ahoward · ahoward · Aug 19, 2026 · Aug 19, 2026 · Aug 19, 2026 · Aug 19, 2026
diff --git a/.specify/memory/constitution.md b/.specify/memory/constitution.md
@@ -1,8 +1,10 @@
 <!--
 Sync Impact Report
 ==================
-Version change: 0.0.0 → 1.0.0
-Modified principles: N/A (initial constitution)
+Version change: 1.0.0 → 1.1.0
+Modified principles:
+  - IV. Antagonistic Testing — antagonist agent changed from Gemini (external CLI)
+    to Fable (Claude model, run as a subagent). Role and rules unchanged.
 Added sections:
   - Core Principles (6 principles)
   - Architecture Constraints
@@ -65,11 +67,11 @@ Every `sys.call` MUST return the same shape:
 
 ### IV. Antagonistic Testing
 
-Tests are specifications. Claude designs, Gemini challenges, then implement.
+Tests are specifications. Claude designs, Fable challenges, then implement.
 
 - Claude (primary agent) designs first pass of tc tests
-- Gemini (antagonist) reviews tests, finds blind spots, suggests harder cases
-- After Gemini review, tests are LOCKED
+- Fable (antagonist) reviews tests, finds blind spots, suggests harder cases
+- After the antagonist review, tests are LOCKED
 - Human checkpoint ONLY when stuck (cannot make tests pass)
 - Tests MUST exist before implementation
 
@@ -130,7 +132,7 @@ Brane uses two databases, each optimized for its domain:
 
 1. **Design Interface** — Define sys.call paths (public API only)
 2. **Design Tests (Claude)** — Write tc test cases with input.json/expected.json
-3. **Review Tests (Gemini)** — Antagonist review, incorporate suggestions
+3. **Review Tests (Fable)** — Antagonist review, incorporate suggestions
 4. **Implement** — Write code (handlers thin, logic functional)
 5. **Loop Until Green** — Fix failures, re-run tests
 6. **If Stuck → Human Checkpoint** — Only when tests cannot pass
@@ -152,7 +154,7 @@ tests/{handler}/
 ### Never
 
 - Implement without tests
-- Skip Gemini review
+- Skip the antagonist review
 - Change tests after review without human approval
 - Move to next task with failing tests
 - Use classes for data structures
@@ -181,4 +183,4 @@ Constitution follows semantic versioning:
 - Complexity MUST be justified against Principle VI
 - Violations require explicit justification and human approval
 
-**Version**: 1.0.0 | **Ratified**: 2026-01-27 | **Last Amended**: 2026-01-27
+**Version**: 1.1.0 | **Ratified**: 2026-01-27 | **Last Amended**: 2026-08-19
diff --git a/CLAUDE.md b/CLAUDE.md
@@ -26,6 +26,7 @@ const result = await sys.call("/namespace/method", data)
 
 - `./dna/product/ROADMAP.md` - **START HERE** — Driving task list for all development
 - `./dna/product/prd.md` - Full PRD (Split-Brain architecture)
+- `./dna/product/vision-spec-machine.md` - **Vision v4.0** — brane as the regenerative specification substrate (#112)
 - `./.specify/memory/constitution.md` - Project principles (6 core rules)
 - `./dna/technical/development-loop.md` - Antagonistic Testing process
 - `./ai/MEMORY.md` - AI's long-term memory
@@ -35,10 +36,11 @@ const result = await sys.call("/namespace/method", data)
 
 **See:** `./dna/technical/development-loop.md` and `.specify/memory/constitution.md`
 
-1. Design interface → 2. Design tests (Claude) → 3. Review tests (Gemini)
+1. Design interface → 2. Design tests (Claude) → 3. Review tests (Fable)
 4. Implement → 5. Loop until green → 6. **⛔ HUMAN CHECKPOINT** (only if stuck)
 
-**Gemini** = antagonist agent. Reviews tests, finds blind spots.
+**Fable** = antagonist agent. Reviews tests, finds blind spots. Run as a subagent
+(Agent tool, `model: fable`), not an external CLI.
 Human checkpoint is for failure resolution, not pre-approval.
 
 ## Coding Conventions Summary
@@ -79,7 +81,7 @@ ai/                   # AI agent resources
 2. Run `/speckit.specify` — creates `specs/{feature}/spec.md`
 3. Open PR for human review
 4. After approval: `/speckit.plan` → `/speckit.tasks`
-5. Review tests with Gemini (antagonist)
+5. Review tests with Fable (antagonist)
 6. Implement via `/speckit.implement`
 7. If stuck (tests won't pass) → Human checkpoint
 8. On completion → Update ROADMAP.md, mark feature complete
@@ -91,7 +93,7 @@ ai/                   # AI agent resources
 - Return different shapes from handlers
 - Skip the Result envelope
 - Implement without tests
-- Skip Gemini review
+- Skip the antagonist review
 - Change tests after review without human approval
 
 ## Active Technologies
@@ -126,6 +128,7 @@ ai/                   # AI agent resources
 - CozoDB mind.db (RocksDB backend) — new `episodes` relation + HNSW index (034-episodic-memory)
 
 ## Recent Changes
+- 067-claim-authority: First-class claims carrying authority tier + source; contradiction representable as data (#113). Schema v1.13.0, 8 sys.call paths, `contradictions` built-in rule, `brane claim` / `brane authority` CLI
 - 021-vector-search: Added semantic search via `/mind/search` endpoint with local embeddings (fastembed-js BGESmallEN, 384 dims)
 - 016-rules-define: Added TypeScript (Bun 1.x) + CozoDB (Datalog), existing mind.ts utilities
 
@@ -164,3 +167,54 @@ echo '{"query": "authentication", "limit": 5}' | bun run src/cli.ts /mind/search
   }
 }
 ```
+
+
+## Spec Machine Reframe (#112)
+
+**See:** `dna/product/vision-spec-machine.md` (v4.0)
+
+Brane is being repositioned from "memory for agents" to **the regenerative specification substrate** —
+the durable, queryable, provenance-backed graph that implementations are generated FROM and validated
+AGAINST. The existing graph, provenance, Datalog rules, lenses, and extraction pipeline are ~70% of it.
+
+Four gaps close the rest:
+
+| Gap | Feature | Issue | Status |
+|---|---|---|---|
+| Claim + authority model (contradiction as data) | `067-claim-authority` | [#113](https://github.com/ahoward/brane/issues/113) | ✅ Done (PR #117) |
+| Observation → requirement promotion gate | `068-promotion-gate` | [#114](https://github.com/ahoward/brane/issues/114) | Next — #113 is done, so it is unblocked |
+| Regeneration → test → feedback (**keystone**) | `069-regeneration-spike` | [#115](https://github.com/ahoward/brane/issues/115) | Unblocked; not started |
+| Production-as-teacher ingestion | `070-production-teacher` | [#116](https://github.com/ahoward/brane/issues/116) | Unblocked; not started |
+
+Two principles govern the claim work and should be preserved by anything built on it:
+
+1. **Strict about authority, loose about vocabulary.** Authority tiers are registered and ranked;
+   predicates and assertions are never validated against a vocabulary.
+2. **Contradiction is data, not a defect.** Competing claims coexist. Resolution is a read-time
+   projection by authority rank — never a write-time deletion. Ties do not resolve.
+
+### Claims (067-claim-authority)
+
+```bash
+brane claim create --concept 1 --predicate refund_window \
+  --assertion "30 days" --authority product --source dna/product/prd.md
+brane claim conflicts          # where the graph contradicts itself
+brane claim list --concept 1 --resolve   # one answer, losers still stored
+brane authority list           # observation < implementation < product < legal < manual
+brane verify --rule contradictions
+```
+
+Invariants anything built on claims must preserve:
+
+- Authority tiers are registered and ranked (**strict**); predicates and assertions are never
+  validated against a vocabulary (**loose**).
+- Rank is joined at read time; claims store the tier *name* only, so re-ranking never rewrites history.
+- Resolution is a read-time projection. Ties at the top rank do **not** resolve.
+- Claims are immutable — correction is delete + re-assert.
+- `cascade_claims()` in `src/lib/claims.ts` is the single deletion seam. Every path that removes a
+  concept or edge (delete handlers, `prune`, re-extraction) calls it.
+- The `contradictions` rule body positionally matches the 8-column `claims` relation. Adding a column
+  (#114's binding flag) means updating that body in the same migration.
+
+**Cozo gotcha:** string literals use backslash escapes, not SQL doubling. Use `esc_cozo()` from
+`src/lib/mind.ts`. `'it''s'` is a parse error.
diff --git a/README.md b/README.md
@@ -229,6 +229,7 @@ bun run repl          # interactive mode
 ## more
 
 - [dna/product/ROADMAP.md](dna/product/ROADMAP.md) — what's next
+- [dna/product/vision-spec-machine.md](dna/product/vision-spec-machine.md) — where this is going: brane as the regenerative specification substrate
 - [dna/product/prd.md](dna/product/prd.md) — full PRD
 - [.specify/memory/constitution.md](.specify/memory/constitution.md) — project principles
 

diff --git a/ai/MEMORY.md b/ai/MEMORY.md
@@ -133,7 +133,7 @@ We aim for Unix-clean, not JS-clean:
 
 **Antagonistic Testing** - Having one agent suggest tests for the other to find blind spots.
 
-The primary agent (Claude) has the most context, so it designs the first pass. The review agent (Gemini) plays adversary - finding edge cases, questioning assumptions, suggesting harder tests.
+The primary agent (Claude) has the most context, so it designs the first pass. The review agent (Fable) plays adversary - finding edge cases, questioning assumptions, suggesting harder tests.
 
 This creates tension that produces better specifications.
 
@@ -142,38 +142,44 @@ This creates tension that produces better specifications.
 | Agent | Platform | Role |
 |-------|----------|------|
 | **Claude** | Claude Code | Primary. Has context. Designs first pass of tests. Implements. |
-| **Gemini** | Google Gemini CLI | Antagonist. Reviews tests. Suggests harder cases. Finds blind spots. |
+| **Fable** | Claude Code subagent (`model: fable`) | Antagonist. Reviews tests. Suggests harder cases. Finds blind spots. |
 | **Human** | - | Checkpoint authority. Final say on test boundaries. |
 
-### Running Gemini
+### Running the Antagonist (Fable)
 
-Claude runs Gemini directly via CLI with `--resume` for session continuity:
+**Changed 2026-08-19.** The antagonist was Google Gemini via CLI. It is now **Fable**, run as a
+Claude Code subagent. Constitution 1.1.0 records the amendment. `gemini` is not installed and is no
+longer part of the loop.
 
-```bash
-gemini --resume latest "review prompt here"
+```
+Agent tool → subagent_type: "general-purpose", model: "fable"
 ```
 
-**Context to include in review prompts:**
-- Relevant conventions from `ai/MEMORY.md`
-- The specific test files being reviewed
-- What kind of feedback is wanted (edge cases, consistency, etc.)
+**What to put in the review prompt:**
+- Point it at the actual files (spec, tests, fixtures) - it reads the repo itself
+- Name the conventions it should check against (`.specify/memory/constitution.md`, `ai/MEMORY.md`)
+- Name the existing code it should check the design *fits* (handlers, migrations, tc fixtures)
+- Say what kind of feedback is wanted, and tell it to be genuinely adversarial rather than agreeable
+- Ask for severity-ranked findings with concrete fixes, and a ready/not-ready verdict
 
-This allows Gemini to build memory of project conventions across reviews.
+**Always verify its findings against the tree before acting.** It is a reviewer, not an oracle. In
+practice its blockers have been real, but the cost of checking is one grep.
 
-### Gemini (Antagonist) - IMPORTANT
+### Fable (Antagonist) - IMPORTANT
 
-- **Platform:** Google Gemini CLI
+- **Platform:** Claude Code subagent, `model: fable`
 - **Role:** Test reviewer, antagonist
 - **When to use:** After designing tc tests, BEFORE implementing
-- **How to run:** `gemini --continue` (Claude runs this directly)
+- **Also useful:** reviewing a spec/plan package before implementation starts
 
-Gemini reviews test designs and suggests:
+Fable reviews test designs and suggests:
 - Missing edge cases
 - Interface inconsistencies
 - Better error scenarios
 - Clearer expected outputs
+- Fixtures the change breaks that nobody listed
 
-**Never proceed past test boundary without Gemini's review.**
+**Never proceed past the test boundary without the antagonist review.**
 
 ---
 
@@ -182,7 +188,7 @@ Gemini reviews test designs and suggests:
 See: `dna/technical/development-loop.md`
 
 ```
-Design Interface → Design Tests (Claude) → Review Tests (Gemini)
+Design Interface → Design Tests (Claude) → Review Tests (Fable)
     → Implement → Loop Until Green → ⛔ HUMAN CHECKPOINT (only if stuck)
 ```
 
@@ -191,7 +197,7 @@ Design Interface → Design Tests (Claude) → Review Tests (Gemini)
 The human checkpoint is for **failure resolution**, not pre-approval:
 
 1. Design tests (Claude)
-2. Review tests (Gemini antagonist)
+2. Review tests (Fable antagonist)
 3. Implement code
 4. Run tests, loop until green
 5. **ONLY IF STUCK** (can't make tests pass) → Human decides: fix tests or fix code
@@ -200,8 +206,8 @@ Tests are the spec. Claude implements against them. Human intervenes only when t
 
 ### Never
 
-- Alter tests after Gemini review without human approval
-- Skip Gemini review
+- Alter tests after the antagonist review without human approval
+- Skip the antagonist review
 - Move to next task with failing tests
 - Treat test failure as "change the tests" (ask human first)
 
@@ -228,7 +234,7 @@ Tests are the spec. Claude implements against them. Human intervenes only when t
 
 1. Design sys.call interface
 2. Write tc tests (input.json, expected.json)
-3. Send to Ali (Gemini) for review
+3. Send to the antagonist (Fable) for review
 4. Incorporate feedback
 5. Create skip tests
 6. **STOP** - Wait for human
@@ -314,6 +320,45 @@ const native = require('./native/6/cozo_node_prebuilt.node');  // static
 
 ---
 
+### Spec Machine Reframe (2026-08-19)
+
+**Decision:** brane is repositioned from "memory for agents" to **the regenerative specification
+substrate** (Chad Fowler, "The Specification Is Not a Document"). Vision doc:
+`dna/product/vision-spec-machine.md` (v4.0). Umbrella issue #112.
+
+The existing substrate — graph, provenance, Datalog rules, lenses, extraction pipeline — is ~70% of
+Fowler's spec machine. Four gaps close it: claims+authority (#113), promotion gate (#114), regeneration
+(#115, keystone + research risk), production-as-teacher (#116).
+
+**Invariants established in `067-claim-authority` that later work must not break:**
+
+- Authority tiers are registered and ranked (strict); predicates and assertions are free-form (loose).
+- Competing claims coexist. Resolution is a **read-time projection**, never a write-time deletion.
+- Authority rank is joined at read time — claims store the tier *name* only, so re-ranking never
+  rewrites history.
+- Ties at the top rank do **not** resolve. `resolution: null`, `unresolved: true`. No tiebreak by
+  recency or ID.
+- Claims are immutable. Correction = delete + re-assert.
+- Conflict comparison is trim-then-exact and **case-sensitive** — case folding would silently hide
+  conflicts, which is the failure this feature exists to prevent.
+
+**Gotchas found while implementing #113:**
+
+- The built-in `contradictions` Datalog rule positionally matches the 8-column `claims` arity. Any
+  feature adding a claims column (e.g. #114's binding flag) must update the rule body in the same
+  migration.
+- **CozoDB string literals use backslash escapes, not SQL doubling.** `'it''s'` is a parse error;
+  `'it\\'s'` is correct. `src/handlers/mind/rules/create.ts` had been doing it the SQL way since it
+  was written — any rule body containing an apostrophe silently failed to save. Fixed via
+  `esc_cozo()` in `src/lib/mind.ts`; use it for every interpolated string. Query parameters (`$val`)
+  also work and avoid the question entirely.
+- `/mind/provenance/create` validates the file against body.db, so tests that need *stale* provenance
+  must add the body row, create the link, then delete the row (see `tests/mind/prune/run`).
+- `prune` only removes concepts whose provenance is **entirely** stale. A concept with no provenance
+  at all is never an orphan and is never pruned.
+
+---
+
 ## Version History
 
 | Date | Change |
@@ -322,3 +367,6 @@ const native = require('./native/6/cozo_node_prebuilt.node');  // static
 | 2026-01-26 | Added multi-agent system (Ali/Gemini). Development loop defined. |
 | 2026-01-28 | CozoDB decision: Keep it, fix bundler via PR to upstream. |
 | 2026-01-28 | Multi-platform compilation working. src/lib/cozo.ts + CI workflow. |
+| 2026-08-19 | Spec machine reframe (#112). vision-spec-machine.md v4.0. 067-claim-authority spec + plan (PR #117). |
+| 2026-08-19 | Antagonist changed from Gemini CLI to Fable subagent. Constitution 1.0.0 → 1.1.0. |
+| 2026-08-21 | 067-claim-authority implemented. mind.db schema v1.13.0. Cozo escaping bug found and fixed. |
diff --git a/dna/product/ROADMAP.md b/dna/product/ROADMAP.md
@@ -12,7 +12,7 @@ specs/{NN-feature}/spec.md      ← /speckit.specify → PR for human review
 specs/{NN-feature}/plan.md      ← /speckit.plan
 specs/{NN-feature}/tasks.md     ← /speckit.tasks
     ↓
-Gemini review                   ← Antagonistic test review (tc tests)
+Fable review                    ← Antagonistic test review (tc tests)
     ↓
 Implementation                  ← /speckit.implement
     ↓
@@ -91,6 +91,15 @@ The hippocampus: 3 verbs, dual-write graph + audit trail, trust tiers. The agent
 - [x] `108-graph-backed-recall` — Auto-connect on remember, graph context in recall ([#108](https://github.com/ahoward/brane/issues/108))
 - [x] `109-append-only-consolidation` — New concepts + CAUSED_BY edges, never mutate sources ([#109](https://github.com/ahoward/brane/issues/109))
 
+### Spec Machine — brane as the regenerative specification substrate ([#112](https://github.com/ahoward/brane/issues/112))
+
+Vision: `dna/product/vision-spec-machine.md` (v4.0). Closing the 30% gap between "memory for agents" and Fowler's spec machine.
+
+- [x] `067-claim-authority` — First-class claims carrying authority tier + source; contradiction as data ([#113](https://github.com/ahoward/brane/issues/113), [PR #117](https://github.com/ahoward/brane/pull/117))
+- [ ] `068-promotion-gate` — Observation → requirement promotion with human ratification ([#114](https://github.com/ahoward/brane/issues/114))
+- [ ] `069-regeneration-spike` — **Keystone**: regenerate from graph → test → feed failures back ([#115](https://github.com/ahoward/brane/issues/115))
+- [ ] `070-production-teacher` — Ingest prod traces / incidents / evals as evidence ([#116](https://github.com/ahoward/brane/issues/116))
+
 ### Deferred
 
 - [ ] `028-verifier-node` — Headless verification node
@@ -104,6 +113,7 @@ The hippocampus: 3 verbs, dual-write graph + audit trail, trust tiers. The agent
 
 | Feature | PR | Date |
 |---------|-----|------|
+| `067-claim-authority` | #117 | 2026-08-21 |
 | `102-109-hippocampus-v2` | #110 | 2026-05-30 |
 | `058-status-dashboard` | #81 | 2026-03-27 |
 | `057-cli-ingest-sessions` | #80 | 2026-03-27 |
@@ -185,3 +195,4 @@ The hippocampus: 3 verbs, dual-write graph + audit trail, trust tiers. The agent
 - **032-corpus complete!** — Adversarial test material: surveillance, dark patterns, COPPA violations, alignment-washing
 - **049-schema-migrations complete!** — Versioned mind.db upgrades with backup/restore, Gemini-reviewed
 - **053-fuzzy-dedup complete!** — Length-scaled Levenshtein prevents false matches on short concept names
+- **067-claim-authority complete!** — Claims carry authority tier + source; contradiction is data. Schema v1.13.0, `contradictions` built-in rule, Fable-reviewed (2 rounds, 3 blockers), 432 tests