---
---


# The Trace Ledger

The ledger is ContextStack's canonical work truth. Every organizational state
transition and execution attempt appends a record. Chat can present or steer the
work, but the transcript is never the only place where ownership, authority, or
outcome exists.

## What the destination records

- Signals, assignments, and the accountable Bot
- Commitments, deadlines, state transitions, and terminal dispositions
- Handoff requests and their acceptance, rejection, or renegotiation
- Approval requests, decisions, policy versions, and authority envelopes
- Orders, runs, attempts, tools, artifacts, and gate verdicts
- No-tool cognition and model-routing receipts
- Provider, model, cost, latency, retries, escalation, and constraint profile
- Evidence bundles, external effects, reconciliation, and accepted outcomes
- Learning proposals, promotion decisions, rollback, and autonomy changes

This is more than a run log. A completed cross-Bot outcome should be
reconstructable—who owned it, who contributed, why a route was chosen, what it
cost, which authority permitted the consequence, and what evidence established
acceptance—without reading chat.

## Three jobs, one record

**Audit.** The organization can explain how a lasting effect occurred and which
human or pre-authorization allowed it.

**Diagnosis.** A failed or stalled commitment can be reconstructed from recorded
state even when the exact model behavior is not reproducible.

**Measurement.** Outcomes can be segmented by work class, route, attempt count,
quality, operator intervention, and total cost. That is the evidence required to
route cheaper intelligence safely and graduate a narrow consequence class.

## Receipts make routing legible

The user should not have to choose a provider before delegating an outcome.
Gateway can buy the cheapest sufficient intelligence inside the lane's privacy,
quality, latency, and budget boundaries. Its receipt makes the actual choice and
any escalation inspectable afterward.

The important economic measure is not token price in isolation. It is cost per
accepted outcome at equal-or-better verified quality, including retries and human
intervention.

## Replay and governed learning

Recorded work can be replayed against a candidate gate, route, skill, or routine.
Only evidence-backed, reviewed, reversible promotions change future behavior.
The ledger therefore supports learning without allowing a model or memory store
to rewrite itself silently.

## Current limitation

The destination record above is not fully implemented. Run-level traces are
ahead of commitment, handoff, approval, and evidence-bundle records. Real-path
token spend is not yet fully instrumented end to end, and routing remains largely
static. Figures presented elsewhere as *modelled* are not ContextStack production
measurements.