/sentinel

user: "/sentinel Should I accept this acquisition offer for $18M?" assistant: "[Runs triage -> FULL protocol -> structure-builder first, then questioner, reality-checker, failure-finder in parallel, then synthesis with calibration prediction]" user: "/sentinel Which restaurant for the team dinner?" assistant: "[Triage -> QUICK -> just asks 2-3 clarifying questions, no heavy analysis]"

/sentinel - Decision Analysis

Step 1: Triage — score inline

Assess the 5 dimensions from the user's description (score 1–5 each):

• stakes : How bad is a wrong decision? (1=trivial, 5=catastrophic)
• complexity : How many interdependent factors? (1=simple, 5=very complex)
• time_pressure : How urgent? (1=plenty of time, 5=deciding right now)
• expertise : How well does the user know this domain? (1=expert, 5=novice)
• reversibility : How hard to undo? (1=easy to reverse, 5=permanent)

Optional modifiers:

• group : Multiple decision-makers? (true/false) → +8.0 if true
• advocacy_mode : User is pitching rather than evaluating? → +5.0 if true
• novelty : How unprecedented is this decision? (1–5) → +(novelty−1) × 1.5
• commitment_horizon : Duration of commitment (1=days, 5=years+) → +(horizon−1) × 2.0

Formula:

expertise_score = 6 − expertise

total = (stakes × 3) + (complexity × 2) + (time_pressure × 1)
      + (expertise_score × 1.5) + (reversibility × 2)
      + modifiers

normalized = (total / 74.5) × 100

Routing:

time_pressure ≥ 5  →  RAPID_STRUCTURE
normalized ≥ 78    →  FULL
normalized ≥ 60    →  STANDARD
normalized ≥ 40    →  LITE
normalized ≥ 20    →  QUICK
otherwise          →  NO_PROTOCOL (just answer directly)

Note on thresholds: The cutoffs 40/60/78 are calibrated heuristics, not empirically derived boundaries. A decision scoring 59 is not meaningfully different from one scoring 61. Use the routing as a starting point — if context suggests a higher protocol is warranted (e.g., a LITE-scoring decision with an irreversible consequence), escalate. The triage score surfaces the relevant dimensions; the routing is a default, not a rule.

Agent flags to activate:

• activate_group_facilitator if group = true
• activate_temporal_auditor if commitment_horizon ≥ 2 OR protocol is FULL/STANDARD
• activate_scope_checker if protocol is FULL or STANDARD
• reduce_reality_checker_confidence if novelty ≥ 4
• advocacy_mode_warning if advocacy_mode = true

Show triage result:

Triage: stakes=[n] complexity=[n] time=[n] expertise=[n] reversibility=[n]
Score: [raw] → [normalized]%  |  Protocol: [RAPID_STRUCTURE / FULL / STANDARD / LITE / QUICK]
[active flags]

Personal profile: Read data/bias-profile.json if it exists. Apply always_add bias IDs to pre-selection, confidence_deflation to MAP scores, timeline_multiplier to estimates. Announce blind spot domain if relevant.

Step 2: Route to protocol

Before routing, pre-select 7 bias IDs relevant to this decision and pass them to the questioner agent as context. Use this mapping (expanded in v8.1 with full Wikipedia catalog):

Decision type signals	Bias IDs to pass
Financial / investment	1 (Anchoring), 7 (Sunk Cost), 23 (Loss Aversion), 9 (Planning Fallacy), 44 (Optimism Bias), 59 (Disposition Effect), 63 (Escalation of Commitment)
M&A / acquisition	1 (Anchoring), 44 (Optimism Bias), 9 (Planning Fallacy), 63 (Escalation), 76 (Survivorship Bias), 81 (Pro-Innovation Bias), 85 (G.I. Joe Fallacy)
People / hiring / team	17 (Halo Effect), 15 (Groupthink), 27 (Attribution Error), 3 (Representativeness), 86 (Ingroup Bias), 48 (False Consensus), 66 (Shared Information Bias)
Strategy / expansion	9 (Planning Fallacy), 6 (Narrow Framing), 15 (Groupthink), 1 (Anchoring), 71 (Groupshift), 76 (Survivorship Bias), 82 (System Justification)
Product / build-buy	7 (Sunk Cost), 1 (Anchoring), 8 (Overconfidence), 51 (Default Effect), 57 (Additive Bias), 73 (Quantification Bias), 63 (Escalation)
Risk / crisis / emergency	65 (Normalcy Bias), 18 (Neglect of Probability), 74 (Salience Bias), 84 (Ostrich Effect), 55 (Risk Compensation), 5 (Framing Asymmetry), 79 (Illusion of Control)
Innovation / technology	81 (Pro-Innovation Bias), 50 (Automation Bias), 41 (Dunning-Kruger), 37 (Attribute Substitution), 72 (Illusion of Explanatory Depth), 68 (Availability Cascade), 57 (Additive Bias)
Group / committee decision	15 (Groupthink), 71 (Groupshift), 66 (Shared Information Bias), 87 (Bandwagon Effect), 89 (Courtesy Bias), 48 (False Consensus), 95 (Bias Blind Spot)
Post-decision review	92 (Hindsight Bias), 91 (Choice-Supportive Bias), 93 (Recency Bias), 14 (Outcome Bias), 28 (Self-Serving Bias), 85 (G.I. Joe Fallacy), 94 (Rosy Retrospection)
Default (unclear)	1 (Anchoring), 7 (Sunk Cost), 8 (Overconfidence), 9 (Planning Fallacy), 15 (Groupthink), 44 (Optimism Bias), 95 (Bias Blind Spot)

Pass the selected IDs to questioner as: bias_ids: [n, n, n, n, n, n, n]

Merge personal profile overrides (if bias-profile.json exists): Add always_add bias IDs from profile to the pre-selected set. Apply confidence_deflation and timeline_multiplier to all MAP outputs. Announce blind spot domain if the decision falls in that category.

---

RAPID_STRUCTURE — time_pressure ≥ 5 (REPLACES old QUICK_CHECK emergency override)

High time pressure = maximum bias risk. Structure is MORE needed, not less. Target completion: 8-10 minutes.

1. Framing neutralization: One neutral sentence restatement of the decision
2. scope-checker: Quick proportionality test — is the response sized right?
3. structure-builder: 3 most critical dimensions only. Score independently.
4. questioner: Exactly 2 questions — the highest-risk assumption and the kill condition
5. failure-finder: 3 failure modes only (not 5-7 — time is genuinely short)
6. Single decision gate: "What one thing, if wrong, makes this clearly incorrect? Is it knowable now?"
7. Output: decision or "acquire X first"

Do NOT add temporal-auditor or group-facilitator in RAPID_STRUCTURE. Do NOT run reality-checker (base rates take time to present properly).

---

QUICK (normalized_score < 40)

Low-stakes, reversible, simple.

1. questioner: 2-3 quick verification questions
2. Done. Don't over-analyze small decisions.

---

LITE (normalized_score 40-60)

1. questioner: 3-4 targeted questions
2. scope-checker: Proportionality check (lightweight)
3. Record prediction if stakes ≥ 3: Read data/decision-ledger.json, append entry, Write data/decision-ledger.json

---

STANDARD (normalized_score 60-78)

1. structure-builder: 3-4 dimensions, MAP scoring
2. questioner: 4-5 questions targeting lowest-confidence dimensions + logic check
3. scope-checker: Full proportionality test
4. temporal-auditor: If activate_temporal_auditor = true (commitment_horizon ≥ 2)
5. Record prediction: Read data/decision-ledger.json, append entry with predicted_probability + review_date, Write data/decision-ledger.json. Then surface to the user: "Prediction recorded. Review date: [date]. Run /sentinel-post [decision-id] then to close the loop."

Output: structured comparison + key questions + prediction

---

FULL (normalized_score ≥ 78)

1. structure-builder: 4-5 dimensions, full MAP scoring (PARALLEL with questioner)
2. questioner: 5-7 questions + logic check + explanatory depth probe
3. reality-checker: Base rates + survivorship audit + reference class + sources
4. failure-finder: Full pre-mortem (5-7 failure modes) + steelman of losing option
5. scope-checker: Full proportionality test
6. temporal-auditor: Always on FULL (commitment horizon audit)
7. group-facilitator: If activate_group_facilitator = true — generate Phase 1 sheets + meeting guide
8. noise-calculator: If 3+ options with numeric MAP scores diverging by > 2 points
9. Disruption mode: Apply one random disruption intervention (see CLAUDE.md)
10. Record prediction + set review date: Read data/decision-ledger.json, append entry, Write data/decision-ledger.json. Surface to user: "Prediction recorded (confidence: [X]%). Review date: [date]. Run /sentinel-post [decision-id] then."

Output: Full analysis with synthesis

NOTE: logic-tester is NO LONGER a separate step in FULL protocol. questioner now includes a logic_check section covering the strongest argument's hidden premises. Invoke logic-tester as a standalone agent only if the user explicitly asks for argument verification.

---

New agent activation summary (v8.1)

Agent	When activated
temporal-auditor	FULL always; STANDARD if horizon ≥ 2; any commitment > 12 months
group-facilitator	group=true flag; or /sentinel-group command
scope-checker	STANDARD + FULL; any resource allocation decision
[disruption mode]	FULL always (one random intervention)

Step 3: Synthesis

After all agents run, provide a SYNTHESIS that:

• States the MAP result (scores + confidence)
• Highlights the most important QUESTIONS (from questioner)
• Shows the reality gap (from reality-checker)
• Lists top 3 failure modes (from failure-finder)
• Ends with: "Before deciding, answer these [2-3] questions first."

Template offer (always append after synthesis): Based on the active domain skill, offer the most relevant template to fill out:

Active domain	Primary template to offer	Read from
hiring	candidate-scorecard	skills/hiring/templates/candidate-scorecard.md
ma	ma-target-evaluation	skills/ma/templates/ma-target-evaluation.md
ma (committee)	noise-audit-committee	skills/ma/templates/noise-audit-committee.md
ma (integration)	integration-premortem	skills/ma/templates/integration-premortem.md
product-management	feature-scorecard	skills/product-management/templates/feature-scorecard.md
retrospective	postmortem-scorecard	skills/retrospective/templates/postmortem-scorecard.md
strategy-marketing	campaign-premortem	skills/strategy-marketing/templates/campaign-premortem.md
none / generic	decision-record	templates/decision-record.md

After synthesis output, add:

📋 TEMPLATE AVAILABLE
Based on this analysis, I can fill in [template-name] using the MAP scores and
failure modes above. Say "fill the template" to generate a completed document,
or "skip" to continue without it.

When user says "fill the template": Read the template file, populate all fields using data from the session (MAP scores, failure modes, predictions, questions), and Write the completed document to data/1-Projets/[decision-id]-[template-name].md.

CRITICAL: Structure FIRST, questions SECOND

The structure-builder (MAP) runs BEFORE or IN PARALLEL with the questioner. This is the key architectural choice: the structure forces independent evaluation, and the questions fill in the gaps the structure reveals.

Previous versions ran bias-scanning first. That's backwards. Choice architecture first, debiasing questions second.

Language

Detect and respond in the user's language.

Honesty clause

If the decision is obvious and doesn't need analysis, say so: "This doesn't need Sentinel. Just [do X]." Don't over-process simple choices.