Tolowa Technology / 01

The right work.
The right model.
Proof it ran.

BATON is a measured routing layer for AI work. It keeps consequential decisions with strong models, sends bounded execution to the right harness, and leaves a record proving what actually happened.

See the routing modelInspect the proof standard
Claude Code, Cursor, and Codex can call Baton. Baton separates decision from execution, routes the work, and records a measured event.
calling clientClaude Code
calling clientCursor
calling clientCodex
BATON / 01
DECIDEintent · risk · reversibility
DOplan · lane · feedback
Claudejudgment
Cursorplan execution
Codexlocal machine
GLMbulk throughput
MEASURED EVENTrequested → executed → status → evidence
TYPEMeasured routing layer
CALL FROMClaude · Cursor · Codex
DEPENDENCYConductor not required
BEHAVIORFail closed
00 / THE FAILURE

The expensive mistake is not using an expensive model.

It is letting the wrong model make the decision—and having no evidence when routing quietly fails.

Frontier everywhere

High-cost judgment models burn time on bulk work.

Cheap model decides

Plausible execution follows a weak, unreviewed plan.

Silent fallback

The calling model does the work and still reports a routed run.

BATON v1 taught us the hard lesson: self-reported routing is not evidence. BATON v2 removed that claim entirely. Only the dispatcher writes the execution event.

01 / DECIDE VS DO

Route by consequence,
not by whichever app is open.

Four classes keep judgment, planning, execution, and throughput from collapsing into one expensive prompt.

01
Reserved

Judgment

Strategy, architecture, public actions, irreversible work

default laneClaudeKeep the decision and final review on a frontier judgment lane.
02
Plan required

Plan then do

Code and system changes where a weak decision creates rework

default laneCursorA strong model writes the plan. Cursor executes inside its boundaries.
03
Feedback loop

Mechanical

Shell, files, browser, desktop, and tests with immediate feedback

default laneCodexThe machine supplies the truth: commands fail, tests turn red, clicks miss.
04
Low blast radius

Bulk

Research, extraction, classification, synthesis, second opinions

default laneGLM / CloudflareHigh-volume work moves off the frontier without pretending it is free.
02 / DEEP THINKING

Think deeply where the decision deserves it.

Deep thinking is a control policy, not a token setting. Ambiguous or high-blast-radius work moves upward. Code execution waits for a plan. Low-risk volume moves down to the least expensive capable lane.

Strong model: decide what should be built.
Execution model: follow the plan within explicit boundaries.

Confidential and client work is hard-filtered away from models marked no-ZDR. Safety constraints are routing rules, not suggestions.

03 / ANY CALLING CLIENT

Baton is the layer behind the workbench.

Claude Code, Cursor, and Codex can each make a manual call. None of them is BATON.

CLAUDE CODE

Judgment and orchestration

Shapes consequential plans, calls measured lanes, and handles final judgment. Claude may be the calling harness rather than a subprocess.

CURSOR

Plan-bounded implementation

The new Cursor lane calls cursor-agent with explicit model attribution. It does not depend on Conductor.

CODEX

Local truth and manual dispatch

Uses codex exec for local-machine work and can call GLM or Cursor lanes. Missing token telemetry is marked unavailable, never zero.

CONDUCTOR ≠ BATON

Conductor coordinates agents and worktrees. BATON routes measured AI work. They can work together; neither requires the other.

04 / THE EVIDENCE

A run is a row,
not a story.

BATON records the requested lane beside the harness that actually executed. Missing steps, mismatches, empty results, and truncation become inspectable failures.

requested_laneWhat BATON asked for
executed_harnessWhat actually ran
duration_msHow long the real call took
statusDone, error, or truncated
token_provenanceObserved or unavailable—never invented
costReal cost; savings only when supportable
A claim has to survive the ledger.
1Planned step
2Dispatcher
3Real harness
4Event row
5Audit
SUPPORTEDone row per step · requested matches executed · usable finish
NO ROUTEzero rows
BYPASSmissing step
FAILEDlength or empty output
05 / PORTABILITY TEST

It left the lab.

A new Mac

Jeramey handed Baton to his brother as part of a portable workstation setup. The system became useful somewhere other than the machine where it was invented.

Cursor, correctly wired

The measured dispatcher gained a real Cursor lane through cursor-agent, with model attribution and fail-closed behavior.

Manual from any workbench

Claude Code, Cursor, or Codex can pass the baton. Conductor is optional.

06 / BOUNDARIES

What BATON is not.

Not a model

It routes work between models and harnesses.

Not an autonomous agent

It enforces a routing and measurement contract.

Not universal token accounting

Evidence is qualified by what each provider exposes.

Not guaranteed savings

Failed and truncated calls keep their cost and claim zero saved.

Not a silent fallback

Unsupported lanes stop instead of becoming the model already open.

Not Conductor

No worktree or handoff tool is required to make a measured call.

Tolowa Technology / 01 of 03

Baton works because the rest of the system remembers what matters.

See the technology collectionExplore MemoryExplore the Skills Library