Shared memory and context tools for agentic work.
Code Rooms
# THE m1nd ORGANISM — the constitution
> **Mandate:** close the PRD + UML for every system and subsystem, so it can be implemented
> cleanly — no spaghetti code — connecting with what has already been mapped, but going deep first
> to find still more useful connections that can genuinely close this loop.
>
> **2026-07-05. Inputs adjudicated as law-drafts:** the Organism
> Map (joint-mapper, 2026-07-05 — subsystem inventory, JOINTs A–K, contradictions, homeless list,
> dependency reality) and the Adversarial Critique (26-finding ledger; 7 blockers). **Every finding
> is adopted or refused with reasons — Appendix ADJ.** Ground: repo @ `main` `5b1a37d` (MEDULLA PRD
> merged, #267); tool census **112** re-measured this session from `server.rs` (name-array, unique);
> the served owner probed this session (binary 1.3.2) — the probe itself reproduced two of the
> critique's live defects (F10 duplicate serialization; the MED-INV-6 false absence, third witness).
> **SOUL-PRD: absent from main at write time** (re-checked); this constitution writes its SLOT and
> binding constraints (§C8), never its internals.
> **Status of this document: SOURCE OF LAW.** The feeder PRDs (NEXTGEN, TWO-TIER, MEDULLA,
> HUMAN-LAYER, FOCUS, X360, HOST-MATRIX) remain the mechanism specs — richer, older, and each
> internally honest — but where they disagree with each other or with this document, **this document
> wins**, and each carries an amendment pointer saying so (§C11). This is the last blueprint of the
> design era; everything after it climbs the ladder (§C10).
---
## C0 — How to read this constitution (the documentary order)
**Truth vs law, separated once.** For **facts**, the shop hierarchy is unchanged: code + git +
runtime > PATHOS > any document, including this one. For **design law** — which grammar, which
verb, which order — this document is the apex; feeder PRDs are mechanism.
The documentary laws (adopting F22, all four prescriptions):
1. **One load-bearing map.** This document is the organism's map. Feeder PRDs are demoted to
feeder specs: their "current-state" sections are non-authoritative (PATHOS + code own current
state); their mechanism sections stay canonical for their subsystem until superseded here.
2. **Anchor policy: the symbol is the contract.** Durable docs cite symbols
(`light_author_handlers::handle_light_author`, `project_brains.rs`, `classify_edge`) — line
numbers are hints that rot in days and are never load-bearing. This document carries no line
numbers by design.
3. **Cross-PRD references land in the same PR or carry an explicit branch stamp.** The agent-docs
CI gate (#229 family) may grep for `§`-refs into sections absent on the merge target. (The one
live instance of this failure class — MEDULLA §9.2 ↔ HUMAN §4A.11 citing each other across
branches — self-resolved when #267 landed both; the policy remains law because the class will
recur.)
4. **Censuses are re-measured at landing.** Any doc that states a count (tools, letters, claims,
nodes) states it *as measured at a named pin*, and the battery pins the counts that are law
(§C6, §C1.5).
**The organism in one diagram** — the spine (one packet out), the write door (one path in), the
four grammars as the only vocabularies, and every subsystem placed:
```mermaid
graph TD
subgraph READERS["THE SPINE — one packet, five readers (C1)"]
AG[agent · north]
PF[human · Pre-Flight Card]
DC[child · delegation packet]
RC[first contact · reception block]
SH[soul headline · reserved slot]
end
PKT[["m1nd-north-packet — the atom<br/>binding · context · memory · sufficiency · honest_gaps · non_claims<br/>BUDGET-BOUND (C1)"]]
PKT --> AG & PF & DC & RC & SH
subgraph STORES["the stores (belief lifecycle — grammar 2)"]
PB[(project brains<br/>project_private)]
MED[(medulla<br/>promoted · doctrine-born)]
COMPOSE[packet composer<br/>project beat + medulla beat<br/>pull-not-push] --> PKT
RERANK[seek/focus rerank<br/>handle_seek · trust × tremor × conformance] --> COMPOSE
PB --> RERANK
MED --> RERANK
DOOR{{"memorize — THE WRITE DOOR (C4)<br/>supersession gate · .history/ · flock ·<br/>provenance stamps · hygiene floor ·<br/>letters-never-evidence refusal"}}
DOOR --> PB
DOOR --> MED
STOP[Stop-hook distiller<br/>ambient, every turn] --> DOOR
MC[mission_close<br/>write_light_memory] --> DOOR
DB[debrief step 6] --> DOOR
PRV[promote — audited crossing] --> DOOR
RL[reception policy-learning] --> DOOR
MAN[manual memorize<br/>the one agent ritual C3] --> DOOR
SOUL[soul_update — future<br/>MUST route here] -.-> DOOR
subgraph TEL["telemetry (letter fates — grammar 3)"]
SPOOL[/global spool<br/>field-reports.jsonl/]
BOX[/per-project boxes<br/>.m1nd inbox/]
SPOOL --> BOX
TEL -.->|honesty class =<br/>calibration ground truth| CAL[calibration<br/>trust ladder — grammar 1]
XRAY[X-RAY conformance<br/>grammar 4] -.->|conformance_boost<br/>BEDROCK +0.20 · EROSION −0.30| RERANK
CAL -.->|verdicts on every answer| PKT
```
## C1 — The Spine: one packet, N readers (JOINT-E · F10 · F17)
### C1.1 The atom
The organism has **one packet**: the north packet (`m1nd-north-packet-v0`). Its shared sub-atoms —
`binding` (trust_mode + fingerprint), `memory[]` (claim + real age + author, absent-never-faked),
`sufficiency`, `honest_gaps`, `non_claims` — are **one contract with N views**. The schemas that
exist today (`m1nd-reception-v0`/`-degraded-v0`, `m1nd-delegation-packet-v0`,
`m1nd-project-brain-bootstrap-v0`) are renderings and sub-shapes of this atom, not siblings: they
version their *view*, never a divergent copy of the sub-atoms. A new field enters the atom once and
appears in views by selection — never invented per-view. (Map contradiction #3, collapsed.)
### C1.2 The five readers — what each MAY add, what each may NEVER add
| Reader | Rendering | MAY add | NEVER adds |
|---|---|---|---|
| **Agent** | `north` result | `next_move`, `recovery_playbook`, verb-level `_m1nd` hints | a fabricated age or presence (absent-never-faked); another brain's memory unlabeled (MED-INV-1) |
| **Human** | the Pre-Flight Card (HUMAN §4.2) | rendering only: action language, fate-lines, calm badges | **data absent from the packet** — a card field with no packet field is fabrication; the card renders, never widens |
| **Delegation child** | the §O.12.4 packet | the orchestrator's brief (always the head), scope + duties, respawn blocks | unlabeled cross-brain memory (M7 tier labels mandatory); quoted code bodies (stale shadows); dropping the duties section |
| **First contact** | the `reception` block | `options[]`, each with the exact call + consequence | an interruption on match (TT-INV-12: silence on match); an auto-bind on mismatch (the caller-root-mismatch rule) |
| **Soul headline** | reserved slot (§C8.6) | **ONE line** — "what this brain IS" — plus the pull verb | a second line; any claim that bypasses grammar 2 or the two-tissues law |
The reception match verdict additionally echoes on the Hall's Brain Chip and the Card's binding
header (JOINT-I) — sub-renderings of reader 4, same verdict, zero new data.
### C1.3 THE PACKET BUDGET LAW (adopting F10, numbers binding)
Pull-not-push graduates from a memory law (MEDULLA law 2) to **the packet law**:
1. **Size:** north ≤ **2,000 tokens** on the MCP path (the organism's own budget doctrine, §O.12.4:
default 2k, hard ceiling 8k, applied to its flagship packet); ≤ **1,200 chars** on the hook path
(TT §10.1, restated as binding on the packet composer, not just the shell script).
2. **Sufficiency-gating with recorded drops:** omitted sections drop out of the JSON and each drop
adds a `non_claims` line (§O.12.4's law, generalized; north already carries `non_claims` — wire it).
3. **The satellite one-line rule:** reception (mismatch only), inbox, soul, doctrine each enter
north as **at most one line: a count/headline + the exact pull verb**
(`inbox: 5 abertas, 2 OLD — inbox_sweep to read`). Anything longer is one verb away.
4. **No duplicate serialization:** `fingerprint`/`graph_state` deduplicate (roots once, or count +
dirs). **The write-path fix is mandatory:** `memorize` stops registering every claim file as a
standalone ingest root — per-file roots collapse into the store-dir root **at write time**
(M5a's migration sweep fixes the stock; without the flow fix the sprawl regrows immediately).
*Measured live this session: the two root arrays byte-identical × 2, 21 entries, 20 of them
individual `.light.md` files.*
5. **The budget is a gate, not prose:** a battery case pins packet size on a reference graph;
growth beyond budget fails CI, forcing rule 3 on whoever adds the next block.
### C1.4 The two trust surfaces stay visually distinct (F17 rider)
`trust_mode` (binding trust: "am I talking to the right brain") and the calibration cap (verdict
trust: "may answers say `act` here") are different claims and render differently in every view.
On an uncalibrated brain the packet **says the cap in words**: *"not measured on this repo yet —
one `calibrate_predict` arms it."* The clone-day demo carries its own honest footnote.
## C2 — The Four Grammars Law (JOINT-B · F5 · F1 · F3 · F4)
### C2.1 The law
The organism speaks **exactly four state grammars**. Every state word in any surface, doc, or
schema maps onto one of the four or is killed in review.
| # | Grammar | Domain | The words |
| **1** | **Trust ladder** | answers (how much to rely) | `act \| reverify \| abstain \| unprovable` |
| **2** | **Belief lifecycle** | stored claims (where a belief lives) | storage: `project_private → promoted → superseded` (+ birth exception `doctrine-born`); overlay: `aged` (computed, never stored); file rendering: `State: authored \| verified \| outdated` |
| **3** | **Letter fates** | telemetry (what happened to a signal) | `wet_ink \| in_flight \| fired_clay \| external` — **derived from the reply graph, never stored**; receipt disposition: `fixed \| declined \| moot`; grouping overlay: `case` |
| **4** | **X-RAY conformance** | architecture (intent vs reality) | `BEDROCK \| BLUEPRINT \| EROSION \| OVERGROWTH \| UNPROVABLE{reason}` |
**Grammar 1 is universal on answers:** envelope, `predict`, `seek`, `xray_gate`, the
delegation-abstain, and `underwrite` all speak it. **`underwrite`'s verdicts are renamed onto the
ladder** (`proceed` → `act`, `require-human` → `reverify`) — it was the same ladder wearing
synonyms (F5.1 adopted).
**Record fields are not grammars.** Mission/delegation record states (`live → debriefed`,
`success|failure|partial`, `outcome_unverified`) and process states (`live|dormant|warming` +
reception verdicts) are ledger-local fields. Any attempt to grow them into an organism-wide fifth
grammar dies at review.
### C2.2 The letter-grammar adjudication (F1 — blocker, resolved)
Three letter grammars were in flight. **The derived grammar wins** (MEDULLA §9.2, the newest
approval, carrying the same law that made `aged` an overlay: *state that can be recomputed
must not be stored where it can drift*). The reconciliations, explicit:
- **Mapping from the sealed doctrine's grammar:** `open → wet_ink` · `triaged(resolves:<id>) →
in_flight` · `resolved → fired_clay`. The `resolves:` stored transition is repealed — the reply
graph (`answers[]`) derives it.
- **The two missing terminals — mapped, not minted.** `rejected` and `expired` do NOT become
fates. The fate axis stays closed at four, derived-only; what they carried moves into the
**receipt disposition** — content of the closing receipt letter, not state of the target:
a receipt that answers-and-declines is `fired_clay` with `disposition: declined`; a moot-check
receipt is `fired_clay` with `disposition: moot`; a fixing receipt is `disposition: fixed`.
One derived axis, zero information lost. *(F1 prescribed "add the terminals or kill them with
reasons" — this is the kill-with-reasons: a terminal that requires storing state on the target
letter would break the derived-only law the winning grammar rests on.)*
- **Case intelligence is a grouping overlay, never a third axis:** a `case` is a set of letter ids
plus one receipt (`case_id`/`fingerprint` fields), layered over fates (→ the Case-Intelligence
PRD, §C10 R11).
- **ORCHESTRATOR ACTION NOTE (owed at M7b's landing, same session):** the sealed
`project-inbox-doctrine` medulla memory is `State: verified` and carries the losing grammar —
it must be **superseded via `memorize`** (same label, the fate grammar, pointer to this section)
when M7b lands. This is a runtime-memory act by the landing orchestrator, not a repo file in any
PR — recorded here so it cannot be forgotten and cannot be recalled as live doctrine afterward.
### C2.3 Vocabulary annexes (enums that mint no synonyms)
- **The stop vocabulary (F3):** ONE enum answers "should I stop gathering?" —
`gathering | sufficient | saturated`, grown exactly two verdicts (`not_here`, `looping`) when the
stop-check rides `focus(mode:"check")` (§C6). `stop_gate` mints no words; Subsystem C's
`BLOCKED` re-definition folds into this enum or dies.
- **The age law (F4):** ONE named constant — `MEMORY_AGE_HALF_LIFE` = 30 d = 720 h — computed at
ONE site. `stale` (north) and `aged_out` (`cross_verify`) are two *renderings* of that one law,
listed as such in every vocabulary table; MEDULLA §3.1's "stale >30d AND/OR aged_out >720h"
reads as one rule, two faces.
### C2.4 Refusals made law (the false unifications, forbidden forever)
1. **Letter age ≠ memory age — NEVER unify.** The semantic inversion is load-bearing: a letter's
age measures **triage neglect** (attention debt — letters never decay, never auto-delete); a
claim's age measures **truth decay**. Folding letters into L1GHT states (supersession,
confidence decay) would destroy the wet-ink/fired-clay distinction the telemetry loop rests on.
2. **X-RAY ≠ belief lifecycle.** `EROSION` is not `superseded`; `BLUEPRINT` is not `wet_ink`.
Code-vs-intent and belief-storage share exactly ONE rung — the honest bottom
(`UNPROVABLE`/`abstain`) — and that is the only joint to standardize.
3. **Belief lifecycle ⊥ trust ladder.** A `promoted` claim can be `abstain`-grade; a
`project_private` claim can be `act`-grade. Storage never couples to calibration.
4. **The Soul gets NO fifth grammar** *(binding constraint on the in-flight SOUL-PRD)*: a soul
claim is a grammar-2 belief whose answers wear grammar 1. §C8.6 carries the full slot.
**The shared deep laws** (why four grammars rhyme without merging): derived-not-stored
freshness/fate; an immutable terminal that keeps history (`.history/`, `fired_clay`, the append-only
ledgers); and a first-class "I cannot decide" (`absent`, `UNPROVABLE`, `insufficient_evidence`,
`external`). The organism refuses to fake certainty — that refusal is the constitution's spine.
## C3 — The Moment Map (JOINT-A · F6)
### C3.1 The collapse
The corpus accreted nine-plus closing rituals; an agent runs three, non-deterministically. The
census collapses to:
> **The agent remembers ONE ritual; the machine owns five; the repo owns the gate.**
- **THE agent ritual — "leave it warmer"** (session close, one composite motion):
`memorize` durable claims + `learn(correct|wrong|partial)` retrieval grades + **one letter only
if m1nd misbehaved**. `mission_close(write_light_memory: true)` **is this same motion in its
with-mission form** — never an additional ritual. (`M1ND_INSTRUCTIONS` §3 already teaches
exactly these three acts; the owed edit is the one sentence naming them ONE motion — rides the
next instructions-touching slice under the era-coherence gate.)
- **THE mechanical gate — the landing gate** (PR/CI): battery green + doc-gate + agent-docs gate +
PATHOS at checkpoints. Already mechanized (#229 family); the only moment with a hard enforcement
surface. Everything repo-truth rides it; nothing session-scoped may hide behind it.
- **Machine-owned, never rituals:** `persist`/snapshot (daemon + SIGTERM/idle-exit) ·
`trail_save` (PreCompact hook) · the distillation gate (Stop hook — §C3.2) · SessionEnd beats
(persist + boot_memory + alerts_ack) · debrief-capture (SubagentStop wave) · consolidation
(daemon cadence). If a closing act fits neither the ritual nor the landing gate, it is wired
into a hook/daemon or killed.
- **`debrief` is the orchestrator's landing gate for a spawn** — it joins the gate conceptually,
not the ritual list. Until SubagentStop mechanizes it, the packet's report protocol
(`[m1nd dlg_…]` + DEVIATIONS/FINDINGS) is the manual half.
- **Promotion is NEVER a closing moment.** It is curation on the medulla's cadence (consolidation
pass / maintainer review). No instructions text may ever add "consider promoting" to session close.
### C3.2 The four "closes", disambiguated (the JOINT-A trap, defused)
| The word says | The machine is | Fires |
|---|---|---|
| `mission_close` | an id-bearing verb: load mission, verify claims, emit handoff, optionally persist | rarely — only on genuinely-open missions |
| "the distillation gate" | the ambient `Stop → cross_verify → memorize` hook; keystone is **`memorize`** (free-form), *provably not* `mission_close` (mission verbs hard-error without `mission_id`; Stop fires every turn) | every turn, once armed (Ω+1 Wave 4) |
| "the curator" | promotion-to-medulla/soul — an explicit `promote` (or future curation pass), on cadence | consolidation cadence |
| "the doc-gate" | a process rule enforced at the landing gate | every PR |
**The one real convergence:** `mission_close{write_light_memory}` and the Stop distiller both
terminate in the same `memorize` door (§C4) — memorize is the shared sink; close and Stop are two
of its callers. The mission lifecycle verbs (`mission_start/next/event/verify/handoff/close`) —
shipped, previously homeless — are **constitutionally homed HERE**: this chapter is their PRD seat;
their mechanics stay in code + `M1ND_INSTRUCTIONS` §3.
### C3.3 A mission's full life (the sequence, distillation gate in its real home)
sequenceDiagram
autonumber
participant H as host (hooks)
participant A as agent
participant M as m1nd owner
participant G as landing gate (CI)
Note over H,M: SESSION OPEN
H->>M: SessionStart hook — north (≤1,200 chars, fail-open)
A->>M: north(task) — first call; reception block on mismatch only
M-->>A: the packet (budget-bound C1) — verdicts wear grammar 1
opt mission-shaped work
A->>M: mission_start … mission_event/verify (id-bearing, explicit)
loop the work loop
A->>M: seek / why / impact … (obey verdicts; focus(mode:"check") to stop)
H->>M: Stop hook (every turn) — cross_verify → memorize (evidence-anchored distiller, Corr 4)
Note right of M: distillation = AMBIENT, machine-owned — not mission_close
Note over A: SESSION CLOSE — the ONE ritual: "leave it warmer"
A->>M: memorize (claims) + learn (grades) + letter (only if m1nd misbehaved)
alt mission open
A->>M: mission_close(write_light_memory: true) — the ritual's with-mission form
H->>M: SessionEnd / PreCompact — persist · trail_save · alerts_ack (machine-owned)
A->>G: PR — battery + doc-gate + agent-docs + PATHOS (THE gate)
Note over M: CADENCE (never per-turn): consolidation pass · promote (curation) · triage sweep
### C3.4 focus ↔ mission ↔ trail — the one working-context, three renderings (JOINT-F)
Three shipped subsystems name "the current working context" and had **no declared relationship**
(the map's highest-priority overlap). The relationship, declared:
> **There is ONE working context. `focus` is its live form, a `mission` is a named `focus` session
> with a handoff, and a `trail` is its persisted, resumable form.**
- **`focus`** owns the goal-conditioned working-set + attention budget while the session runs.
- **`mission_*`** is that focus session given a name, `expected_phases`/`non_goals`/`non_claims`,
and a **handoff packet** — the mission verbs are the focus session's ledger, not a rival notion
of context. (Homed at §C3.2 as their PRD seat.)
- **`trail_*`** (`trail_save`/`resume`/`merge`/`list`) is the **persisted form of the focus
working-set** — the same context frozen so a later session resumes it. `trail_save` is therefore
a machine-owned beat (§C3, PreCompact), never a fourth thing.
- **One "have I gathered enough" signal, three renderings:** `focus.sufficiency`,
`coverage_session` (the blind-spot map), and `mission_verify` are **not three measures** — they
are three surfaces onto the same sufficiency question, and the stop-check that answers it is the
single §C2.3 enum (`gathering | sufficient | saturated`, +`not_here`/`looping`) riding
`focus(mode:"check")`. No surface may mint a fourth sufficiency verdict.
*(This declares the relationship the map flagged; the surface unification — collapsing the three
sufficiency renderings behind one call — is a mechanism the FOCUS/mission family owns, not new law
minted here.)*
## C4 — The Write Spine (JOINT-H)
**ORG-INV-1 · One write door.** `memorize` (the `light_author` path) is the ONLY door through
which durable knowledge enters any store. Its callers, enumerated and closed: the Stop distiller ·
`mission_close{write_light_memory}` · `debrief` (findings under the child's id, lessons under the
grader's) · `promote` (the medulla copy) · reception policy-learning · the agent's own ritual ·
**and every future writer, `soul_update` included** — a subsystem that writes durable memory
without passing this door is unconstitutional.
Everything the organism enforces about knowledge, it enforces **at this one door**:
- the supersession gate (`WouldDowngrade` — weaker writes bounce);
- invalidate-and-keep (`.history/`, `State: outdated`, forever);
- per-slug flock (write serialization);
- provenance stamps (`Created`, `Source-Agent`, `Origin-Brain`, promotion chain §C8);
- the hygiene floor (secret-scan + conflict-marker guard);
- **the witness-tissue refusal (F16):** a letter path in a claim's `evidence:` array is refused at
write time — same mechanism class as the conflict-marker guard (§C8.5).
One door means one place to hold every invariant, one battery surface, zero drift between writers.
## C5 — Identity & Reception (JOINT-C · F2 · F23 · letter#49)
### C5.1 The agent_id law (supersedes TWO-TIER §11 — amendment written there, §C11)
**Canonical grammar: `host:tier:name[@parent]`**, lowercase —
`host:main:orchestrator`, `host:sub:burst-1t@orchestrator`, `human:maintainer`,
`ci:battery`. Field practice + the sealed identity taxonomy win over the older `host:role` sketch.
One grammar, one parser, one test:
- **hosts are open data** (the HOST-MATRIX roster; never a hardcoded four-name enum — §C7.4);
- **the `tier` token carries the seat** (`main` = the orchestrating seat, `sub` = spawned);
- **the `name` token carries the role label** for human readability (`*-orchestrator`, `burst-*`);
- **role-bearing gates parse the TIER token**, never grep the name: `promote` executes when
`tier == "main"` OR `agent_id == human:maintainer` (etiquette-by-audit unchanged, TT-INV-7 — the
predicate is parseable, violations auditable via `Promoted-By`);
- `@parent` chains provenance for spawn trees; absent = top-level.
### C5.2 Reception: the envelope block is canon (F23)
The **shipped form is the law**: reception is an **envelope block attached to** results
(`north`/`health`/`session_handshake`; present on the first result and until answered) — additive,
typed, non-breaking. The designed carrier that returned the reception packet **as the result of
whatever verb the agent first called** is **repealed** — a `seek` that returns a non-seek shape
breaks every typed client on day one. The full `m1nd-reception-v0` packet (`what_exists`
enumeration, walk-up verdicts, options) rides **inside the block** or behind the `bind` answer.
The requirement ("the system should already surface all the information") is satisfied by the block.
### C5.3 The child law: a delegation packet IS a mother-pre-filled reception (JOINT-C, constitutional)
Reception (an agent picks a brain) and delegation (a mother hands a brain) are the write- and
read-sides of one binding contract; `M1nd-Caller-Root` (reception's wire fact) and
`mission.binding` (the packet's wire fact) are the same datum at two hops. Therefore:
- the packet's `mission.binding` **names** the brain — the child never chooses, it **verifies** it
landed on the named brain via reception + `session_handshake`;
- on a match the child's first contact is the SILENT reception (TT-INV-12) — inheritance is
explicit cargo, never routing luck;
- the child's `memorize` lands project-private in the named brain under the child's own id
(MEDULLA §8 made constitutional; M7 carries the tier-labeled memory slice).
### C5.4 Reconnect-rebind (letter#49 — [SHIPPED 2026-07-05 · R13])
After an MCP reconnect, the wire-session bind drops and `caller_root` collapses to the host cwd —
a session that HAD a brain silently loses it. Law: **a rebind after session-id change re-runs
first-contact classification with the disk roster consulted as a candidate** — route to the
existing brain on a match, or surface reception suggesting `project_root=<that repo>`, never the
host-cwd default. Owner: the reception/routing family (TT Slice 2R residue). Ladder: §C10 R13.
**[SHIPPED]** the roster consult landed as `ProjectBrainRegistry::covering_brain` (the UNIQUE brain
related to the caller by ancestry, `None` on zero/ambiguous) + `mcp_http::enrich_reception_with_roster`
at the owner-default mismatch seam: the mismatch reception now carries `known_brain` and its
`ingest_your_repo` call points at the existing repo root (a warm re-bind), not the host cwd — silent
match untouched (TT-INV-12). See §C10 R13 for the full receipt.
## C6 — The Verb Budget (F7 · F8)
### C6.1 The census law
**114 live tools** — was 112 at `5b1a37d`; **+2 at R6** (`delegate` + `debrief`, this landing),
re-measured by the same unique name-array census. (The critique independently measured 112 two days
before R6; NEXTGEN's "119" is corrected by amendment, §C11.) Law: every doc census carries its pin;
**the battery pins the advertised count** (a case asserts `tools/list` count == the documented
number) so drift fails CI, not credibility (TT-INV-11 applied to tool counts).
**The advertised-surface budget:** the ESSENTIAL tier (`ESSENTIAL_TOOLS`, ~25 tools, live) is the
default face every host sees; **ceiling ≤ 40**. The full surface stays callable
(`M1ND_TOOL_TIER=full` + dispatch) but is deliberately tiered — the deferred-tools reality (letter
L14: a session died between ToolSearch-load and first call) proves hosts already tier m1nd
involuntarily; m1nd tiers itself deliberately.
### C6.2 The kill/keep table (adopted whole; capabilities survive, verbs die)
| Designed verb | Verdict | Becomes |
| `delegate`, `debrief` | **KEEP** | genuinely new I/O contracts (composite read; graded mutation) |
| `bind` | **KEEP** | the reception answer — no existing verb means "choose a brain" |
| `promote` | **KEEP** | the audited crossing needs its own recorded act (+ §C8 riders) |
| `underwrite` | **KEEP** | real composite gate; verdicts renamed onto the trust ladder (§C2.1) |
| `envelope` | **KILL as verb** | the §O.4.1 response layer — **§O.4.1 wins its own contradiction** (F8): "not optional and not a separate call" forbids `envelope(any_answer)`; rolled out per-tool behind the Move-0 gate; Move 1's deliverable re-reads "the envelope layer + its calibration harness" |
| `stop_gate` | **KILL as verb** | `focus(mode:"check")` grown `not_here \| looping` (§C2.3) — FOCUS already ships the check mode |
| `quarantine_on_boot` | **KILL as verb** | a `north`/boot behavior — north IS the boot verb; a second boot verb splits "never start cold" in two |
| `handoff_receipt` | **KILL as verb** | fields on `mission_close`/`mission_handoff` — `write_light_memory` is the precedent flag pattern |
| `negative_space` | **KILL as verb** | an `audit`/`missing` mode — pure re-ranking of four read surfaces |
| `hot_blast` | **KILL as verb** | `impact(weight:"runtime")` — a weighting input, not a capability |
| `swarm_collision` | **KILL as verb** | the §O.12.4 packet's `swarm:{collisions:[…]}` field — Slice 3 set algebra; a twin verb is the duplication |
| `inbox_drop` | **KILL as MCP** | letters are **appended, not called** — the global spool IS the drop path (MEDULLA §9.2's own law); resolves the homeless verb by dissolution |
| `inbox_sweep` | **KEEP, demoted to CLI/REST** | its consumers are the triage session and the Hall (`GET /api/mailbox`) — never the in-loop agent; stays OFF the MCP surface |
| `calibrate_delegate` | backlog | §O.12.8 already defers it |
| `soul_*` | **null hypothesis: ZERO new verbs** | a soul that is curated medulla claims under a `kind:` needs no verb; the SOUL-PRD must beat this hypothesis in writing (§C8.6) |
### C6.3 The future-verb rule
**A new capability enters as a flag/mode/field on an existing verb unless its input→output
contract is genuinely new.** The proponent writes the reuse audit (the table above is enforcement
pass #1); a new verb PR that cannot show the audit is refused at review. Tool count changes
re-pin the battery case in the same PR.
## C7 — Universality: the stranger test (F11 · F12 · F17 · the universality law)
### C7.1 The constitutional invariant
**ORG-INV-2 · The stranger test.** Every mechanism must have a birth story on a fresh
`npm i -g` machine: no maintainer file, no launchd, no Claude, no `~/m1nd` checkout. A feature that
only exists on the maintainer's machine is a prototype wearing a feature's name. Every PRD slice
states its stranger story or its honest tier (below).
### C7.2 The medulla birth story
- **`m1nd init --medulla`** (and the first-run doctor offer) creates a **host-neutral medulla** at
`~/.m1nd/medulla/` on machines that never had one — the doctrine tier, the promotion target, and
the medulla box exist for every user, not only the maintainer.
- **Migration note:** the maintainer's `~/.m1nd/runtimes/claude/` remains their medulla via the registry's
`brain_kind` — the `[KILLED: S6 directory move]` decision protected a live migration and stands;
it does **not** define the fresh-install default. New docs and doctor prescriptions name the
host-neutral path; the host-named path is grandfathered, never taught.
- **Doctrine precedence, in m1nd's own terms:** `medulla-absolute > project claim >
medulla-default`. The maintainer's `CLAUDE.md` mapping (ABSOLUTE-global > repo > general-global) is
demoted to an *example instantiation*. MED-INV-8's pruned-to-pointers rule points at **medulla
claims** (portable to every host), never at a host's private rule file.
### C7.3 The OS service tier
Service management is a doctor recipe matrix — launchd / systemd / Task Scheduler — OR the honest
**"project-brains-only" tier**, stated as a SUPPORTED mode, not a gap: on a machine with no service
manager the medulla is born on first use and served on demand; only ambience degrades. `launchctl`
verbatim in prescriptions is a macOS recipe row, never the product's voice.
### C7.4 Hosts are data
The host roster lives in the HOST-INTEGRATION-MATRIX and in registry/config **data** — never in a
hardcoded enum (TT §11's four-name list dies with the §C5.1 amendment; the `host` token is open).
### C7.5 Box-birth consent (F12 — adopting option b)
The mailbox distributor may create `<repo>/.m1nd/` to file a box, **but with an
ignore-by-default `.gitignore` covering `inbox.jsonl` until the repo's own `m1nd init` flips it to
committed** — the loud-warning ceremony remains the ONE consent moment for committed telemetry
(TT-INV-8 amended here by law, not by silence: *boxes get a consent-deferred birth — local
immediately, git-traveling only after init*). The maintainer's ask survives whole: `~/project-d` gets its
box today (local, swept by triage); it starts traveling when project-d consents. The §7.1 secret-scan
floor at filing stands unchanged.
### C7.6 Calibration is metabolism (F17)
Birth runs `calibrate_predict`: `m1nd init` / the one-call bootstrap arm calibration as part of
being born (where co-change history exists to calibrate on; honestly deferred with the cap-in-words
footnote where it does not — §C1.4). The clone-gate battery asserts the composed packet **states
the cap in words** on an uncalibrated brain. Nobody owes the tool a favor for it to be honest.
## C8 — Evidence Integrity: the two tissues (F13 · F14 · F15 · F16 · F26)
### C8.1 The two-tissues law
**ORG-INV-3.** Every claim is either **VERIFIED tissue** (its `evidence` paths re-derive: anchored
to real code nodes, re-hashed at read time) or **DECLARED tissue** (witness/self-report — honest,
but UNPROVABLE by construction). Every surface renders which tissue it is serving; declared tissue
never wears verified formatting. This binds every store, every packet, every future soul.
### C8.2 Promotion re-anchors evidence (F13) — [SHIPPED 2026-07-05 · R4/M6]
`promote` gains **step 2.5**: evidence paths are rewritten **origin-qualified**
(`evidence: <origin_root>#<path>`, beside `Origin-Claim`). Medulla-side freshness then either
**(a)** delegates re-hashing to the origin brain when reachable (the doctrine-read channel,
reversed), or **(b)** stamps the claim **`evidence_unverifiable`** — the state X360 §5.4 already
defines (`UnprovableReason::EvidenceUnverifiable`) — rendered on every recall. **A medulla claim
never reads fresher than it can prove.** Without this, the doctrine tier is un-falsifiable by
construction — declared tissue wearing verified costume at the most-read tier.
> **Shipped (honest):** `promote_handlers::reanchor_evidence` implements both channels — a claim
> carrying code evidence is origin-qualified against its `Origin-Brain` (channel a); a claim whose
> origin root does not resolve is stamped `Evidence-Unverifiable: true` in the medulla copy's
> frontmatter (channel b), rendered so every recall surface can label it declared tissue. A claim
> with NO evidence is neither (it was never verified tissue). **Residue:** the marker is rendered
> frontmatter; the X360 `UnprovableReason::EvidenceUnverifiable` enum the spec references is NOT
> yet a code type — wiring the marker through the trust envelope's verdict (so a recall of an
> unverifiable medulla claim caps its trust) is a later slice. Cases:
> `reanchor_origin_qualifies_code_evidence`, `reanchor_marks_unverifiable_without_origin`,
> `reanchor_no_evidence_is_neither`, + the end-to-end origin-qualified assertion in
> `promote_lands_audited_and_surfaces_cross_brain`.
### C8.3 The promotion evidence-class gate (F14) — [SHIPPED 2026-07-05 · R4/M6]
**P3** joins P1/P2: a claim may promote only if `State: verified` OR `Source-Agent` is
`human:maintainer`. One frontmatter check inside the M6 verb, one battery case. Declared maker
findings stop one verb short of every session's doctrine beat.
> **Shipped:** `promote_handlers::evidence_class_gate` — checked INSIDE the verb before any write;
> a non-verified, non-founder claim is refused with a typed reason that cites C8.3. Cases:
> `c83_gate_allows_verified`, `c83_gate_allows_founder_sourced`, `c83_gate_refuses_unverified_maker_claim`,
> + the end-to-end `c83_refuses_unverified_claim` (nothing reaches the medulla on refusal).
### C8.4 The curator laws (F15) — and who verifies the curator
The consolidation pass (the daemon curator) is bound by three mechanical laws plus one seat law:
1. **Evidence union:** a parent claim's `evidence` = the union of its children's **code** evidence
paths; children ride in `Supersedes`/lineage fields, **never** in `evidence` — claims never cite
claims as evidence.
2. **Merge-and-recite, never re-phrase** (Correction 4 inherited verbatim): any sentence in the
parent not present in a child is a battery failure. The curator has no transcript; it may not
author.
3. **Confidence caps at max(children)** — a parent never exceeds its strongest child.
4. **The seat law (the circularity, answered):** curator output passes a **soul-check pass run by
a DIFFERENT session/agent** than the one that curated — grader ≠ author, the debrief precedent
(findings under the child's id, lessons under the grader's) applied to curation. The check
itself is flag-first (§C6.3): a `cross_verify`/`audit` mode, not a new verb, unless the audit
proves otherwise.
### C8.5 Letters are witness tissue (F16)
A letter is citable as **provenance** ("this was felt, on this date, by this agent") and never as
**verification** (no claim's truth grade rises because a letter asserts it). Concretely: a letter
reference in a claim's body is fine; a letter path in `evidence:` is **refused at the write door**
(§C4). Binding on M7b and on the SOUL-PRD.
### C8.6 The SOUL slot (F26 — the constraints bind the soul, not the reverse) — [S0 + S1 substrate SHIPPED 2026-07-05 · R16]
The SOUL-PRD (in flight, absent from main at this writing) inherits, non-negotiably:
1. **No fifth grammar** (§C2.4): soul claims are grammar-2 beliefs; their answers wear grammar 1.
2. **The headline obeys the one-line rule** (§C1.3.3): one line + the pull verb, in the packet.
3. **`soul_*` must beat the zero-new-verbs null hypothesis** (§C6.2) in writing.
4. **Letters never evidence** (§C8.5).
5. **`soul_update` routes through the write door** (§C4).
6. **Curator output is seat-verified** (§C8.4).
7. **The soul rides LAST on the ladder** (§C10 R16): it is the curated apex of `promoted` claims —
it cannot be designed coherently before the medulla state machine is real.
If the SOUL-PRD needs to break any of these, it argues the exception in writing against this
section — the constitution is the null hypothesis it must beat.
> **Shipped (honest — the seven constraints, checked against the code):** R16's S0 + S1
> substrate landed with every constraint held. (1) **No fifth grammar** — soul claim
> states (`verified-fresh`/`evidence-stale`/`superseded`/`receipt-required`/`declared`)
> are computed at check time from doc+anchors+repo truth, never a stored per-claim field
> (SOUL-INV-7). (2) **The headline obeys the one-line rule** — `soul_read` returns ONE
> authored headline; the north sub-atom (S2) is deferred but bounded ≤220 chars by design.
> (3) **`soul_*` beat the zero-new-verbs null hypothesis in writing** — SOUL-PRD §4 argues
> it: `soul_check`/`soul_read` are read-only composers a `kind:`-tagged medulla claim
> cannot be (they verify a git-tracked DOCUMENT against reality); the curator is NOT a
> verb (a workflow); `soul_update` is a MODE of `memorize`, not a new verb. (4) **Letters
> never evidence** — the soul never reads a letter path as evidence; it verifies fs/git/
> graph anchors only. (5) **`soul_update` routes through the write door** — it is the
> `soul_source` field on `LightAuthorInput`, rendered as `Soul-Source` frontmatter through
> the ONE `memorize` sink (SOUL-INV-8), same gates as every write. (6) **Curator output is
> seat-verified** — `soul_check {verify_curator_report}` refuses grader==author (§C8.4).
> (7) **The soul rides LAST** — R16 is the ladder's finale, mounted on the shipped medulla
> (R2–R4). The one honest exception argued in writing (SOUL-PRD §9): S0 is read-only and
> mounts on organs all shipped BEFORE the medulla, so it ships first as the RED-maker.
## C9 — Capacity & the Operational Envelope (F18 · F19 · F20 · F21)
### C9.1 THE EVICTION GATE (F18 — hard pre-condition, battery-pinned)
The interim owner-hosted topology recreates the blast radius two-tier was built to kill (one owner
crash now takes N brains' warm state — "16:44 at scale"). Law:
> **Before the owner hosts brain #5, OR before `tier:"all-brains"` ships — whichever comes first —
> the owner MUST have (a) an LRU eviction/limit for warm brains and (b) per-brain
> persist-on-evict.** Battery-pinned: bootstrap K+1 brains → owner RSS bounded → `kill -9` the
> owner → **every brain warm-boots from its own snapshot**.
**[SHIPPED 2026-07-05 — ladder R15.]** The warm project-brain map
(`ProjectBrainRegistry.brains`) is now LRU-bounded at `DEFAULT_WARM_BRAIN_CAP = 4`
(configurable via `with_capacity`). Every insert — both the one-call bootstrap and a
warm-boot resolve — routes through `insert_with_eviction`: when the map would exceed the
cap it picks the least-recently-used **project** brain, calls `SessionState::persist()`
on it (the same #230 store-write path a warm-boot re-hydrates from), then drops the `Arc`.
The **bound dev graph is never in this map** (it lives on `AppState::session`), so it can
never be an eviction victim — only project brains evict. Battery case
`eviction_gate_bounds_the_map_and_persists_on_evict_surviving_kill9`
(`tests/two_tier_project_brains.rs`): bootstrap cap+1 brains → the map never exceeds the
cap → the bound graph answers unchanged → `drop(owner)` (the kill-9 stand-in) → a fresh
owner over the same store warm-boots **every** brain, evicted ones included, with matching
node counts. The persist-on-evict step is isolated at the unit level
(`project_brains::eviction_gate_tests::eviction_persists_unpersisted_state`): a node added
to a brain's in-memory graph **after its last persist** reaches the on-disk snapshot only
because eviction flushed it (RED without the flush: the store snapshot is never written).
**Line of intent, recorded:** the interim owner-hosted variant is **topology debt**; TT Slices 2/3
(process-per-repo) are its retirement. Its convenience killed live friction and must not quietly
become the end state — §2's whole argument was against exactly that end state. **Residue:** the cap
bounds warm state but the blast radius is still N-brains-per-owner until the process-per-repo flip;
eviction is LRU-by-touch (no size/age weighting yet); `tier:"all-brains"` (R3) fans out over the
disk roster and must warm-boot through this same gate so a wide fan-out cannot pin more than the cap.
### C9.2 `promote` across the topology flip (F19)
M6's owner-local mechanics ("load the claim from the source brain's store") are stamped
**interim**. At canon topology the direction inverts: the **project owner pushes** the copy over
the same HTTP channel as the doctrine read (the V2 `memorize --promote` path). The battery case is
written against the **contract** (same input/output), so it survives the transport flip unchanged.
### C9.3 The loopback law (F20)
**ORG-INV-4 · Loopback-only is a named invariant, not an implicit premise.** The bind surface
refuses non-loopback binds outright, battery-pinned. Remote access is a V-later design with
**auth required at that door** — filed as design debt by name, so the first tunnel/VPS PR meets a
written refusal instead of an unauthenticated brain carrying the user's cross-project mind.
### C9.4 Door-coverage honesty (F21)
An invariant is stated **with its door coverage** until every door routes:
*MED-INV-1 holds on routed HTTP; the stdio and paramless-REST doors close at TT Slices 2/3.*
An invariant claimed wider than its enforcement is the honesty-overclaim class the letters keep
catching (the L52 class, at the architecture level). This phrasing rule applies to every ORG/TT/MED
invariant with partial enforcement.
## C10 — THE LADDER (the single cross-PRD build order)
**Reading rule:** an implementer reads this chapter alone and knows what to build Monday. Every
rung: **what · RED-first proof · what it unblocks.** The long pole is the medulla state machine
(M5a→M5b→M6); the two integration points every rung composes over are the write door (§C4) and the
packet spine (§C1). No rung lands without its battery case first (RED), its doc pass, and the
landing gate.
R0["R0 · the honesty patch<br/>MED-INV-6 false-absence fix"] --> R2
R1["R1 · the packet diet<br/>Budget Law enforcement"] --> R2
R1 --> R12
R2["R2 · M5a — storage split,<br/>Origin-Brain, migration, refusal"] --> R3
R15["R15 · eviction gate (C9.1)<br/>LRU + persist-on-evict"] --> R3
R2 --> R4
R3["R3 · M5b — tier recall,<br/>no-leak proven, all-brains"] --> R4
R3 --> R7
R4["R4 · M6 — promote verb<br/>+ C8 riders (2.5, P3)"] --> R16
R6["R6 · Delegation Slice 1+2<br/>delegate / debrief"] --> R7
R1 --> R6
R7["R7 · M7 — packet memory slice<br/>tier + origin_brain rows"] --> R16
R8["R8 · M7b — boxes, telemetry class,<br/>fates, sweep, consent birth"] --> R9
R8 --> R11
R9["R9 · Slice 3M — Mailbox view"]
R6 --> R10["R10 · Pre-Flight Card<br/>(HUMAN Slice 1)"]
R11["R11 · CASE-INTELLIGENCE PRD<br/>fingerprint · cases · sentinels ·<br/>claim-vs-measure · abandonment"]
R12["R12 · the ambient wave<br/>Stop distiller · PreCompact · SessionEnd"]
R13["R13 · reconnect-rebind<br/>(letter#49)"]
R14["R14 · §9.5.1 per-brain partition<br/>(ONE owner: TT backend)"]
R5["R5 · Windows CI honesty triage"]
R17["R17 · conformance_boost rerank<br/>X-RAY → seek/focus (JOINT-J)"]
R16["R16 · SOUL-PRD + slices — LAST<br/>bound by C8.6"]
**R0 — the honesty patch (MED-INV-6 hotfix). [SHIPPED 2026-07-05]** *What:* the false-absence fix
carved OUT of M5b and shipped first as a defect fix: a beat over a non-empty store surfaces memory
or stamps `memory_exists: n` — never "No durable memory yet". *RED:* on file three times over —
letter#52, the critic's session, and THIS session's own opening north (store ≥20 light roots,
`memory: []`). *Shipped:* north stamps `memory_exists` = the on-disk L1GHT store count; the false
absence line fires only when the store is truly empty; RED→GREEN unit tests
(`north_over_nonempty_store_never_claims_no_durable_memory` + the empty-store companion) plus
`SessionState::light_memory_count`. *Unblocks:* trust in the flagship packet — every later rung's
packets are honest. *(Adjudicates F24: urgency adopted by the split; the literal M5b-before-M5a
reorder is refused — tier recall needs M5a's `Origin-Brain` labels; the dependency is real.)*
**R1 — the packet diet (Budget Law enforcement, §C1.3). [SHIPPED 2026-07-05]** *What:* dedup
`fingerprint`/`graph_state`; the memorize write-path stops minting per-file ingest roots (flow fix);
size battery-pinned on a reference graph (≤2k tokens MCP / ≤1,200 chars hook). *RED:* this session's
live packet (dup arrays × 21 roots). *Shipped:* `graph_runtime_summary` carries only
`ingest_root_count` (the full array lives once, in the fingerprint); a `.light.md` written into the
`agent-memory` store collapses to the single store-dir root instead of one root per sidecar; the
battery case `north_packet_within_budget` pins the live packet at ~1,419 tokens and fails CI on
dup-arrays / sidecar-roots / >2k growth. RED→GREEN unit tests
(`north_binding_serializes_ingest_roots_once_not_duplicated`,
`memorize_does_not_mint_per_sidecar_ingest_roots`). *Measured:* the binding's roots footprint drops
from ~356 tokens (array × 2, 21 roots) to ~12 (count + a 2-root array). *Unblocks:* the hook path
can finally carry the packet it is doctrine-bound to deliver → R12; every satellite (inbox count,
soul headline) enters under the one-line rule. *(R0 ∥ R1 — both small, both live defects, both
pre-medulla.)*
**R2 — M5a: storage split + `Origin-Brain` + migration + brainless-root refusal** (MEDULLA §11). [SHIPPED 2026-07-05]
**[SHIPPED 2026-07-05]** *What:* the tier IS the directory — `Origin-Brain` stamped on every
`memorize` (`SessionState::origin_brain()` → project root, or `medulla` for the owner store;
rendered in `render_light_markdown`); the brainless-root refusal at the top of
`handle_light_author` (medulla store + known foreign caller root → refused with the typed one-call
bootstrap, never a silent shared-store write); the migration as a standalone, pure-filesystem
`MedullaMigration` (`medulla_migration.rs`) — `plan` (pure-read dry-run default) → `apply`
(backup-first + count-conserving) → `rollback` (byte-for-byte restore), with the ghost-pointer sweep
of `ingest_roots.json`. *RED→GREEN:* today 25 mixed claims, zero `Origin-Brain`, ghost root, and a
brainless-root memorize lands silently → now every claim carries a tier by directory + `Origin-Brain`,
the refusal redirects honestly, and the migration maps claims to the right store with a proven
migrate→rollback round-trip (no data loss). **CODE-LAND-ONLY: the LIVE migration is HELD for the
maintainer — the code lands and is scratch-proven, but no live owner is migrated here.**
*Unblocks:* R3 (labels), R4 (a real store to promote into), all provenance law (§C8).
**R15 — the eviction gate (§C9.1). [SHIPPED 2026-07-05]** *What:* LRU + persist-on-evict in the
owner's warm project-brain map (`ProjectBrainRegistry`, cap `DEFAULT_WARM_BRAIN_CAP = 4`), the bound
dev graph pinned (it is not in the map, so it never evicts). *RED→GREEN:* the map was unbounded —
bootstrapping cap+1 brains grew it to cap+1 (`warm_len` assert, RED by construction); now every
insert routes through `insert_with_eviction`, which persists-then-drops the LRU project brain so the
map never exceeds the cap and a `kill -9` (owner drop) is survived by every brain warm-booting from
its own store (`eviction_gate_bounds_the_map_and_persists_on_evict_surviving_kill9`); the flush of
state mutated after a brain's last persist is isolated in
`eviction_gate_tests::eviction_persists_unpersisted_state` (RED: no snapshot written without the
flush). *Unblocks:* R3's `all-brains` half and any onboarding past brain #4. **Hard pre-condition —
R3's `all-brains` does not ship without it, and its fan-out must warm-boot through this same gate.**
**R3 — M5b: `tier` recall + no-leak proven + `all-brains`** (MEDULLA §11; gated by R15). [SHIPPED 2026-07-05]
**[SHIPPED 2026-07-05.]** *What landed:* pull-only tier recall at the routing seam — a routed
brain X's default beat composes exactly X's own store + the medulla; `tier` (`project` |
`medulla` | `project+medulla` (default) | `all-brains`) selects which stores the memory beat of
`seek`/`north`/`boot_memory` reads. The no-leak law (MED-INV-1) is mechanical: a claim from brain
Y reaches X's default beat ONLY if it is a medulla claim (promoted/doctrine-born). `all-brains`
fans out over every hosted store, **each warm-boot routed through the R15 eviction gate** (so a
wide fan-out never pins more than the cap), grouped/labeled by `origin_brain`. Provenance-in-recall:
`Origin-Brain` flows frontmatter → `light:origin_brain:` graph tag → `SeekResultEntry.origin_brain`
→ every recall row. *Proof:* the leak permutation matrix (seed Y, assert X's default beat never
carries it, assert `all-brains` does — labeled `origin_brain: Y`), the medulla-doctrine-surfaces-
cross-brain case, the default-labeled case, and the `all-brains`-through-eviction (warm map ≤ cap)
case — all RED-first (`m1nd-mcp/tests/medulla_m5b_tier_recall.rs`). *Held for the maintainer:* the
live `:1338` was NOT migrated/restarted — CODE-LAND-ONLY. *Unblocks:* R4's read path, R7's
tier-labeled packet rows, the real doctrine beat.
**R4 — M6: the `promote` verb WITH the C8 riders. [SHIPPED 2026-07-05]** — step 2.5
origin-qualified evidence (§C8.2), P3 verified-only gate (§C8.3), demotion documented,
agent-workflow surfaces in the SAME PR (era-coherence gate). *What landed:* the `promote` verb as
an owner-level cross-store crossing at the routing seam (reads a project brain, writes the medulla)
— it loads the source claim (hard error on unknown slug), runs the C8.3 gate (`State: verified` OR
`Source-Agent: human:maintainer`), the hygiene floor (secret + conflict-marker scan), re-anchors
evidence origin-qualified (`<origin_root>#<path>`) or marks `Evidence-Unverifiable: true` (§C8.2),
writes the medulla copy through the shared supersession core (a weaker re-promotion bounces
`WouldDowngrade`) with the four provenance stamps (`Origin-Brain`, `Origin-Claim`, `Promoted-By`,
`Promotion-Reason`), stamps the project witness `Promoted-To` (elevate, never move), and re-ingests
so the claim surfaces in another brain's default beat under `tier: medulla`. Demotion is documented
as `learn wrong` / consolidation on the medulla copy — never touches the witness. *RED→GREEN:* a
grep proved no promotion surface existed and no recorded act could carry a finding cross-brain; now
13 unit + 6 end-to-end cases (over the real HTTP seam) prove the audited crossing, both evidence
riders, the gate, the bounce, and demotion-preserves-witness. **CODE-LAND-ONLY: the LIVE `:1338`
was NOT migrated/restarted — held for the maintainer (needs a kickstart to serve the verb).**
*Residue:* the TT §8 conflict-composition rendering + the `witness_superseded` consolidation check
are deferred (they belong with the consolidation pass); `Evidence-Unverifiable` is a rendered
frontmatter marker, not yet wired into the trust envelope's verdict. *Unblocks:* the doctrine tier
is real → R16 has ground to stand on.
**R5 — Windows CI honesty triage (F25; parallel, small, anytime — early, because honesty debt
compounds). [SHIPPED 2026-07-05]** *What:* fix the teardown + `display_name` separator, or demote Windows from
"blocking" in the written proof story until it blocks again. *Proves:* a gate described as
blocking blocks. *Owner:* the CI family.
**R6 — Delegation Slices 1–2: `delegate` (project-tier packet) then `debrief`** (§O.12.10). [SHIPPED 2026-07-05]
**[SHIPPED 2026-07-05 · R6.]** *What landed:* `delegate` composes a read-only
`m1nd-delegation-packet-v0` in north's class — the mother's binding (the named brain), the selected
project-tier memory slice as explicit cargo, ranked anchors, a delegate-time staleness header, known
static dependents, and a deterministic string-stable `prompt_markdown`; three abstain classes
(`needs_ingest` / `unscopable` / `seeds_unresolvable`), each with evidence + `next_move`; a dumb
registry record per packet (the debrief join key). `debrief` is the only mutation (via `memorize` /
`learn`): conformance algebra with a worst-of verdict that always carries fence existence, findings
memorized under the subagent + map-miss lessons under the grader, asymmetric `learn`, and one
`outcomes.jsonl` row per debrief (stamped `outcome_unverified` without evidence). The child law
(§C5.3) is real in code: `mission.binding.workspace_root` IS the datum reception verifies
(`covers_root`) — one datum at two hops, silent on match. `debrief` alone joins
`READ_ONLY_DENIED_TOOLS`; `delegate`'s omission is its ambient legality. Census 112 → 114.
*RED (proven):* no delegate/debrief surface existed (grep-proven) — a child could not inherit the
mother's binding + slice by any recorded act; six behavior tests drive the real HTTP seam green,
including the loop-closing demo (a second `delegate` surfaces the first debrief's finding, by author,
with age). *Unblocks:* R7; R10 renders a stable packet; the outcomes ledger starts feeding
calibration. *(The packet's medulla doctrine block waits for R3/R7 — Slice 1 ships project-tier
only, honestly labeled `tier: "project"`.)*
**R7 — M7: the delegation-packet memory slice — [SHIPPED 2026-07-05]** — `tier` + `origin_brain` on
packet rows; the mother-selected slice as explicit cargo (§C5.3). *Shipped variant:* the composer
labels its own project rows (tier + origin from the routed store, legacy → store identity, never
faked); the medulla doctrine feed folds into the nested `context.memory` via the M5b routing seam
(`serve_and_compose`, `delegate` added to `TIER_RECALL_TOOLS`, nested feed path taught to the
fold/strip machinery); the packet is re-rendered post-fold so the labeled doctrine reaches the
child's one reader string. No new verb, no dispatch/schema change. *RED→green:* a two-brain fixture
asserted the rows lacked `tier`/`origin_brain` (both tests red), then green with a `tier=medulla`
doctrine row beside a `tier=project` fact row, each origin-labeled. **CODE-LAND-ONLY: live `:1338`
NOT migrated/restarted.** *Unblocks:* child sessions that inherit exactly what the mother chose,
auditable.
**R8 — M7b: the telemetry class + boxes + fates + sweep** — [SHIPPED 2026-07-05] —
`memory_misdelivery` vocabulary, the
distribution (with §C7.5 consent-deferred box birth), letter ids + `answers[]`, fate derivation
(+ receipt `disposition`, §C2.2), `inbox_sweep` as CLI/REST (§C6.2), `GET /api/mailbox`,
`confusion_rate` in doctor. **Carries the §C2.2 orchestrator note: supersede the sealed
inbox-doctrine memory at this landing.** *RED:* four live confusion rows untyped; 53+ letters in
one file, prose-only linkage. *Unblocks:* R9, R11.
> **Shipped (honest):** landed as the pure-filesystem `mailbox.rs` module (letter
> id = `sha256[0..12]` for cross-machine dedup; distribution files each letter into
> exactly one box by the normalized `repo`/`brain` — repo-side box or the medulla
> box, MED-INV-10 held; consent-deferred `.gitignore` birth, §C7.5; idempotent
> append-with-dedup). Fates derived (`wet_ink`/`in_flight`/`fired_clay`/`external`,
> the §C2.2 grammar; `external` never counted). `inbox_sweep` shipped as
> `--inbox-sweep` CLI + `GET /api/inbox_sweep` (OFF the MCP surface, §C6.2);
> `GET /api/mailbox?brain=` reuses the §4A.9 selector (`served_brain` echo);
> `doctor` gained the `mailbox` counts block; `/api/instances` gained
> `mailbox_open_count` (absent-honest). Proof: 14 unit + 5 REST-e2e cases, NEUTRAL
> fixtures. **CODE-LAND-ONLY:** the live spool was NOT distributed and `:1338` was
> NOT restarted (held for the maintainer; `m1nd-mcp --inbox-sweep` runs the local
> distribution when chosen). The §C2.2 orchestrator memory-supersession + the
> secret-scan hygiene floor at filing remain the landing/next-slice acts (see
> MEDULLA §11 M7b shipped-residue). *Unblocks:* R9 (the Mailbox view renders this
> contract), R11 (case intelligence over the stable ids + `answers[]`).
**R9 — HUMAN Slice 3M: the Mailbox view. [SHIPPED 2026-07-05]** Pure rendering of R8's contract
(INV-17/18: only the viewed brain's letters; receipts always linked). *What landed:* the caixinha
per brain — `lib/mailbox.ts` (fate-line derivation ● / ◍ / ↳ / ◌, day-chapter grouping, the
class→matte-chip palette drawn from the five EXISTING non-violet token families, the in-box
receipt-linkage resolver, the INV-17 echo guard reusing §4A.9's `servedBrainMatches`) +
`components/hall/MailboxView.tsx` (`MailboxBody`, the pure component-testable surface, behind a thin
fetch shell). The Hall card's D3 field becomes the ENTRY — "N open" opens THAT brain's box
(`GET /api/mailbox?brain=<root>`), absent-honest (renders only when `mailbox_open_count` exists,
never a fabricated zero). The medulla box is its own labeled entry from the Hall header
("Medulla — relatos transversais", projectless letters only). Read-only: zero compose affordances
in the DOM. 20 new tests from five REAL captured `/api/mailbox` fixtures (neutral repo names in the
committed fixtures — the live data it displays may name projects, but that is local runtime data,
never committed). `npm test` 227 green; tsc + vite build green; violet-lint + icon-lint green
(external wears stone, not violet); dist regenerated + committed; air-gap grep zero external hosts.
**R10 — the Pre-Flight Card (HUMAN Slice 1). [SHIPPED 2026-07-05]** The §C1 reader-2 rendering, after
R6 stabilizes the packet shape; carries the F17 cap-in-words and the Budget-Law pointer (§C11).
*What landed:* `PreFlightCard` (`components/preflight/`) renders the SAME north packet an agent
receives, read for a human — the pure view-model in `lib/preflight.ts` (every derived value traces
to a real packet field; a field with no packet field behind it renders absent, never invented,
INV-01). The beats, top to bottom: **BINDING** (trust mode in action language + the real fingerprint
node/edge counts + the JOINT-I reception rider, verbatim through the abstain-class `FreshnessBanner`),
the **VERDICT** (`act`/`reverify`/`abstain` — abstain in quarantined iris violet — plus the `impact`
blast line in floor language, INV-08), **ANCHORS** (focus + PageRank as a strip, real node_ids only),
**WHAT AGENTS KNOW** (the R7 memory beat: claim + author + age + tier + origin_brain, absent-honest,
plus the R0 `memory_exists` truthful empty line — recall miss ≠ empty store), **HONEST GAPS**
(first-class violet `GapCard`s, one action each), and the `next_move` button (verbatim, absent → no
button). Entry: the drawer's `[Check before editing]` opens it seeded with the selected node, fetching
the REAL `north` + one `impact`, both scoped to the viewed brain (§4A.9); ESC/backdrop ascends. SOFT
PROOF: matte, nothing glows, icons from the registry only, all violet delegated to the allow-listed
soft components (`TrustDot`/`VerdictChip`/`GapCard`/`PostItChip`/`FreshnessBanner`). *RED (green):*
card golden against two REAL captured packets (`preflight_north.json` warm + `preflight_north_degraded.json`)
+ `north_cold.json`/`impact.json`; 33 net-new tests (card render-gate + view-model), 205 UI tests
green; violet-lint + icon-lint green; `tsc` clean; `vite build` green with the dist air-gap grep at
zero external resource loads; dist regenerated + embedded (rust-embed).
**R11 — the CASE-INTELLIGENCE PRD (the homeless organ, JOINT-K — PRD first, then slices) —
[PRD SHIPPED 2026-07-05 → `docs/CASE-INTELLIGENCE-PRD.md`: five organs under the §C2.2 overlay
law, CASE-INV 1–7, slices S0–S4 defined there (S4 honestly gated on the ambient wave); shipping
the PRD closes this rung].**
After R8 (needs stable letter ids + `answers[]`). Scope, from the sealed doctrine: fingerprint
(tool+symptom) with auto-escalation on 2nd recurrence · cases (N letters → 1 root cause → 1 fix →
linked receipts) as the §C2.2 grouping overlay · absence sentinels in the battery (post-restart
diff — the "diff de permanência" made mechanical) · claim-vs-measure audit · **the abandonment
signal** (agent stopped using m1nd after an error ⇒ automatic honesty-class letter — the corpus's
one auto-generated calibration ground truth, the JOINT-D seam made real).
**R12 — the ambient wave (the distillation gate in its real home).** Stop → `cross_verify` →
`memorize` with the evidence-anchored distiller (Correction 4: anchor to what the turn touched,
never free-summarize); PreCompact `trail_save`; SessionEnd persist. Depends on serve/attach (done)
+ R1 (the hook budget) + **a maintainer green-light for hook install** (named human gate). *RED:* the A/B
harness re-run against the served owner.
**R13 — reconnect-rebind (§C5.4). [SHIPPED 2026-07-05.]** *What landed:* the reconnect misroute
(field letter#49) is closed at the routing seam. After an MCP reconnect the wire session is minted
fresh (`bound_project_root` cleared) and the bridge re-stamps `M1nd-Caller-Root` = the host cwd —
which, when the host was launched ABOVE the repo, is an ANCESTOR of it, not the repo. So the
caller_root had no brain of its own, the bound graph did not cover it, and the call fell to the owner
graph with reception suggesting `ingest project_root=<host cwd>` — the wrong root, blind to the
existing project brain under the caller. The fix consults the disk roster (the R8 cold-listing
machinery) at the ONE seam that holds both the caller_root and the registry: a new
`ProjectBrainRegistry::covering_brain` returns the UNIQUE known brain related to the caller by
ancestry (either direction) and abstains (`None`) on zero (unknown repo) or more than one (ambiguous
— nested brains / a workspace over several repos: honesty over a guess); an exact-match root is
excluded (that path is a silent bind, not a rebind). On the owner-default mismatch path,
`mcp_http::enrich_reception_with_roster` rewrites the reception to name that brain — `known_brain` +
the `ingest_your_repo` call now points at the repo root (a warm re-bind, not a fresh birth). One seam
covers `north` / `health` / `session_handshake`; a matched caller still binds silently (TT-INV-12
preserved). **Proof:** RED→GREEN in `m1nd-mcp/tests/two_tier_project_brains.rs`
(`reconnect_reception_prefers_the_existing_brain_over_the_host_cwd` — the letter#49 shape;
`reconnect_unknown_root_keeps_the_plain_reception_and_a_match_stays_silent` — the unchanged branches)
+ the `covering_brain` unit test in `project_brains.rs` (both ancestry directions, no-relation `None`,
ambiguity `None`, exact-match excluded). Full `cargo test -p m1nd-mcp` green (802); clippy + fmt clean.
**R14 — the §9.5.1 per-brain partition — ONE owner: the TWO-TIER backend. [SHIPPED 2026-07-05.]**
(Resolves map contradiction #5: HUMAN §4A.6/G4 and MEDULLA S6/MED-INV-7 are consumers, not
co-owners.) *What:* the session/query counters stop being owner-global and partition on the
session's bound brain (`session.bound_project_root`) — because each project brain is a full
`SessionState` and a routed call dispatches against the brain that owns the caller
(`mcp_http::route_and_run` → `serve_and_compose`), the counters were already partitioned at the data
layer; R14 exposes the partition at the surface. `http_server.rs::instances_listing` now enriches
every `/api/instances` entry with its OWN `attached_sessions` + `query_count` + `calibration_armed`
(self from its own `SessionState`; a warm project brain via `ProjectBrainRegistry::warm_session_stats`;
a dormant brain absent-honest, never a faked 0), so per-brain surfaces (the Hall G4 line, per-brain
calibration state) stop wearing owner-global numbers. The owner-wide total moves to the owner's own
receipt, labeled owner-wide — not gone, correctly attributed. The 2H residue family (the
`[needs-backend §9.5.1]` card fields, per-brain `calibration_armed`) rides this rung and becomes real;
the interim "across all brains" G4 caption is removed. *RED→GREEN (letter#51's misattribution case,
PROVEN):* `m1nd-mcp/tests/hall_brains_listing.rs::per_brain_counters_partition_on_the_bound_brain_not_owner_global`
(two brains driven to divergent counts, each card wears its OWN count and NEVER the cross-brain sum) +
`dormant_project_brain_omits_live_session_counters_absent_honest`; UI `cardV2.test.ts`/`card-v2.test.tsx`
(qualifier gone, absent-honest). Full `cargo test -p m1nd-mcp` green; UI 207/207; clippy + fmt clean.
**R17 — the conformance_boost rerank (JOINT-J: X-RAY steers attention) — owner: the X-RAY/seek
family; small, parallel. [SHIPPED 2026-07-05]** *What:* grammar-4 conformance becomes the payoff axis it
was designed to be — the manifesto stops being a report and steers what loads. An additive
`conformance_boost` term rides the shared `handle_seek` rerank (BEDROCK **+0.20**, EROSION **−0.30**);
because `focus` is a thin layer over `handle_seek`, the same term boosts seek and focus at once. It
**composes** with the two terms already in that rerank — trust × tremor (damping, multiplicative) ×
conformance (additive) — the one place intent-vs-reality feeds attention. *What landed:* the additive
term + full plumbing (constants, `resolve_node_conformance`, `SeekInput.conformance_aware`,
`SeekOutput.conformance` summary + erosion-in-result-set drift note) was implemented on `main` in
`8244c48`; this rung composed it onto the **corrected base** (below) and pinned the composition with a
RED→GREEN proof + a battery no-regression. Both the BEDROCK up-boost (+0.20) and the EROSION malus
(−0.30) ship — the P1 malus-only leapfrog is superseded (up-boost is live and proven). *Precondition
(LANDED):* the centrality-vs-semantic balance of this rerank was corrected first (#278) — the
`graph_activation` term is now gated by node relevance so a high-PageRank/low-similarity hub can't ride
pure centrality to the top; `conformance_boost` sits on that corrected base. *RED→GREEN (PROVEN):* a
synthetic rerank fixture — with the term OFF, a BEDROCK node and an EROSION node rank by base score only;
with it ON, the BEDROCK node out-ranks its pre-boost position (+0.20) and the EROSION node drops (−0.30).
Composition pinned: a BEDROCK-but-semantically-irrelevant node still can't ride the boost above the
relevant pool (base relevance + the activation gate gate it), the exact per-node delta equals
`boost × heuristic_factor` (trust × tremor still composes, never dominated), and absence/opt-out stays
byte-identical. Battery: `m1nd_wins` unchanged (20/20, 38/38 pass) before vs after. *Residue:* X360 §5.5
P3 (freshness-aware `Unprovable` downgrade) + P4 (calibration surface for the three constants) remain
deferred — see the X360 §5.5 note.
**R16 — SOUL: the PRD, then slices — LAST. [S0 + S1 substrate SHIPPED 2026-07-05]** Bound by
§C8.6's seven constraints; designed against a working medulla (R2–R4 landed). *What shipped:*
`soul_handlers.rs` — `soul_check` (parse `docs/PATHOS.md` → anchored claims → classify
path/line-hint/symbol/git/consistency → verify → the honesty report + one-line **freshness
receipt**: N-fresh/M-stale/K-priced @sha) + `soul_read` (the explicit pull), both read-only; plus
S1's substrate — `soul_update` as a `memorize` mode carrying `Soul-Source` provenance through the
ONE write door, and the §C8.4 curator SEAT CHECK (`soul_check {verify_curator_report}` refuses
grader==author, silent prunes, declared-tissue removal, a missing `still_stale` valve). *RED-first
proof:* no soul surface existed on `main` (grep-proven); `soul_check_behavior.rs` is the GREEN that
could not exist before — 6 cases + an ignored live-PATHOS probe. *Seed battery, on the LIVE cp10
soul:* **13 fresh · 14 stale · 61 declared** (the Access Map cites a deleted probe — the drift the
PRD predicted, now measured). §C8.6's seven constraints verified against the code (see the C8.6
shipped note). *Deferred, honest:* the automated curator SWEEP + ENXUTO budget enforcement (S1
automation — agent-executed by design), the north-packet `soul` sub-atom + Hall D4 (S2), the
pathos-skill call-through (S3). If the SOUL-PRD row lands here at merge, the §C8.6 checkmarks are
verified against its text.
**Parked (owner named, no rung — parked is a state, not a euphemism for dropped):**
| Item | Owner | Why parked |
| Cross-machine zombie claims (supersession forks once memory travels by git) | TWO-TIER V2 (named at TT §21.3 + MEDULLA §12.3) | blocked on git-travel being real (TT Slice 3); revisit at that landing |
| Full per-brain calibration mechanism | TT §9.5.1 family (R14) + the §C7.6 birth rider | R14's partition landed (per-brain `calibration_armed` now surfaces on each `/api/instances` entry, 2026-07-05); the FULL per-brain calibration *measurement* mechanism still rides the birth rider — R14 exposes armed-vs-not, not a per-brain re-measure |
| `inbox_drop` | **dissolved** (§C6.2) | the spool append IS the drop; a CLI convenience may ride R8 only if a real consumer proves need |
| SSE pure-reader relay gap | Living Tree Slice 1 (PATHOS Known Problems) | already owned there; not this constitution's to re-own |
| PATHOS auto-refresh push-back | the maintainer (decision A/B — PAT secret vs checks bypass) | human decision, named |
## C11 — Amendments applied (the surgical ledger)
Written into the feeder PRDs **in this same PR**, each ≤3 lines, each pointing here as source of
law (F22's "superseded-by" pointers, instantiated):
| # | File · section | The line | Law here |
| A1 | `TWO-TIER-BRAIN-PRD.md` §11 | agent_id `host:role` **superseded** → `host:tier:name[@parent]`; promote-gate parses the tier token | §C5.1 |
| A2 | `TWO-TIER-BRAIN-PRD.md` §19 | pointer: letter grammar + box-birth consent are constitutional law (MEDULLA §9.2 stays the mechanism spec) | §C2.2 · §C7.5 |
| A3 | `NEXTGEN-AGENT-PRD.md` §O.2 | census corrected: ~~119~~ → **112** @ `5b1a37d`, re-measured at every landing, battery-pinned | §C6.1 |
| A4 | `NEXTGEN-AGENT-PRD.md` §O.4.2 | killed-verb amendment note (strikethroughs, not deletions): only `underwrite` survives as a verb; `envelope` = the §O.4.1 layer (§O.4.1 wins) | §C6.2 |
| A5 | `HUMAN-LAYER-PRD.md` §4.2 | pointer: the card renders a budget-bound packet (≤2k/≤1,200; one-line satellites; drops in `non_claims`); renders, never widens | §C1.3 |
**Owed beyond this PR (named so they cannot be forgotten):**
- `M1ND_INSTRUCTIONS` §3 gains the one sentence naming memorize+learn+letter **one ritual**
("leave it warmer"; `mission_close` = its with-mission form) — rides the next slice that touches
the instructions, under the era-coherence gate (code change; this PR is docs-only).
- The §C2.2 orchestrator action note (supersede the sealed inbox-doctrine memory) — executes at
R8's landing, a runtime-memory act.
- PATHOS's OMEGA-floor blurb still describes Move 2 in its pre-reframe "solvency/token-budget"
wording — one curated line for the next PATHOS checkpoint (PATHOS is a hand-curated handoff file,
not amended from a mission PR).
## Appendix ADJ — the adjudication table (every finding, no silent drops)
**Duty discharged:** all findings present in the critique are adjudicated below. The critique's
header counts "26 findings"; its body numbers **F1–F8 and F10–F26 (F9 does not exist in the
document; F5 is the unification position)** — 25 numbered items, all here. Verdicts:
**ADOPTED** (into law as prescribed) · **ADOPTED-AMENDED** (into law, shape changed, reason given)
· **REFUSED** (with reasons). Tally: **22 ADOPTED · 3 ADOPTED-AMENDED (F1, F22, F24) · 0 refused
outright** — the mandate ordered the blockers' prescriptions adopted, and the shapers survived
scrutiny; where I bent one, the bend and its reason are visible.
| F# | Sev | Verdict | Where | Reason (one line) |
|---|---|---|---|---|
| F1 | blocker | **ADOPTED-AMENDED** | §C2.2 | derived grammar wins; `rejected`/`expired` become receipt **dispositions**, not fates — new fate words would need stored state on the target, breaking the derived-only law that made this grammar win; case = overlay; doctrine-memory supersession scheduled at R8 |
| F2 | shaper | **ADOPTED** | §C5.1 · A1 | `host:tier:name[@parent]` canon; TT §11 amended; the gate parses `tier == "main"` or `human:maintainer` — one grammar, one parser, one test |
| F3 | shaper | **ADOPTED** | §C2.3 | one stop enum; `stop_gate` mints no words; the check rides `focus(mode:"check")` |
| F4 | shaper | **ADOPTED** | §C2.3 | `MEMORY_AGE_HALF_LIFE`: one constant, one computing site, two renderings at most |
| F5 | position | **ADOPTED AS LAW** | §C2 | the four grammars + the three must-not-unify refusals, verbatim into the constitution |
| F6 | blocker | **ADOPTED** | §C3 | one ritual ("leave it warmer"), one gate (landing), five machine-owned beats, promotion = cadence; the sentence carried verbatim |
| F7 | blocker | **ADOPTED** | §C6 | kill/keep table whole (incl. `hot_blast`); census law (112 @ pin, battery-pinned); ESSENTIAL default face ≤40 ceiling; future-verb rule standing |
| F8 | shaper | **ADOPTED** | §C6.2 · A4 | §O.4.1 wins: the envelope is a response layer, never a verb; Move 1 re-read accordingly |
| F10 | blocker | **ADOPTED** | §C1.3 | all five budget clauses, the §O.12.4 numbers binding; the live dup + root-sprawl reproduced in THIS session's own packet — flow fix mandated, not just stock sweep |
| F11 | blocker | **ADOPTED** | §C7.1–.4 | stranger test as ORG-INV; `m1nd init --medulla` at host-neutral path; precedence in m1nd's own terms; hosts are data; OS tier honest |
| F12 | shaper | **ADOPTED** (option b) | §C7.5 | consent-deferred box birth: box now, git-travel after init's loud warning — TT-INV-8 amended by law, the maintainer's box preserved |
| F13 | blocker | **ADOPTED** | §C8.2 | promote step 2.5: origin-qualified evidence + delegated re-hash or `evidence_unverifiable` (X360's existing state); a medulla claim never reads fresher than it can prove |
| F14 | shaper | **ADOPTED** | §C8.3 | P3: `State: verified` or maintainer-sourced, checked inside the verb |
| F15 | shaper | **ADOPTED** | §C8.4 | evidence-union parents; merge-and-recite never re-phrase; confidence caps at max(children) |
| F16 | shaper | **ADOPTED** | §C8.5 · §C4 | letters = witness tissue; `evidence:` refusal at the write door, same mechanism class as the conflict-marker guard |
| F17 | shaper | **ADOPTED** | §C7.6 · §C1.4 | calibration at birth; the cap said in words; binding-trust vs verdict-cap visually distinct |
| F18 | blocker | **ADOPTED · SHIPPED 2026-07-05 (R15)** | §C9.1 | eviction gate hard at brain #5 or `all-brains`, whichever first; kill-9 battery case; interim named topology debt with its retirement. Shipped: LRU cap (4) + persist-on-evict in `ProjectBrainRegistry`, bound graph pinned, kill-9 warm-boot battery green |
| F19 | shaper | **ADOPTED** | §C9.2 | promote direction inverts at canon topology; contract-level battery survives the flip |
| F20 | shaper | **ADOPTED** | §C9.3 | loopback-only as named ORG-INV; non-loopback binds refused; auth required at the future remote door |
| F21 | shaper | **ADOPTED** | §C9.4 | invariants stated with door coverage until all doors route |
| F22 | shaper | **ADOPTED-AMENDED** | §C0 · §C11 | all four documentary laws adopted; amendment: the cited cross-branch dangle (MEDULLA §9.2 ↔ §4A.11) self-resolved at #267's landing — the policy stands because the class recurs |
| F23 | shaper | **ADOPTED** | §C5.2 | the shipped envelope block is canon; the §9.5.2 result-replacement carrier repealed |
| F24 | polish | **ADOPTED-AMENDED** | §C10 R0 | urgency adopted by carving the MED-INV-6 fix out as R0 (ships before everything); the literal M5b-before-M5a reorder refused — `tier` recall needs M5a's `Origin-Brain` labels, the dependency is real |
| F25 | polish | **ADOPTED** | §C10 R5 | Windows triage rung: fix or honestly demote — a gate described as blocking must block |
| F26 | shaper | **ADOPTED** | §C8.6 | the soul slot bound by seven constraints; the SOUL-PRD argues exceptions in writing or inherits |
*Every law above traces to a map JOINT (A–K) or a critique finding (F#); every count to a named
pin. Where this constitution is wrong, reality wins — and the correction lands here first, then
ripples out through §C11-style amendments, never by silent divergence.*
