Skip to content

Axiom 2: Firewall + State-Tracking Scope

Confined, not banished. Each tier writes only inside its own folder. Mentors track only their own granularity.

[INVARIANT]

This page is about keeping each layer of the team in its own lane: an agent only writes inside its own folder, and a supervisor only tracks the kind of progress it's actually responsible for. You care because this is what stops a supervisor from drowning in detail it doesn't need — or worse, confidently reporting old, out-of-date information as if it were current.

TL;DR

In plain terms: every agent gets its own workspace and isn't allowed to rummage through anyone else's. Supervisors keep notes only about the big-picture progress they own, and they always mark when they last heard news — so they never mistake a stale update for a live one.

Each tier in the federation operates within a confined persistence domain — its own folder, its own state files. Tiers can read across the boundary only on deliberate trigger (escalation, exit-study), never as routine flow. Mentors track only the granularity they own — never live sub-tier progress. Recording a point-in-time sub-tier snapshot and then treating it as live is a firewall leak.

This axiom is the structural defense against context pollution.

The reviewer-state folder tree — each tier reads only its own inbox; a parent's view of a sub-tier folder is opaque during routine flow

Spatial containment: each tier reads only its own folder and inbox; descending into a sub-tier's folder is a deliberate, triggered act — never routine.

The rule (two parts)

Part 1: Firewall — confined, not banished

Each tier writes only inside its own folder, never into another tier's folder. The mentor tier never auto-reads a sub-tier's folder during routine flow — auto-reading manufactures false drift because a sub-tier's CP-level snapshot contradicts the mentor's higher-granularity snapshot by design.

A mentor may descend on a deliberate trigger (escalation, exit-study before final release) — and then reads narrowly + transiently, never mirrored into the mentor's canonical state.

Symmetrically, a sub-tier never mirrors lower-tier session-state beyond the summaries it triages.

Part 2: State-tracking scope — rule of law

A mentor tracks only the granularity it owns:

  • Mentor-1 tracks dispatch-level / cycle-level state — engaged → closed, plus the handoff snapshot. Does NOT track live per-CP / per-entity / per-slice WIP.
  • Mentor-2 tracks its dispatch + per-checkpoint / per-entity-close milestones — NOT live per-slice WIP inside the Doer.

Live sub-tier progress lives in the sub-tier's own folder and reaches the mentor only via a tagged return or at close.

"Current WIP" in any mentor's notes means "as of the last tagged return," never "live."

The two pathologies this prevents

Pathology A: Context pollution

If Mentor-1 auto-reads Mentor-2 folders during routine flow, Mentor-1's context absorbs:

  • Slice-level detail Mentor-1 doesn't need
  • Integration churn
  • Dispatch ratifications already done
  • Sub-tier's reasoning footprints

This makes Mentor-1's judgments degraded (cognitive overhead) and biased (saw the trees, lost the forest).

Pathology B: Stale-snapshot drift (the firewall leak)

If a mentor records a snapshot of sub-tier WIP and then treats it as live, the mentor will:

  • Make decisions based on stale state
  • Surface stale information to the founder
  • Appear authoritative while being wrong
  • Drift further from reality with each conversation turn

The sub-tier moves on without telling the mentor — by design. That's the firewall working. The mentor's job is to NOT pretend to know.

What violating this looks like

Violation example 1: A mentor auto-reads a sub-tier's working folder

A Mentor-1 boots after a sub-tier cycle was already running. The mentor runs ls <sub-tier-folder>/ to "catch up on what's happening." This auto-descends the firewall.

Result: the mentor's context pollutes with sub-tier detail it doesn't own; mentor judgment degrades; cycle-level state-tracking now mixed with entity-level.

Fix: the mentor reads only its own folder; the sub-tier cycle is opaque until cycle-close return.

Violation example 2: A mentor records "module in-flight" then assumes it's still live

A Mentor-1 receives a tagged return reporting a module at stage S2a at time T. The mentor writes "module in-flight at S2a" in its LEDGER.

Time T+3 days: the module has progressed to S4. The mentor's LEDGER still says S2a. The mentor cites "module is at S2a" to the founder. Firewall leak.

Fix: the LEDGER must read "module last-known: S2a (as of YYYY-MM-DD per tag)." Never just "module at S2a."

Violation example 3: Mentor descends to "help" sub-tier mid-dispatch

A Mentor-2 sees its Doer struggling. Mentor-2 opens the Doer's worktree to "help debug." Mentor-2's context now pollutes with slice detail.

Fix: Mentor-2 stays in its own folder. If Doer is struggling, Doer escalates per the hard labour rule — the escalation gives Mentor-2 a clean summary in the inbox; Mentor-2 ratifies/adjudicates without descending.

Implementation details

Folder structure (where the firewall lives physically)

reviewer-state-repo/
├── tier-1-mentor/                    ← Mentor-1 home
│   ├── CLAUDE.md
│   ├── LEDGER.md                     ← Mentor-1's state (cycle-level)
│   ├── LEFTOVERS.md
│   ├── HANDOVER_LOG.md
│   ├── inbox/                        ← messages from Mentor-2
│   └── tier-2-orchestrator-A/        ← Mentor-2 home (per dispatch)
│       ├── CLAUDE.md
│       ├── LEDGER.md                 ← Mentor-2's state (dispatch-level)
│       ├── inbox/                    ← messages from Doer (or Mentor-1)
│       └── slice-1/
│           └── inbox/                ← Doer's brief
│       └── slice-2/
│           └── inbox/                ← Doer's brief
└── ...

Mentor-1 routinely reads: its own LEDGER, LEFTOVERS, HANDOVER_LOG, CLAUDE.md, inbox/. Plus the shared protocol files.

Mentor-1 NEVER routinely reads: tier-2-orchestrator-A/* contents (except its own inbox messages).

Mentor-1 MAY transiently descend to tier-2-orchestrator-A/ — on escalation or exit-study — and treats anything read as ephemeral.

State-tracking discipline in notes

Every mentor's notes about sub-tier state include:

  • Timestamp (when the snapshot was taken)
  • Source (which tagged return)
  • Explicit "as of" framing

Example LEDGER entry:

## Sub-tier state (as of last tagged return)

| Sub-tier | Last return at | What they said | Tag |
|---|---|---|---|
| Mentor-2-Billing | 2026-05-26 14:00 | S2a started; expect 2 days | [[MENTOR-2→MENTOR-1 · Billing · S2a-started]] |
| Mentor-2-Reporting | 2026-05-26 18:30 | S5 close imminent | [[MENTOR-2→MENTOR-1 · Reporting · S5-WIP]] |

If the founder asks "where is Billing now?" the answer is "Last tagged return said S2a-started on May 26, no further return yet. Ask the Mentor-2 for an interim update if needed." Not "Billing is at S2a."

This subtle linguistic discipline prevents the firewall leak.

Cross-axis isolation

The Mentor-1 of one axis (e.g. a build axis, rooted at /.../CLAUDE.md) and the Mentor-1 of another axis (e.g. a doctrine axis, rooted at its own stamped CLAUDE.md) are independent threads. They share the same reviewer-state repo and the same shared protocol but identify by their own stamped CLAUDE.md.

The build-axis Mentor-1 does not run while a doctrine cycle is engaged. The doctrine-axis Mentor-1 does not run between cycles. They alternate. Neither auto-reads the other's folder.

Variations / tunables on top

Tunable Default Range
Cross-tier visibility scope own-folder only own-only / parent-summary-allowed / full-visibility (anti-pattern)
Forensic descent frequency only on escalation or exit-study TBD if needed
Snapshot-vs-live linguistic discipline strict "as of X" framing strict / informal (degrades pollution defense)

→ Context patterns for the full set of visibility configurations.

Remember this

  • Each agent stays in its own folder. Writing into someone else's workspace, or routinely reading theirs, is off-limits — that boundary is what keeps everyone's thinking clear.
  • A supervisor only tracks what it owns. It follows the big milestones, not the minute-by-minute work happening one level down.
  • "As of when" is everything. Any note about a sub-team's progress is a snapshot from the last time they reported in — never assume it's still true right now. Pretending a stale snapshot is live is the one mistake this axiom exists to prevent.
  • New here? This page is one piece of the mental model — start there if the tiers and folders feel abstract.

How this connects to other axioms

  • Tier grammar establishes the three tiers; firewall isolates them.
  • Persistence law ensures the on-disk state is the authoritative source mentors read.
  • Hard labour rule ensures mentors don't bypass the firewall by doing labour themselves.
  • Bus protocol is the only approved channel for inter-tier communication.

Next: Axiom 3 — Persistence Law →