decision-logger（决策记录器）

Decision Logger

Two-layer memory system for executive decisions. Layer 1 stores everything discussed. Layer 2 stores only what the founder approved. Future sessions read Layer 2 only -- this prevents hallucinated consensus from past debates bleeding into new deliberations.

Keywords

decision log, memory, approved decisions, action items, board minutes, conflict detection, DO_NOT_RESURFACE, decision history, overdue, supersession, decision search, decision tracking, accountability

Two-Layer Architecture

Why Two Layers?

Single-layer decision logs create a dangerous problem: agents read old debates, rejected proposals, and discarded ideas, then treat them as context for new decisions. This causes "hallucinated consensus" where rejected ideas gradually become accepted through repetition.

The two-layer system prevents this by strictly separating raw discussion from approved decisions.

Layer Architecture

Layer 1: Raw Transcripts (NEVER auto-loaded)
  Location: memory/board-meetings/YYYY-MM-DD-raw.md
  Contains: Full deliberation, all perspectives, rejected arguments
  Loaded: Only on explicit founder request
  Retention: Active 90 days, then archived

Layer 2: Approved Decisions (AUTO-LOADED every session)
  Location: memory/board-meetings/decisions.md
  Contains: Only founder-approved decisions and action items
  Loaded: Automatically at start of every board meeting (Phase 1)
  Mutation: Append-only. Decisions are never deleted, only superseded.

Layer Interaction Rules

Decision Entry Format

Standard Decision Record

## [YYYY-MM-DD] -- [DECISION TITLE]

**Decision:** [One clear statement of what was decided]
**Context:** [1-2 sentences on why this decision was needed]
**Owner:** [One person or role accountable for execution]
**Deadline:** [YYYY-MM-DD]
**Review Date:** [YYYY-MM-DD]
**Confidence:** [High / Medium / Low]
**Rationale:** [Why this option over alternatives, 1-2 sentences]

**User Override:** [If founder changed agent recommendation -- what and why]

**Rejected Alternatives:**
- [Proposal] -- [reason for rejection] [DO_NOT_RESURFACE]
- [Proposal] -- [reason for rejection]

**Action Items:**
- [ ] [Action] -- Owner: [name] -- Due: [YYYY-MM-DD]
- [ ] [Action] -- Owner: [name] -- Due: [YYYY-MM-DD]

**Dependencies:** [Other decisions this depends on]
**Supersedes:** [DATE of previous decision on same topic, if any]
**Superseded by:** [Filled retroactively if overridden later]
**Raw transcript:** memory/board-meetings/[DATE]-raw.md
**Tags:** [topic tags for search -- e.g., pricing, hiring, market-entry]

Completed Action Item Format

- [x] [Action] -- Owner: [name] -- Completed: [YYYY-MM-DD] -- Result: [one sentence]

Conflict Detection System

Before logging any new decision, the system checks for three types of conflicts.

Conflict Type 1: DO_NOT_RESURFACE Violation

A new decision matches a previously rejected proposal.

Detection: New proposal text similarity > 70% to a rejected proposal

Response:
  BLOCKED: "[Proposal]" was rejected on [DATE].
  Reason: [original rejection reason]

  To reopen: Founder must explicitly say "reopen [topic] from [DATE]"
  This cannot be overridden by agents.

Conflict Type 2: Topic Contradiction

Two active decisions on the same topic reach different conclusions.

Detection: Same tags + contradictory conclusions

Response:
  DECISION CONFLICT DETECTED

  Active decision (older): [DATE] -- [decision text]
  New decision: [DATE] -- [decision text]

  These decisions contradict each other.

  Options:
  1. Supersede old decision (new replaces old)
  2. Merge decisions (reconcile the conflict)
  3. Defer to founder (present both, let founder choose)

Conflict Type 3: Owner Conflict

Same action assigned to different people in different decisions.

Detection: Same action description, different owners

Response:
  OWNER CONFLICT

  Action: "[action text]"
  Decision 1 ([DATE]): Owner = [Person A]
  Decision 2 ([DATE]): Owner = [Person B]

  Resolve: Which owner is correct?

Conflict Resolution Decision Tree

START: Conflict detected
  |
  v
[What type of conflict?]
  |
  +-- DO_NOT_RESURFACE --> Block automatically. Only founder can reopen.
  |
  +-- Topic contradiction --> [Is the new decision from a board meeting?]
  |                           |
  |                           +-- YES --> Supersede old by default (board > individual)
  |                           +-- NO  --> Present both to founder for resolution
  |
  +-- Owner conflict --> [Which decision is more recent?]
                         |
                         +-- Flag to founder with both dates
                         +-- Default to more recent unless founder overrides

Decision Lifecycle

States

PROPOSED --> APPROVED --> ACTIVE --> [COMPLETED | SUPERSEDED | EXPIRED]

PROPOSED:    Agent synthesis presented to founder
APPROVED:    Founder explicitly approved
ACTIVE:      Being executed, action items in progress
COMPLETED:   All action items done, review confirmed success
SUPERSEDED:  New decision replaced this one
EXPIRED:     Review date passed without renewal

State Transitions

Logging Workflow

Post-Decision Logging (after Board Meeting Phase 5)

Step 1: Founder approves synthesis
  |
Step 2: Write Layer 1 raw transcript
  --> memory/board-meetings/YYYY-MM-DD-raw.md
  |
Step 3: Run conflict detection against decisions.md
  |
  +-- Conflicts found --> Surface to founder, wait for resolution
  +-- No conflicts --> Continue
  |
Step 4: Append approved entries to decisions.md (Layer 2)
  |
Step 5: Set review dates and action item deadlines
  |
Step 6: Confirm to founder:
  "Logged: [N] decisions, [M] action items tracked, [K] flags added"

Action Item Management

Overdue Detection

At the start of every session, scan for:

1. Action items past their deadline
2. Decisions with review dates that have passed
3. Decisions older than 90 days with no completion status

Alert Format

OVERDUE ITEMS (as of [today's date])

Action Items Past Deadline:
  1. [Action] -- Owner: [name] -- Due: [date] -- [X] days overdue
     From decision: [decision title] ([date])

  2. [Action] -- Owner: [name] -- Due: [date] -- [X] days overdue

Decisions Pending Review:
  1. [Decision title] -- Review was due: [date]
     Original decision: [summary]
     Prompt: "You decided [X] on [date]. Worth a check-in?"

Stale Decisions (> 90 days, no status update):
  1. [Decision title] -- Decided: [date] -- Last update: [date]

Action Item Priority Matrix

Search and Retrieval

Search Capabilities

Decision Summary Views

File Structure

memory/
  board-meetings/
    decisions.md           # Layer 2: append-only, founder-approved
    YYYY-MM-DD-raw.md      # Layer 1: full transcript per meeting
    archive/
      YYYY/                # Raw transcripts after 90 days

Integration with Other Skills

Red Flags

• Same topic discussed 3+ times without a logged decision -- decision avoidance
• Action items consistently overdue by the same owner -- capacity or accountability issue
• Decisions made without checking history -- risk of contradiction
• Layer 1 being loaded without explicit request -- hallucinated consensus risk
• No review dates set on decisions -- decisions age without re-evaluation
• Rejected proposals resurfacing in new language -- DO_NOT_RESURFACE not enforced
• Decision log not consulted at start of board meetings -- institutional memory not used
• All decisions owned by one person -- bottleneck or delegation failure

Proactive Triggers

• Review date passed on a decision -- prompt: "You decided [X] on [date]. Worth checking in?"
• Action item overdue > 7 days -- escalate to founder with owner context
• Same topic area has 3+ active decisions -- consolidation review needed
• 30+ days without any logged decision -- is the system being used?
• New decision proposed that matches DO_NOT_RESURFACE -- block and explain
• Decision from 6+ months ago with no status update -- mark as stale, prompt review

Output Artifacts

Tool Reference

1. decision_tracker.py

Tracks executive decisions with full lifecycle management (Proposed > Approved > Active > Completed/Superseded/Expired). Scans for overdue action items, stale decisions, and generates status summaries.

python scripts/decision_tracker.py --input decisions.json --json
python scripts/decision_tracker.py --input decisions.json

2. decision_quality_scorer.py

Scores decision quality across 6 dimensions: framing (problem definition), alternatives (options considered), information (evidence quality), reasoning (logic soundness), commitment (action clarity), and metacognition (awareness of uncertainty). Generates improvement recommendations.

python scripts/decision_quality_scorer.py --input decision_assessments.json --json
python scripts/decision_quality_scorer.py --input decision_assessments.json

3. decision_tree_builder.py

Builds decision trees with expected value analysis. Calculates optimal paths through probability-weighted outcomes, identifies highest-value decisions, and generates sensitivity analysis on key assumptions.

python scripts/decision_tree_builder.py --input tree_data.json --json
python scripts/decision_tree_builder.py --input tree_data.json

Troubleshooting

Success Criteria

• All board meeting decisions logged in Layer 2 within 24 hours of approval
• Action item completion rate exceeds 80% within stated deadlines
• Zero DO_NOT_RESURFACE violations (rejected proposals do not re-enter decision flow)
• Decision review dates honored for 90%+ of active decisions
• Decision quality score averages above 7/10 across all 6 dimensions
• Conflict detection catches 100% of topic contradictions before new decisions are logged
• Decision log consulted at the start of every board meeting session

Scope & Limitations

In scope: Two-layer decision memory architecture, decision entry and lifecycle management (Proposed > Approved > Active > Completed/Superseded/Expired), conflict detection (DO_NOT_RESURFACE, topic contradiction, owner conflict), action item tracking with overdue alerting, decision search and retrieval by topic/owner/date/status, decision quality scoring, and expected value analysis via decision trees.

Out of scope: CRM or project management tool integration (tools consume JSON exports), meeting transcription or recording, team-level task management (use project-management/ skills), strategic planning or OKR tracking (use strategic-alignment or ceo-advisor), and automated decision-making. This skill tracks and evaluates decisions; it does not make them.

Limitations: Conflict detection uses tag and text matching; semantically similar but differently worded proposals may not be caught. Decision quality scoring is retrospective and depends on honest self-assessment. Decision tree expected value calculations assume probabilities are estimable; highly uncertain environments may make probability assignment misleading. The two-layer architecture requires discipline to maintain; if Layer 2 is not consistently updated, institutional memory degrades.

Integration Points

• chief-of-staff -- Manages the logging workflow; single point of control for Layer 2 writes
• board-meeting -- Triggers decision logging after Phase 5 approval; decision log loaded at Phase 1
• strategic-alignment -- Checks if decisions cascade properly to team goals and OKRs
• executive-mentor -- Reviews stale decisions for re-evaluation; coaches on decision quality improvement
• ceo-advisor -- Strategic decisions logged and tracked; decision patterns inform leadership coaching