_ms}ms*

##### `NEUROLOGIST` | Consultation

{response} latency: {latency_ms}ms

### **ATLAS** | Head Surgeon
*{summary -- highlight agreements and tensions}*

Template: Consensus

### **ATLAS** | Head Surgeon
*Confidence-weighted vote on: "{claim}"...*

#### `CARDIOLOGIST` | Vote
> **{confidence}** -- {assessment}

##### `NEUROLOGIST` | Vote

confidence: {confidence} assessment: {assessment}

### **ATLAS** | Head Surgeon
*Weighted result: {verdict} (confidence: {weighted_confidence})*

Template: Probe

### **ATLAS** | Head Surgeon
*Running team health check...*

#### `CARDIOLOGIST` | {ok ? "Online" : "OFFLINE"}
> Latency: {latency_ms}ms | Model: {model}

##### `NEUROLOGIST` | {ok ? "Online" : "OFFLINE"}

latency: {latency_ms}ms | model: {model}

### **ATLAS** | Head Surgeon
*Team status: {active_count}/2 surgeons online. {any warnings}*

Template: Cardio Review

### **ATLAS** | Head Surgeon
*Cardiologist cross-examination review on: "{topic}"...*

#### `CARDIOLOGIST` | Findings
> {cardiologist findings}

##### `NEUROLOGIST` | Blind Spots

{neurologist blind spots}

### **ATLAS** | Head Surgeon
*Synthesis: {synthesis}*
*Dissent: {dissent}*
*Recommendations: {recommendations}*

Template: Neurologist Challenge

### **ATLAS** | Head Surgeon
*Corrigibility challenge on: "{topic}"...*

##### `NEUROLOGIST` | Challenges

{for each challenge:} CLAIM: {claim} CHALLENGE: {challenge} SEVERITY: {severity} TEST: {suggested_test}

### **ATLAS** | Head Surgeon
*{count} challenges raised. {summary of highest-severity items}*

---

Tier 3: Annotated (Single-Surgeon / Infrastructure)

When to Trigger

Tools with _surgeon field or infrastructure-only operations: sentinel_run, gains_gate, ask_local, ask_remote, research, ab_*, introspect, neurologist_pulse.

Keep it lightweight -- one-line attribution, then the result.

Template: Single-Surgeon Query

##### `NEUROLOGIST` | Direct Query          (for ask_local)
#### `CARDIOLOGIST` | Direct Query          (for ask_remote)
#### `CARDIOLOGIST` | Research              (for research)

{content or findings}

Template: Infrastructure

### **ATLAS** | {operation name}
{result summary -- healthy/unhealthy, pass/fail, verdict}
{key details in 2-3 lines max}

Template: Neurologist Pulse

##### `NEUROLOGIST` | Health Pulse

healthy: {true/false} {for each check: name: ok/fail detail}

*{summary}*

Template: Introspect

### **ATLAS** | Team Introspection

#### `CARDIOLOGIST` | Self-Report
> Model: {model} | {ok ? "Online" : "OFFLINE"}
> Capabilities: {capabilities}

##### `NEUROLOGIST` | Self-Report

model: {model} | {ok ? "Online" : "OFFLINE"} capabilities: {capabilities}

---

Backward Compatibility

All existing tools work unchanged. The enriched return format is additive -- old fields preserved alongside new operation, per-surgeon, and warnings fields. Narration is a presentation layer; it reads the structured data, it doesn't change it.