Discipline: This command follows references/execution-discipline.md — the Iron Law of verification, the never-start-on-main rule, and the Red Flags below.

Plain Language: Instead of planning forward ("build X, then Y, then Z"), we work backward from the goal: "For our goal to be achieved, what MUST be true?" This surfaces hidden requirements early.

Key terms: A required truth (RT) is a precondition — something that MUST be true for the outcome to succeed. An anchor is the result of backward reasoning from the outcome. The binding constraint is the single hardest-to-close required truth that, if unresolved, blocks all solution options. Convergence is the point where all required truths are satisfied.

Scope Guard

This phase ONLY updates .manifold/<feature>.json and .manifold/<feature>.md with required truths, gaps, and solution options. Solution options are PROPOSALS, not instructions to build -- m4-generate handles implementation. After updating manifold files: display anchoring summary, suggest next step, STOP.

Schema Compliance

Field	Valid Values
Sets Phase	`ANCHORED`
Next Phase	`GENERATED` (via /manifold:m4-generate)
Required Truth Statuses	`SATISFIED`, `PARTIAL`, `NOT_SATISFIED`, `SPECIFICATION_READY`
Required Truth ID Prefix	RT-1, RT-2, RT-3...

See SCHEMA_REFERENCE.md for all valid values. Do NOT invent new statuses.

Output Format: JSON+Markdown Hybrid

CRITICAL: Generate TWO outputs, not one YAML file.

1. JSON Structure (IDs, statuses, maps ONLY)

Update .manifold/<feature>.json with required truth references:

{
  "anchors": {
    "required_truths": [
      {
        "id": "RT-1",
        "status": "NOT_SATISFIED",
        "maps_to": ["B1", "T1"],
        "parent": "OUTCOME",
        "relevance": 0.95,
        "confidence": 1.0,
        "evidence": [
          {"id": "E1", "type": "file_exists", "path": "src/retry/error-classifier.ts", "status": "PENDING"}
        ]
      }

Key rule: JSON contains NO text content. Only IDs, statuses, and constraint mappings.

2. Markdown Content (statements and gaps)

Update .manifold/<feature>.md with required truth content:

## Required Truths

### RT-1: Error Classification System

Can distinguish transient from permanent failures.

**Gap:** No current error taxonomy exists.

### RT-2: Idempotent Retries

Retries are idempotent via transaction idempotency keys.

**Gap:** Current system lacks idempotency implementation.

---

## Solution Space

### Option A: Client-side Exponential Backoff
- Satisfies: RT-3
- Gaps: RT-2, RT-4, RT-5
- Complexity: Low

### Option B: Server-side Workflow Engine
- Satisfies: RT-1, RT-2, RT-3, RT-4, RT-5
- Gaps: None
- Complexity: High

### Option C: Hybrid (Client retry + Server queue) ← Recommended
- Satisfies: RT-1, RT-3, RT-5
- Gaps: None (with implementation)
- Complexity: Medium

Markdown Heading Rules

ID Pattern	Markdown Heading Level	Example
RT-1, RT-2	`###` (h3)	`### RT-1: Error Classification`

Iteration Recording

Append to "iterations" array in JSON when phase completes:

{
  "iterations": [
    {
      "number": 3,
      "phase": "anchor",
      "timestamp": "2026-04-04T00:00:00Z",
      "result": "Derived 6 required truths, recommended Option A",
      "required_truths": 6,
      "by_status": { "SATISFIED": 0, "PARTIAL": 1, "NOT_SATISFIED": 3, "SPECIFICATION_READY": 2 },
      "solution_options": 3,
      "selected_option": "Option A"
    }
  ]
}

REQUIRED: Every iteration MUST have number, phase, timestamp, and result (string). Omitting result fails schema validation.

Evidence Types (v3):

file_exists - Verify file exists on disk
content_match - Grep for pattern in file
test_passes - Run test and check exit code
metric_value - Check runtime metric threshold
manual_review - Requires human verification

Usage

/manifold:m3-anchor <feature-name> [--outcome="<statement>"] [--skip-lookup]

Why Backward Reasoning?

Forward planning starts with a spec and discovers missing requirements late. Backward reasoning starts with the outcome and asks "What must be TRUE?" — this surfaces hidden assumptions and implicit requirements early, before implementation begins.

Process

State the outcome - Clear, measurable success criteria
Check blocking dependencies - Read blocking_dependencies from JSON (written by m2). If present, derive required truths for the BLOCKER constraints FIRST -- these are the strongest candidates for the binding constraint. If draft_required_truths (from m1) exist for blocking constraints, validate those as starting seeds.
Ask "What must be TRUE?" - Necessary conditions for outcome
Derive required truths - Chain backward from each condition
Score each edge - For every RT, record its parent ("OUTCOME" or a parent RT) and the edge's relevance and confidence (see Edge Map below)
Identify gaps - What's missing between current state and requirements?
Generate solution space - Options that satisfy all required truths

Edge Map: Relevance & Confidence (the backward-reasoning tree is a map, not a list)

A flat list of required truths hides the reasoning. Each RT is a node with one parent — either OUTCOME (top-level necessary condition) or another RT (a sub-truth the parent depends on). Every parent→child edge carries two scores in [0,1]:

Field	Question it answers	Anchors
`relevance`	Does "child TRUE" genuinely make "parent TRUE" more achievable — is it a real prerequisite, not decorative?	`1.0` = parent is impossible without it · `0.5` = materially helps · `≤0.3` = tangential ⇒ drop the RT
`confidence`	How necessary is the child — invariant-grade necessity vs. merely one helpful option?	`1.0` = must hold, no substitute · `0.7` = strong but alternatives exist · `0.4` = merely helpful

Derivation rules:

Assign parent, relevance, confidence as you derive each RT — not post-hoc. The scores are part of the reasoning, not decoration on top of it.
Calibrate to what an independent reviewer would assign, not to your hopes. Before writing each edge, ask: "If a skeptical reviewer who has never seen my reasoning re-scored this edge from the outcome and the constraints alone, what number would they pick?" Emit that number. Your score is correct only when it survives that re-scoring — systematic over- or under-confidence is the primary defect this phase is graded on.
Use the full scale; mid-range is the common case. Snap each score to the nearest anchor in the tables above (1.0 / 0.85 / 0.7 / 0.5 / 0.4 / 0.3) rather than rounding everything to 1.0. Reserve confidence = 1.0 for genuinely substitutable-by-nothing, invariant-grade truths; most real prerequisites land at 0.7–0.85 because an alternative path usually exists. A tree whose edges are all 0.9–1.0 is almost always over-confident and will lose calibration points.
Raise the tree's true strength by selecting necessary RTs, not by inflating scores. Two scores are graded together: a map-strength score that rewards edges a skeptical reviewer would also rate high (rel & conf near 1.0), and a calibration score that rewards matching that reviewer. The way to score well on BOTH is to keep the tree to RTs that are genuinely load-bearing — invariant-grade, no-substitute preconditions the outcome literally cannot hold without — and report their high rel/conf honestly. Do not pad the tree with merely-helpful RTs carried at conf 0.4–0.5: each one drags mean(confidence) down and weakens map strength. If an RT is only one helpful option among several, it belongs in solution space, not the truth tree — move it there. The ideal tree is small, fully necessary, and honestly high-scored, not large and defensively low-scored. Inflating a weak edge to 1.0 to chase strength is self-defeating: the reviewer re-scores it low and you lose more on calibration than you gain on strength.
Justify each score in one clause (in the derivation, e.g. "conf 0.7: alternative via queue exists"). An edge you cannot justify in a clause is a guessed score — re-derive it.
Prune low-relevance edges. Any edge with means the RT is not actually required — remove it (it belongs in solution space, not the truth tree).

Self-check before writing (quality gate):

□ Does every RT have a parent, relevance, and confidence?   (else: incomplete map)
□ Is the outcome decomposed into ≥2 top-level RTs?           (else: single-RT trap)
□ Is each RT phrased as a TRUTH that must hold, not a build step?  (else: forward-plan leak)
□ Are confidences spread (not all 1.0)?                      (else: uncalibrated)
□ Does every RT have ≥1 evidence item?                       (else: unverifiable)

A map that passes all five is high-quality; one that fails any is incomplete — fix before continuing.

Recursive Backward Chaining

Default: flat mode (--depth=1). The always-loaded body of this file covers flat-mode elicitation and produces the first-pass required truths.

For multi-level decomposition (when --depth > 1, range 1-4), load references/recursive-decomposition.md. That file covers recursion rules, tree output format, termination conditions, schema additions (depth, children), and variance guardrails (target RT band [10, 28] from the eval baseline).

Binding Constraint (Enhancement 5)

Among RTs with status PARTIAL or NOT_SATISFIED, identify the single binding constraint — the one that is hardest to close, unlocks the others when closed, and blocks all options when unclosed. Surface it as BINDING CONSTRAINT: [RT-ID] [reason] — chain: RT-[n], RT-[n]. Order solution options by how directly they address it; solutions that skip it are deprioritized.

Schema: {"anchors": {"binding_constraint": {"required_truth_id": "RT-3", "reason": "...", "dependency_chain": ["RT-1", "RT-5"]}}}

Handoff to m4: m4-generate reads anchors.binding_constraint, tags artifacts satisfying its RT with "priority": "binding", and generates them first (when context is freshest).

Reversibility Tagging (Enhancement 4)

Tag every solution option: TWO_WAY (reversible), REVERSIBLE_WITH_COST (note the cost), ONE_WAY (flag consequences, require explicit acknowledgment). Surface ONE_WAY options distinctly and list what they close.

Example

/manifold:m3-anchor payment-retry --outcome="95% retry success for transient failures"

OUTCOME ANCHORING: payment-retry

  (edge = parent → child, [rel=relevance, conf=confidence])
RT-1: Can distinguish transient from permanent failures  ← OUTCOME [rel=0.95 conf=1.0]  [gap: no error taxonomy]
RT-2: Retries are idempotent                             ← OUTCOME [rel=1.0  conf=1.0]  [gap: no idempotency]
RT-3: Retry state persists across failures               ← RT-2    [rel=0.85 conf=0.9]  [gap: in-memory only]
RT-4: Downstream services recoverable before budget exhausts ← OUTCOME [rel=0.7 conf=0.7]  [gap: no circuit breaker]
  (note: "client-side exponential backoff" was considered but is one helpful option, not a
   necessary truth — it lives in solution space below, not as an RT. Pruning it keeps the tree
   honestly high-scored rather than padding it with a conf-0.4 edge.)

SOLUTION SPACE:
  Option A: Client-side backoff     — satisfies RT-4      [TWO_WAY]
  Option B: Server-side workflow    — satisfies all       [ONE_WAY ⚠️]
  Option C: Hybrid (client+queue)   — satisfies RT-1/2/3  [TWO_WAY]

RECOMMENDATION: Option C  →  Next: /manifold:m4-generate payment-retry --option=C

Context Lookup

Before anchoring, use WebSearch to research the feature domain (architectural patterns, library versions, external service limits) so required truths and solution options reflect current reality. Summarize findings in a "DOMAIN CONTEXT" block before anchoring. Skip if --skip-lookup is passed or context lookup was done in m1/m2 within the same session.

Execution Instructions

For JSON+Markdown Format (Default)

Run Context Lookup (see above) — research the feature domain via WebSearch unless already done in m1/m2
Read structure from .manifold/<feature>.json
Read content from .manifold/<feature>.md
Get outcome from --outcome flag or Markdown ## Outcome section
For the outcome, recursively ask "What must be TRUE?"
Each truth becomes an RT-N (Required Truth)
Identify gaps between current state and requirement
Generate 2-4 solution options
Recommend best option with rationale
Update TWO files:

.manifold/<feature>.json — Add required truths to anchors.required_truths with id, status, maps_to, parent, relevance, confidence
.manifold/<feature>.md — Add ### RT-1: Title + statement + gap under ## Required Truths

Set phase to ANCHORED in JSON
Run manifold validate <feature> -- fix any errors before showing results. Format lock: if .json exists, never create/update .yaml.

Solution-Tension Validation

After selecting the recommended option, validate it against m2's tensions. For each RESOLVED tension, check whether the option honors the recorded resolution — mark CONFIRMED if yes, REOPENED if no. Record in JSON under anchors.tension_validation (array of {tension_id, status, by_option | reason}). If any tension is REOPENED, surface via AskUserQuestion with options to accept, change option, or modify the option.

Red Flags

Thought	Reality
"I'll forward-plan the steps"	m3 reasons backward: ask what must be TRUE for the outcome.
"One required truth is enough"	Decompose until truths are concrete and verifiable.
"Skip evidence on the truths"	Every required truth needs at least one evidence item.
"A flat list of truths is fine"	The tree is a map: every RT needs a `parent` + `relevance` + `confidence` edge.
"Mark everything confidence 1.0"	Calibrate — over-confident edges are a defect; spread the scores honestly.

Run manifold validate <feature> after updates. Shared directives (output format, interaction rules, validation) injected by phase-commons hook.

name	manifold-m3-anchor
description	Backward reasoning from desired outcome. Derives required conditions by asking 'What must be TRUE?'