forge-deliberation

Deliberation Methodology

Loaded during the Reason phase. Governs how the Reasoner generates, tests, and scores competing hypotheses.

When to use

Load this skill at the start of Phase R (Reason) for standard single-trace deliberation. Use it to generate ≥3 genuinely distinct hypotheses, stress-test each with the four adversarial checks (Inversion, Boundary, Pre-Mortem, Dependency), score on the 5-dimension rubric (Evidence Alignment 30%, Constraint Satisfaction 25%, Risk Profile 20%, Reversibility 15%, Second-Order Clarity 10%), and execute 1–3 passes as set by the depth score from Phase F. Do not use when Self-Consistency Mode (G2/TRANCE) fires — load forge-self-consistency instead.

---

Hypothesis Generation

The 3-Hypothesis Minimum

Generate ≥3 genuinely distinct hypotheses. "Distinct" means they lead to different actions, not just different phrasings of the same recommendation.

Quality test: If removing one hypothesis doesn't change the deliberation outcome, it was a strawman. Replace it.

Generation Strategies by Decision Type

Decision Type	How to Generate Hypotheses
TRADE-OFF	One hypothesis per viable option. Include "do nothing" or "defer" if applicable.
FEASIBILITY	H1: Feasible as specified. H2: Feasible with modifications (name them). H3: Not feasible (name the blocking constraint).
ROOT-CAUSE	One hypothesis per plausible causal chain. Each must name the initiating event, the propagation path, and why other safeguards didn't catch it.
CONFLICT-RESOLUTION	One hypothesis per stakeholder position, plus a synthesis position that partially satisfies all.
CONSTRAINT-SATISFACTION	H1: Satisfy all constraints (name the solution shape). H2: Relax soft constraint S1 (name what improves). H3: Relax soft constraint S2 (name what improves).
RISK-ASSESSMENT	One hypothesis per risk cluster. Group related failure modes rather than listing them atomically.

Hypothesis Structure

Each hypothesis follows this format:

### Hypothesis H-N: [Name]

**Position**: [One sentence: what this hypothesis asserts]

**Requires to be true**:
- [Condition 1]
- [Condition 2]

**Evidence supporting**:
- [E-ID]: [How it supports] (reliability: H/M/L)

**Evidence opposing**:
- [E-ID]: [How it opposes] (reliability: H/M/L)

**Second-order effects**:
- [Effect 1: consequence the requester may not have considered]
- [Effect 2]

**Falsification test**: This hypothesis is wrong if [specific observable condition].

---

Stress-Testing Protocol

For each hypothesis, run these adversarial checks:

1. Inversion Test

"If the opposite of this hypothesis were true, what evidence would I expect to see? Do I see it?"

If you see evidence consistent with the inversion and cannot explain it away, reduce confidence.

2. Boundary Condition Test

"Under what extreme but realistic conditions does this hypothesis break?"

Name the conditions explicitly. If the hypothesis only works in a narrow band of conditions, that's a risk.

3. Pre-Mortem Test

"Assume we acted on this hypothesis and it failed. What was the most likely cause of failure?"

The answer is a [RISK] marker in the final verdict.

4. Dependency Test

"What external factors must remain stable for this hypothesis to hold?"

Each dependency that could change is a [REVERSAL-CONDITION].

---

Scoring Rubric

Score each hypothesis on 5 dimensions. Each dimension is 1–5.

Dimension	1	3	5	Weight
Evidence Alignment	Contradicted by high-reliability evidence	Mixed or partial support	Strongly supported by multiple H-reliability sources	30%
Constraint Satisfaction	Violates ≥1 hard constraint	Satisfies hard, violates soft	Satisfies all hard and soft constraints	25%
Risk Profile	High-probability, high-impact failure modes	Moderate risks with known mitigations	Low residual risk after mitigations	20%
Reversibility	Irreversible; high sunk cost if wrong	Partially reversible with effort	Easily reversible; low switching cost	15%
Second-Order Clarity	Unpredictable downstream effects	Some known, some uncertain	Well-understood consequences	10%

Composite score = Σ(dimension_score × weight)

Scoring Rules

1. Score before comparing. Score each hypothesis independently, then compare. Do not anchor on the first hypothesis scored.
2. Evidence-gated scoring. If you cannot justify a score with specific evidence, score 3 (neutral) and flag with [GAP].
3. Tied scores. If the top two hypotheses score within 0.3 of each other, the verdict must acknowledge genuine ambiguity rather than forcing a winner.
4. Sensitivity check. After scoring, ask: "If I changed any single dimension score by ±1, would the winner change?" If yes, the verdict is sensitive and confidence should be reduced.

---

Reasoning Patterns for Specific Decision Types

Trade-Off Reasoning

Build a decision matrix:

| Criterion (weight) | Option A | Option B | Option C |
|---------------------|----------|----------|----------|
| Performance (30%) | 4 — meets p99 target | 3 — marginal | 5 — exceeds |
| Operational cost (25%) | 2 — new infra required | 4 — fits existing | 3 — moderate change |
| Team familiarity (20%) | 5 — known stack | 2 — learning curve | 3 — partial overlap |
| Reversibility (15%) | 3 — moderate | 4 — easy rollback | 2 — schema migration |
| Ecosystem maturity (10%) | 4 — stable | 5 — dominant | 2 — early stage |
| **Weighted total** | **3.35** | **3.35** | **3.30** |

When scores are close, shift to qualitative tiebreakers: which option has fewer [ASSUMPTION] markers? Which has clearer [REVERSAL-CONDITION]s?

Root-Cause Reasoning

Build a causal chain for each hypothesis:

[Trigger Event] → [Propagation Step 1] → [Propagation Step 2] → [Observed Failure]
     ↑                    ↑                      ↑
 Evidence: E-3        Evidence: E-1           Evidence: E-5

Eliminate hypotheses whose causal chains have gaps (no evidence for a propagation step) unless the gap is explicitly marked and the chain is otherwise the strongest.

Feasibility Reasoning

For each hypothesis, walk the constraint list:

| Constraint | Status | Evidence | Notes |
|-----------|--------|----------|-------|
| C1: p99 < 200ms | ✅ PASS | E-2: benchmark shows 140ms | Margin: 30% |
| C2: No new spend | ⚠️ CONDITIONAL | E-4: requires Redis, ~$200/mo | Soft constraint — flagged |
| C3: PostgreSQL 14 | ✅ PASS | E-1: schema compatible | Verified |

Feasibility = all hard constraints pass. Conditional feasibility = hard pass, soft fail.

---

Multi-Pass Deliberation

When deliberation depth is Standard (2 passes) or Deep (3 passes):

Pass 1: Generate hypotheses, score, select top 2. Pass 2: Expand top 2. For each, enumerate implementation implications, edge cases, and failure modes not visible in Pass 1. Re-score. Pass 3 (Deep only): Red-team the leading hypothesis. Actively construct the strongest possible argument against it. If the argument holds, the verdict must acknowledge it as a live risk.

Convergence Check Between Passes

If the winning hypothesis changes between passes, this is a signal of genuine ambiguity. The verdict's confidence score should decrease by 10–15 points.

If the winning hypothesis is the same across all passes, confidence increases by 5–10 points.

---

Plan Checkpoints (CRYSTALIUM execution layer)

During multi-pass deliberation, FORGE checkpoints each reasoning branch to the CRYSTALIUM execution layer. This makes the deliberation history auditable and recoverable — if context is exhausted mid-deliberation, a plan_replan call records which alternative was selected and why.

Checkpoint at each pass boundary

After scoring is complete for a pass, record the deliberation state:

mcp__crystalium__plan_checkpoint(
  plan_id  = <decision-id, e.g. "forge-<thread_id>-pass<N>">,
  state    = {
    pass:         <pass number>,
    scope:        <framed question>,
    top_hypotheses: [{ id, name, score }],
    gaps:         [<active GAP markers>]
  },
  step     = "pass<N>-scored",
  metadata = { depth: <simple|standard|deep>, project: <cwd-project> }
)

Replan when a branch is revised

If the winning hypothesis changes between passes (or a REFORGE revises the leading alternative), record the branching decision before continuing:

mcp__crystalium__plan_replan(
  plan_id            = <same plan_id as prior checkpoint>,
  from_checkpoint_id = <checkpoint_id returned by prior plan_checkpoint>,
  new_plan           = {
    selected_hypothesis: <id and name>,
    supersedes_id:       <id of demoted hypothesis>,
    reason:              <why the branch changed>
  }
)

The old checkpoint is preserved (bi-temporal); the replan creates a new version. This produces an auditable deliberation trace: which alternatives were live at each pass, and which branch was promoted and why.

Graceful skip: if mcp__crystalium__* tools are unavailable (CRYSTALIUM not installed), proceed without checkpoints — never hard-fail. Deliberation continues normally without execution-layer recording.

---

Anti-Patterns in Reasoning

Anti-Pattern	Signal	Remedy
Anchoring	First hypothesis is always the winner	Score all before comparing; randomize evaluation order mentally
False dichotomy	Only 2 options considered	Force a third: "What if we did neither? What if we did both?"
Survivorship reasoning	"X worked for Company Y"	Ask: "How many companies tried X and failed? Would we know?"
Complexity bias	More sophisticated option scores higher by default	Add "simplest viable option" as a mandatory hypothesis
Sunk cost reasoning	Prior investment influences scoring	Score based on forward-looking costs only; mark prior investment as context, not evidence
Premature convergence	Decided after one pass on a deep problem	Check deliberation depth; enforce required passes

---

Reasoner — Deliberation Skill