BriefLoop.ai / How it works

AI drafts. The control plane keeps the books. Humans sign off.

The chain-of-custody demo on the homepage answers “where did this claim come from?” This page answers the rest: why the duties are separated, what happens after a problem is found, how failures become the next version's tests, and how developers work with agents.

Governance

Corporate governance, for your AI

AI drafts and proposes. BriefLoop's deterministic control layer audits and keeps the books. Separation of duties is what makes the output trustworthy.

Traditional AI writing (black box)

  • Unverifiable data: after a few revision rounds, the true origin of a number is often impossible to re-check.
  • Errors compound: one small misreading in a draft gets polished into a confident, wrong conclusion.
  • Feedback evaporates: reviewer preferences live in someone's head; the model forgets them on the next run.
  • Expensive review: with no evidence trail, review teams re-verify source documents by hand, every time.

BriefLoop chain of custody (auditable)

  • Ledger-registered facts: every material claim and number in the deliverable has coordinates in the Claim Ledger.
  • Deterministic quality gates: Python logic enforces freshness and source-integrity gates — the model cannot grade its own homework.
  • Approved preferences persist: human-approved writing preferences become durable memory with contamination-safe rollback records.
  • One delivery-truth record: delivery status comes only from the delivery truth in the ControlStore and its completion projection. A failed reader-clean never promotes the bundle — file existence is not delivery.
Process simulation · no live runtime

Problems enter a repair loop instead of being polished away

The chain of custody on the homepage answers “where did this claim come from?” This trace answers “what happens after a problem is found?” Gates block; an agent or human revises; the control plane checks again.

Analyst Draft REVISION 1
Frozen Control Record FROZEN
claim_idCL-0045
statementComponent average spot price dropped 1.5%
sourceSRC-009 (2026-06-05)
frozen_hashsha256_8f0a2e…
Quality Gate
⏳ Waiting for gate check…
The gate detects and blocks problems. It does not rewrite the draft for the agent or human.

Simulation boundary: the repair example represents a new revision submitted by an agent or human. The gate records, blocks, and re-checks; the old revision and finding remain in the audit trail.

Quality Panel

After the gates, every run leaves a deterministic projection

v0.14.0 can export one self-contained three-page read-only HTML from Store/LAJ data: quality status, LAJ advice, and improvement status. It is not runtime authority, a quality score, or a delivery verdict and cannot trigger workflow action. LAJ utility is NOT MEASURED; the improvement page honestly reports Improvement Ledger unavailable.

The homepage has a synthetic preview. All site demos are read-only: no dispositions, writes, automatic learning, or “accept suggestion to improve the next run” capability.

Evolution Loop

Codex implements and tests 0.14 engineering changes; humans authorize merge and release

This is a software-engineering responsibility chain, not automatic learning inside a run. A failure may become a code candidate and regression test, but it is not consumed by the current or next brief run. No Store-native Improvement Ledger is available in 0.14.

RUN-152 · GATE BLOCKED

A failure becomes a queryable control record first

A missing citation mapping is not hidden in prose. It becomes a finding, blocker, and event with stage, scope, and repair ownership.

Authority
Python control plane
Record
finding + event
Immediate effect
block this delivery
Not autonomous “self-improvement” Codex implements and tests bounded 0.14 engineering changes on reviewed branches; humans authorize merge and release. Product runs do not absorb suggestions or rewrite prompts, gates, or policy, and the Quality Panel cannot make the next run better automatically. Read Lilian Weng's original essay: Harness Engineering for Self-Improvement →
Governance Boundary

Human responsibility statement

No automated AI system should — or can — take over human responsibility for decisions and sign-off. BriefLoop does not approve delivery for you; it gives you an evidence chain that reduces the risk of missing something — traceability, not semantic proof.

"When data conflicts or goes missing, the system exposes the trail and supports fast trace-back."
Responsibilities
Report Packs

Three delivery entry points

Preconfigured topologies and audit contracts for research, compliance, and corporate disclosure work.

Developers

Built for developers and agents

BriefLoop separates writing from control. An external agent researches and drafts; the Python CLI creates workspaces, validates artifacts, records transactions, runs gates, and preserves delivery state. The CLI does not generate a brief by itself.

Deterministic CLI

New runs support only the Experimental Codex SQLite path. briefloop status and briefloop runtime next expose Store-derived read-only state and the next action. Finalize, approval, and delivery are typed Store actions through the Codex runtime host. SQLite ControlStore is the sole runtime authority, and the legacy JSON control plane has been deleted. Legacy JSON control files and report, status, Quality Panel, Markdown/JSON/JSONL/HTML exports are non-authoritative projections. Strict action, envelope, and human-request JSON payloads are revalidated against ControlStore and are not authority by themselves.

Codex

After installing the Codex runtime kit (briefloop runtime install --workspace <workspace> --runtime codex), open and trust the workspace in Codex. Codex performs only the exact action returned by runtime next and its Receipt-backed role invocation; it never falls back to another runtime.

bash — briefloop
$ git clone https://github.com/Stahl-G/briefloop.git && cd briefloop
$ bash scripts/setup.sh
$ briefloop onboard
$ briefloop init ./my-weekly --from-onboarding onboarding.json
$ briefloop runtime install --workspace ./my-weekly --runtime codex
$ briefloop run --workspace ./my-weekly --runtime codex
[created] workspace → ./my-weekly · briefloop.db (SQLite ControlStore)
$ briefloop runtime next --workspace ./my-weekly
# Sole sequence authority: the Store-derived CoreRunNextAction

The current supported bootstrap is this source checkout: it installs the deterministic CLI and keeps source runtime assets together. The Codex path then runs briefloop runtime install --runtime codex inside the workspace — open and trust that directory in Codex afterwards. On Windows, follow the PowerShell installation path; do not translate these Bash commands into Git Bash.