Stakes Assessor

Stakes Assessor — Fast-Lane vs Deep-Lane Gate

Problem construction is automatic (Reiter-Palmon). A friction-heavy suite that overlaps with cognitive default will get bypassed precisely when most needed. The fast-lane / deep-lane split is binding-vow's structural answer: low-stakes statements clear in ≤30 seconds; high-stakes statements get the full seven-phase audit. This skill makes the call.

The skill is opinionated: it defaults toward higher stakes when uncertain. The cost of running standard mode on a low-stakes statement is wasted seconds; the cost of running quick mode on a high-stakes statement is a misframed decision.

The Three Levels

| Stakes | Mode default | Time budget | Phase coverage | |---|---|---|---| | Low | quick | ≤30s | Phases 1, 5, 6 (statement-grader only) | | Medium | standard | 2–5 min | All phases, single pass | | High | deep | 5–15 min | All phases + frame rotation ≥3 alternatives + dual-audience compression + full audit |

The mode is a default. The orchestrator can upgrade (e.g., quick → standard if early grader signals fail) but rarely downgrades.

Five-Axis Scoring

Stakes are scored across five axes. Each axis returns low / medium / high; aggregate is the maximum (worst-case rules).

| Axis | Low signal | Medium signal | High signal | |---|---|---|---| | Reversibility | Easily reversible (text edit, draft) | Reversible with effort (project pivot, role change) | Irreversible (public publication, capital commitment, signed contract) | | Magnitude | Personal/team scale, low resource | Mid-scope: a project, a quarter, a hire | Org-wide, multi-quarter, multi-million, public-facing | | Time horizon | This week | This quarter | Multi-year or career-long | | Audience reach | Self or 1–3 people | A team / department | Public, board, organization-wide, regulator | | Optionality | Many easy alternatives if this is wrong | A few alternatives, with switching cost | Few or no alternatives; locked in if committed |

Aggregate rule: the worst (highest-stakes) of the five axes drives the verdict. A statement that's low on four axes and high on one (e.g., irreversible) is high stakes.

Detection Signals

Quick triage before running the five-axis scoring:

| Phrase / context | Likely stakes | |---|---| | "Quick thought on...", "FYI...", "rough draft" | Low | | "I'm noodling on...", "thinking through..." | Low or Medium | | "We need to decide whether..." | Medium | | "This is for the board / regulator / public release" | High | | "Once we ship this we can't roll back" | High (irreversibility) | | "Multi-quarter program", "8-figure decision", "headcount allocation" | High (magnitude) | | "Career decision", "company strategy", "legal exposure" | High | | Statement is about an essay angle (not yet drafted) | Low to Medium (draft is reversible) | | Statement is about a public-facing artifact about to ship | High (irreversibility + reach) |

Process

1. Quick triage — scan for any High signal in the table above. If present, jump to step 4.
2. Score the five axes explicitly. Each axis: low / medium / high.
3. Aggregate — take the worst-case (highest) axis as the verdict.
4. Surface irreversibility as a separate flag — even at medium aggregate, if any axis is "irreversible," flag it. The orchestrator may upgrade to deep mode based on this alone.
5. Return verdict + axis scores + irreversibility flag.

Output Format

STAKES — [first 60 chars of statement...]
─────────────────────────────────────────────
Verdict: [low | medium | high]
Mode default recommendation: [quick | standard | deep]

Axis scores:
  Reversibility:     [low | medium | high]   [signal]
  Magnitude:         [low | medium | high]   [signal]
  Time horizon:      [low | medium | high]   [signal]
  Audience reach:    [low | medium | high]   [signal]
  Optionality:       [low | medium | high]   [signal]

Irreversibility flag: [none | present — note]
Rationale: [which axis drove the verdict]

When to Upgrade

The orchestrator may upgrade the mode beyond the stakes default in these cases:

| Trigger | Upgrade | |---|---| | Irreversibility flag is set, even at medium aggregate | medium → deep | | Audience-classifier returns public with skeptical sub-context | upgrade by one level | | Problem-typology returns wicked or mess | upgrade by one level | | Early Phase 6 grader signals show multiple-axis failure | trigger Re-state Level 3 (deep) |

Conversely, the orchestrator may not downgrade against the stakes verdict. A high-stakes problem doesn't get quick mode just because the user is in a hurry.

Edge Cases

| Pattern | Handling | |---|---| | User explicitly requests quick mode on a high-stakes statement | Run quick mode AND surface the stakes verdict prominently. Don't silently override; let the user own the trade-off | | Statement is high on one axis (e.g., audience reach) but the user dismisses it ("it's just a blog post") | Apply the worst-case rule; flag the disagreement. Public publication is genuinely high-reach even if the author treats it casually | | Stakes are unclear because the statement is too vague to assess | Return "deferred — need more specificity"; route back to Phase 1 | | Stakes shift across iterations (statement starts low-stakes but becomes high-stakes after reformulation) | Re-run stakes-assessor on the reformulated statement; mode may change |

Output Contract for six-eyes

Called from Phase 2 (Diagnose), in parallel with problem-typology and audience-classifier. Returns:

• Stakes verdict (low / medium / high)
• Mode default recommendation
• Per-axis scores
• Irreversibility flag

six-eyes uses these to set the mode for the run. If irreversibility is flagged, the orchestrator upgrades the mode even if aggregate stakes are medium. This is the Reiter-Palmon mitigation: low-stakes statements clear fast, but anything irreversible gets full audit regardless.

Connections

• problem-typology (binding-vow) — runs in parallel; some typologies (wicked, mess, adaptive) imply high stakes by default
• audience-classifier (binding-vow) — runs in parallel; public audience often implies high stakes
• statement-grader (binding-vow) — independent skill but mode-selection feedback loop: catastrophic grader failures may upgrade stakes retroactively
• xy-detector (binding-vow) — XY patterns often hide high-stakes decisions inside low-stakes-shaped questions

Sources

• Reiter-Palmon, R., Mumford, M. D., & Redmond, M. R. (1994). Problem construction model — the "automatic problem construction" finding is the empirical anchor for the fast/deep split.
• The five-axis scoring is derived from synthesizing irreversibility-focused decision theory (Bezos's "Type 1 vs Type 2 decisions") with magnitude-and-reach concepts from operational risk management.
• See [[binding-vow-research-findings]] vault note for the empirical foundation discussion.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: npbuilds
• Source: npbuilds/skill-library
• License: MIT
• Homepage: https://skill-library-prod.web.app/

Install and usage instructions live in the source repository linked above.