probe

name	probe
description	Run the composed fan→funnel pipeline N times on the same situation to map the decision landscape. Produces a variance report (what's stable vs. variable across runs) and a decision landscape map (basins, ridges, load-bearing assumptions). Writes to <situation-dir>/probes/ with self-contained run-0N/ subdirectories. Use when the user types '/probe [situation]' or asks to probe a decision space.
license	MIT
compatibility	["codex","claude-code","opencode"]
metadata	{"author":"cyberneutics","version":"0.1.0","tags":["deliberation"],"tested-tools":["claude-code"]}

Probe Skill

Run the composed scenarios→committee pipeline (the deliberated choice) multiple times on the same situation. Compare the resolutions. Identify what persists across runs (eigenforms — robust decisions) and what varies (residues — decisions sensitive to framing, parameters, or stochasticity). Produce a variance report and decision landscape map.

This is the iteration operation: repeated walks through the decision space. Each run actualizes a different trajectory, and the differences reveal structure.

Situation resolution

The skill determines where to write output using this precedence:

1. --situation <path> — if the user provides this on invocation, use <path> as the situation directory directly.
2. Config file — read situations_root from .agents/committee-config.yml (path relative to repo root), then append <topic-slug>/.
3. Default — use ~/situations/<topic-slug>/.

Once the situation directory is resolved:

• Create the directory if it doesn't exist.
• If situation.md doesn't exist, create it (see scenarios or committee skill for template).
• Write probe output to <situation-dir>/probes/.
• Each run's self-contained artifacts go in <situation-dir>/probes/run-0N/.

Rosters (references/roster.md, references/scenario-roster.md) are bundled with this skill — they are part of the methodology, not the situation output.

When to use

• User types /probe [situation]
• User asks to "map the decision landscape for [X]" or "test stability of [Y]"
• High-stakes decisions where understanding the topology of the decision space matters more than any single resolution
• When a single deliberated choice produced a result the user suspects is sensitive to framing or assumptions
• When the user wants to know: "Would we reach the same conclusion if we ran this again?"

When NOT to use

• The decision is low-stakes or time-sensitive → single /committee or /scenarios + /committee is sufficient
• The user wants to explore possible futures → /scenarios
• The user wants one deliberation → /committee
• Token budget is constrained — Probe runs N full pipelines

Resource equation

(situation → resolution)^N → variance-report  [Probe]
variance-report → decision-landscape-map  [Map]

The Probe runs the composed pipeline N times, collects the resolutions, and synthesizes a variance report. The Map operation converts the report into an actionable decision landscape.

What the skill does

When invoked, the skill:

1. Resolves the situation directory (see Situation resolution above) and frames the situation (or reuses an existing framing if the situation already contains scenarios)
2. Runs N independent deliberated choice cycles (default N=3):
   • Each run generates scenarios independently (fresh fan)
   • Each run deliberates independently (fresh funnel)
   • Runs do not share context — no cross-contamination
3. Compares the N resolutions to produce a variance report
4. Synthesizes the variance report into a decision landscape map
5. Presents findings to the user: what's stable, what's variable, what's load-bearing

Run independence

Each run must be genuinely independent:

• Fresh scenario generation (different stochastic trajectories)
• Fresh committee deliberation (different debate dynamics)
• No reference to other runs during generation
• The situation framing is the only shared input

This ensures that agreement across runs reflects the decision landscape's structure, not cross-contamination.

Parameters

Parameter	Default	Description
N (run count)	3	Number of independent runs. Minimum 2, recommended 3, maximum 5 unless user overrides.
Scenario parameters	Per-run defaults	Optionally fix scenario parameters across runs (same axes, same horizon) to isolate committee-level variance, or let them vary to capture full pipeline variance.
scenario_context	None	Optionally point to an existing scenario set (e.g. <situation-dir>/scenarios/) to reuse across runs (only the committee varies). This isolates funnel-level instability.

Variance isolation modes

Mode	Scenarios	Committee	What it measures
Full variance (default)	Fresh per run	Fresh per run	Total pipeline instability
Funnel-only	Fixed (reuse one scenario set)	Fresh per run	Committee-level instability given fixed futures
Fan-only	Fresh per run	Fixed parameters	Scenario-level instability (how different the futures are across runs)

Full variance is the default because it maps the complete decision landscape. Funnel-only and fan-only modes are diagnostic — use them when full variance reveals instability and you want to locate its source.

Probe record directory

Location: <situation-dir>/probes/ (see Situation resolution above for how <situation-dir> is determined).

Files:

00-probe-charter.md

---
probe:
  date: YYYY-MM-DD
  situation: "short description"
  N: 3
  mode: full_variance | funnel_only | fan_only
  scenario_parameters: {} | {fixed parameters}
  situation_source: "user input" | "<situation-dir>/scenarios/"
---

Body: the framed situation (or reference to existing framing).

run-01/ through run-0N/

Each subdirectory contains the full artifacts from one deliberated choice cycle:

run-01/
  scenarios/
    00-situation.md
    01-roster.md
    01-parameters.md
    02-scenarios.md
    03-assessment.md
  deliberation/
    00-charter.md
    01-roster.md
    01-convening.md
    02-deliberation.md
    03-resolution.md

These are the standard /scenarios and /committee outputs, nested per run.

04-variance-report.md

The core analytical output. Compares the N resolutions.

---
variance_report:
  date: YYYY-MM-DD
  situation: "short description"
  runs: 3
  mode: full_variance

  eigenforms:
    - finding: "description of what persisted across all runs"
      present_in: [1, 2, 3]
      confidence: high
    - finding: "description of another persistent finding"
      present_in: [1, 2, 3]
      confidence: high

  residues:
    - finding: "description of what varied"
      run_1: "what run 1 said"
      run_2: "what run 2 said"
      run_3: "what run 3 said"
      likely_source: situation_framing | parameters | operational_noise
    - finding: "another variable finding"
      run_1: "..."
      run_2: "..."
      run_3: "..."
      likely_source: parameters

  instability_sources:
    situation_framing:
      - "description of framing ambiguity that drove divergence"
    parameters:
      - assumption: "the load-bearing assumption"
        effect: "when X, resolution says A; when not-X, resolution says B"
    operational_noise:
      - "description of stochastic variation not attributable to substance"

  summary: "2-3 sentence synthesis"
---

Body: narrative analysis structured as:

1. Eigenforms — what every run found (the stable territory): For each eigenform, cite which runs found it and why it's robust.

2. Residues — what varied across runs (the interesting instability): For each residue, describe the variation, diagnose the source, and explain what the variation reveals about the decision landscape.

3. Instability diagnosis — which of the three sources drove the variance:

   • Situation framing: premises underdetermine the possibility space
   • Parameters: load-bearing assumptions that flip outcomes
   • Operational noise: LLM stochasticity not washing out

4. Implications — what the user should do with this information:

   • Eigenforms → commit with confidence
   • Parameter-driven residues → investigate the load-bearing assumptions
   • Framing-driven residues → tighten the situation description
   • Noise-driven residues → ignore (or increase N)

05-landscape-map.md

The topological synthesis. Converts the variance report into actionable structure.

---
landscape_map:
  date: YYYY-MM-DD
  situation: "short description"

  basins:
    - basin_id: A
      description: "description of this cluster of resolutions"
      runs: [1, 3]
      key_commitment: "the decision this basin represents"
      assumptions_required: ["assumption 1", "assumption 2"]
    - basin_id: B
      description: "description of alternative cluster"
      runs: [2]
      key_commitment: "the alternative decision"
      assumptions_required: ["assumption 3"]

  ridges:
    - between: [A, B]
      pivots_on: "the assumption that separates the basins"
      sensitivity: "how small a change flips the outcome"

  load_bearing_assumptions:
    - assumption: "the critical assumption"
      if_true: "basin A"
      if_false: "basin B"
      currently_testable: true | false
      test: "how to test it"

  robust_actions:
    - action: "something that appears across multiple basins"
      present_in_basins: [A, B]
      rationale: "why it's robust"

  monitoring_plan:
    - signal: "early warning sign"
      indicates: "which basin is materializing"
      source_scenario: "which scenario identified this signal"
---

Body: narrative landscape summary structured as:

1. Terrain overview — how many distinct basins? How fragmented is the decision landscape?

2. Basin descriptions — what each cluster of resolutions looks like, what assumptions it requires, what it optimizes for.

3. Ridge analysis — where do outcomes flip? Which assumptions are load-bearing? How sensitive are the transitions?

4. Robust commitments — actions that survive across basins. These are safe to take regardless of which future materializes.

5. Contingent commitments — actions that depend on which basin you're in, with trigger conditions for switching.

6. Monitoring plan — early warning signs that indicate which basin is materializing, sourced from the scenarios.

Execution approach

Given token and time costs, the probe should be transparent about its process:

1. Before starting: Tell the user how many runs will be executed and the expected cost/time. Ask for confirmation if N > 3.
2. During each run: Briefly note which run is in progress ("Run 2 of 3").
3. Between runs: Do not carry forward any specific content from previous runs. The situation framing is the only shared input.
4. After all runs: Generate the variance report and landscape map, then present the key findings inline.

Abbreviated runs

For probing, individual runs can be more compact than standalone deliberated choices:

• Scenarios: 2–3 paragraphs per scenario (not 3–5)
• Committee: standard format (not rigorous/extended)
• Resolution: full format (this is what gets compared)

The resolution is the unit of comparison. Scenario and deliberation detail matters for the individual run but the probe's analytical value comes from comparing resolutions.

What success looks like

After a probe, the user should be able to:

• Name the eigenforms: "Every run agreed that [X]"
• Name the ridges: "The decision hinges on whether [Y]"
• Distinguish robust from fragile commitments: "We can safely do [A] regardless; [B] depends on [Y]"
• Know what to monitor: "If we see [signal], basin [Z] is materializing"
• Decide whether to commit: The landscape map either gives confidence (broad basin, clear eigenforms) or identifies what needs more investigation (narrow ridges, load-bearing assumptions that are untested)

Integration with other skills

Skill	Relationship
/scenarios	Probe runs this internally per run (fan half)
/committee	Probe runs this internally per run (funnel half)
/review	Can be run on individual probe runs' deliberations if quality is in question
/string-diagram	The probe pipeline can be visualized as a string diagram

Files reference

• Committee roster: references/roster.md
• Scenario roster: references/scenario-roster.md
• Deliberated choice workflow: references/deliberated-choice-workflow.md
• Cross-scenario learning: references/cross-scenario-learning.md