Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1,151 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

🧾 BriefLoop

AI-assisted business briefings you can question later.

English | 简体中文

Website: briefloop.ai · Contact: contact@briefloop.ai

OpenAI Build Week · 15-Minute Pilot · Getting Started · Weekly Loop · Troubleshooting · Reference Workspace · Contact

Writer entry: use /briefloop in Claude Code, or briefloop in a shell.


OpenAI Build Week 2026

BriefLoop was built with both Codex and GPT-5.6. They served different roles, and neither could make a model output authoritative on its own.

Participant How it was used Authority boundary
Codex The primary engineering environment for architecture decomposition, Python implementation, testing, adversarial review, failure analysis, documentation, and scoped repair. Codex is also the host for BriefLoop’s current Experimental SQLite runtime path. Codex could propose and implement bounded changes, but it could not approve its own work, merge changes, or authorize product and research claims.
GPT-5.6 Generated candidate claims and weekly-brief drafts for the controlled comparison. GPT-5.6 Sol with maximum reasoning in Codex, together with an Academic Research Skill, supported related-work research, candidate protocol design, and technical-report drafting. A separate GPT-5.6 Sol Pro discussion challenged the controls, falsification criteria, citations, and claim boundaries. Model outputs remained proposals. They were not automatically treated as evidence, accepted experimental results, or publication-ready conclusions.
Human maintainer Selected the research question, verified primary sources, defined invariants and acceptance criteria, froze the experimental protocol, accepted or rejected revisions, and approved merges and publication. Final architecture, experiment, risk-acceptance, merge, and publication authority remained human-owned.

The Academic Research Skill https://github.com/imbad0202/academic-research-skills organized the research workflow; it was not used as source evidence. The underlying papers, official documentation, repository artifacts, and primary publications remained the cited sources.

Judge quickstart — no API key required

Requires Python 3.12.

git clone https://github.com/Stahl-G/briefloop.git
cd briefloop
bash scripts/setup.sh
source .venv/bin/activate
bash scripts/demo.sh

The deterministic demo creates a public-safe reference workspace with a reader brief, Claim Ledger, Quality Panel, source appendix, and event-log excerpt. It does not call a model, fetch live sources, or prove output quality.

Evidence boundary

The Prompt, Skill, and BriefLoop comparison is reported only from frozen artifacts, hashes, and completed review records. This README does not claim that BriefLoop has already won the comparison, automatically resolves every knowledge conflict, or removes the need for human review. It does not prove semantic truth.


✨ In one sentence

BriefLoop is an open-source workflow for recurring business briefings.

It does not try to be “a better prompt.” It keeps track of the process behind a brief:

  • Which source did this number come from?
  • Which claims were registered?
  • Which checks passed or failed?
  • Which reader preferences were approved by a human?
  • What should be reused next time, and what should not?

When someone asks where a number came from, BriefLoop does not ask the model to improvise an explanation. It opens the ledger.


🧯 The problem

Many teams write the same kind of material every week:

  • market updates
  • industry briefings
  • competitor tracking
  • policy monitoring
  • equity research notes
  • investor-relations materials
  • executive briefings
  • project status reports

LLMs can draft these documents quickly. The hard part is not drafting. The hard part is trust.

Common problems:

  1. Sources disappear. A number enters the final brief, and two weeks later nobody knows where it came from.

  2. Small errors become confident conclusions. A weak source, stale data point, or misread paragraph can survive several editing passes and look authoritative.

  3. Feedback evaporates. “Lead with the business impact,” “do not use generic wording,” or “verify this category against the original filing” often stays in someone’s head instead of the workflow.

  4. Handoffs are painful. The real rules for writing the brief live with one experienced person, not in a visible process.

BriefLoop turns those hidden rules into a workflow that can be inspected, repeated, and improved.


👥 Who is this for?

BriefLoop is useful for:

  • analysts, associates, management trainees, IR teams, strategy teams, and market-intelligence teams who write recurring briefings;
  • teams that want AI-assisted writing to be traceable instead of merely fluent;
  • researchers studying agent workflows, human-in-the-loop systems, and auditable AI processes.

It is not the right tool if you only want:

  • a one-click pretty report generator;
  • an autonomous agent that publishes without review;
  • a system that proves every claim is true;
  • a way to make an externally generated AI report “safe” after the fact.

🧱 What BriefLoop actually does

Think of BriefLoop as a briefing pipeline with ledgers.

Step What happens Why it matters
1. Prepare materials Collect local files, source packs, or search results The model should not start from nothing
2. Register claims Important numbers, dates, entities, and claims are written into a Claim Ledger Later you can ask where a claim came from
3. Draft by roles Scout, Analyst, Editor, Auditor, and related roles work within boundaries Writing becomes staged work, not one giant prompt
4. Run gates Quality gates check source age, new facts, missing support, and delivery state Deterministic checks do not rely on prompt memory
5. Deliver by human action Final delivery is explicitly triggered by a human The system does not publish or bypass review
6. Inspect the audit trail Accepted actions leave SQLite receipts; JSON, Markdown, and HTML views remain replaceable projections Reviewability does not create a second runtime authority

The design rule is simple:

Agents may write and propose. Deterministic services accept authoritative effects, and delivery remains human-triggered.


📚 The four questions it keeps answerable

Question What BriefLoop records Where to look
Where is this run? Current Store revision, stage, blockers, and next safe action briefloop status, briefloop runtime next, SQLite ControlStore receipts
Where did each number come from? Claim Ledger entries, source dates, source appendix, gate findings claim_ledger.json, source_appendix.md, quality_gate_report.json
What took effect? Accepted strict requests, transaction receipts, and invocation lineage briefloop.db through supported status and runtime views
What guards delivery? Store-backed gate evaluations, package readiness, and explicit human approval Receipt-backed runtime actions and read-only status projections

Agents can observe and propose. Only strict requests accepted by deterministic services change the Store, and delivery stays human-controlled. A Store-native reusable-guidance or Improvement Ledger surface is not shipped yet.


📦 What you get from a run

A normal delivered brief is usually:

  • output/delivery/brief.md
  • output/delivery/<report-name>.docx

The SQLite ControlStore is the sole runtime authority for a fresh Codex run, stored in the workspace-local briefloop.db. Depending on the run stage, BriefLoop may also write replaceable audit and reader projections such as:

  • output/intermediate/claim_ledger.json
  • output/source_appendix.md
  • output/intermediate/quality_gate_report.json
  • output/intermediate/quality_panel.html

Those projections are not meant to be read by every end user, and they are not read back as runtime legality. They help a team answer questions, debug failures, review decisions, and hand the process to someone else.


🔎 A tiny example

A delivered brief might say:

This week, the sample PV module spot price fell 1.8% week over week,
the third consecutive weekly decline.

BriefLoop expects that claim to have a registered entry:

{
  "claim_id": "CL-0012",
  "statement": "The sample PV module spot price fell 1.8% week over week.",
  "source_id": "SRC-003",
  "evidence_text": "Example source excerpt showing the week-over-week price change.",
  "metadata": {
    "published_at": "2026-06-05",
    "source_title": "Example PV price sheet"
  }
}

If a source is stale, a number is unregistered, or an editor adds a new material fact late in the process, gates should surface the issue instead of letting it silently enter the final document.


🚀 Quick start

macOS / Linux

git clone https://github.com/Stahl-G/briefloop.git
cd briefloop
bash scripts/setup.sh
source .venv/bin/activate
bash scripts/demo.sh

scripts/demo.sh creates an API-free reference workspace with a final brief, Claim Ledger, Quality Panel HTML, Quality Summary, source appendix, and event-log excerpt.

For the shortest walkthrough, use the 15-Minute Pilot.

Package-index installs are not the launch path yet. pipx / PyPI packaging is tracked in pipx And PyPI Packaging Prep; do not use pipx install briefloop until release notes say a real package-index artifact has been published and smoked.

When you are ready to use your own materials, create your first briefing workspace:

briefloop onboard
briefloop init ~/briefloop-workspace --from-onboarding onboarding.json
briefloop runtime install --workspace ~/briefloop-workspace --runtime codex
briefloop run --workspace ~/briefloop-workspace --runtime codex

New runs use a fresh SQLite briefloop.db as their only runtime authority. JSON-only workspaces are not migrated or accepted as control input. The Codex adapter is Experimental: it freezes role_topology=single_session, executes separate Receipt-backed role invocations in one shared Codex session, and keeps JSON, Markdown, HTML, status and Quality Panel files as replaceable projections.

Common setup helpers:

briefloop init --from-onboarding onboarding.json <workspace>
briefloop runtime install --workspace <workspace> --runtime codex
briefloop run --workspace <workspace> --runtime codex

When a run needs a frozen set of local materials, the runtime accepts one strict 1–256 member source-pack request and records all members atomically in a single SQLite Receipt. The request binds the frozen manifest hash and preserves its source IDs, canonical URLs, and incident opened_at metadata. It never silently registers only the first file. The workspace-local Codex kit is also hash-bound: edits, deletion, extra role files, or symlinks fail closed before runtime work continues.

For workspaces that use the llm_decide source profile, source discovery runs through the runtime-host route (run --runtime codexruntime next).

Windows PowerShell

Windows does not require WSL or Git Bash. PowerShell is the recommended path.

winget install Python.Python.3.12

git clone https://github.com/Stahl-G/briefloop.git
cd briefloop

.\scripts\setup.ps1
.\.venv\Scripts\Activate.ps1

briefloop version

If PowerShell blocks script execution:

powershell -ExecutionPolicy Bypass -File .\scripts\setup.ps1

🤖 Runtime transition

The former Claude/Hermes/OpenCode/operator JSON-control paths are not entrypoints for a new SQLite run. Their retained assets are historical surfaces pending the separate legacy deletion unit. Do not use them to continue or migrate a JSON-only workspace.

For the active Experimental Codex path, install the packaged workspace kit:

briefloop runtime install --workspace <workspace> --runtime codex
briefloop run --workspace <workspace> --runtime codex

Good next reads:


🧪 Three ways to try it

Start from the product entry that matches your report:

Report job Start with Best for
Industry or market weekly industry-weekly recurring market updates, competitor tracking, policy monitoring
Management monthly management-monthly executive reviews, monthly operating updates, management briefing packs
Document review document-review reviewing a set of documents with page/source traceability
briefloop new industry-weekly ./weekly-brief
briefloop new management-monthly ./monthly-review
briefloop new document-review ./document-review

document-review prepares a document evidence review workspace. It does not make legal, compliance, or disclosure judgments.

Start with the three supported entries above. Other experimental entries are documented later for users who already understand the basic loop.

Path Best for What to do
Look once You want to understand the project quickly Run the demos and read the reference runs
Run once You want to test it on a few local files Create a workspace, add materials, run one brief
Use weekly You want a recurring workflow Configure sources, sections, reader preferences, and feedback

Optional demo commands:

bash scripts/demo.sh
bash scripts/demo-deep-dive.sh

The demos use synthetic materials. They show the evidence chain and gate behavior. Real use starts with your own workspace and sources.


🧭 Current status

Current version: v0.14.0

Current main entrypoints:

  • CLI: briefloop
  • Experimental SQLite-only Codex runtime: briefloop run --workspace <path> --runtime codex, followed by briefloop runtime next, invocation-start, invocation-accept|fail, and apply
  • Experimental one-shot web initialization: briefloop init <path> --web
  • read-only three-page report view: briefloop quality html --workspace <path> [--open]
  • experimental offline-shadow LAJ: briefloop experiments laj shadow-run and briefloop experiments laj present for public/synthetic advisory evaluation and standalone JSON/Markdown/HTML presentation; an explicitly supplied current-report-bound laj.json may be displayed read-only with briefloop quality summarize --laj-view <laj.json>

v0.14.0 completes the SQLite-only cutover and adds read-only interaction surfaces:

  • SQLite ControlStore (briefloop.db), accepted strict requests, Receipts, and ledger relations are the sole runtime authority. JSON-only workspaces are unsupported, with no importer, migration, dual read/write, or fallback.
  • the legacy JSON runtime-state stack and its dead consumers are deleted; legacy control files and report/status/Quality Panel exports are non-authoritative projections. Strict action, envelope, and human-request JSON payloads are revalidated against the Store and are not authority by themselves.
  • the packaged Codex Skill follows the exact Store-derived next action and Receipt-backed invocation protocol. It does not fall back to operator or another runtime.
  • the loopback init wizard and three-page HTML are read-only interaction surfaces. LAJ remains Experimental and NOT MEASURED; the Improvement Ledger page reports unavailable and cannot write guidance into a later run.
  • the v0.14 engineering changes were implemented and tested with Codex. Human maintainers authorized the merges and release.

The carried-forward supported report tooling and advisory quality surfaces include:

  • ReportSpec, ReportPack, ReportTemplate, and PolicyProfile contracts
  • workspace skeletons and deterministic PolicyProfile resolution
  • delivery / audit bundle manifests and clean bundle archives
  • supported industry-weekly, management-monthly, and document-review product entrypoints
  • bounded evidence_extract source/scope registration, source locks, logical page inventory seeds, and text-span seed registries
  • experimental SourceHub Lite setup for local files, RSS feeds, and runtime web-search handoff tasks
  • durable source evidence pack materialization and source taxonomy normalization
  • internal release-mode approval records
  • Quality Panel JSON / Markdown / HTML projections and audit-bundle integration
  • standalone experimental LAJ JSON / Markdown / HTML second-opinion artifacts and an optional read-only Quality Panel section for an explicit bound view; they do not affect workflow status, gates, finalization, delivery, repair, approval, recommended authoritative actions, or next-action authority, and evaluator efficacy is not measured
  • reader-quality warning/projection surfaces for template conformance, materiality selection, support-calibrated wording, citation profiles, coverage/omission, and scoped final-abstract diagnostics
  • trajectory-regulation decision narrowing for repeated retry/repair/blocker loops
  • proposal-only Semantic Support Auditor surfaces and human adjudication records that do not create support truth, gates, delivery approval, or release authority
  • public-safe reference, synthetic regression, minimal comparative evaluation, launch smoke, and release checklist guardrails

These features are meant to make report types, source evidence, default policies, delivery bundles, and operator quality visibility more product-like.

These features are still contracts, metadata, defaults, setup paths, approval records, deterministic warnings, and projection controls. They do not parse PDFs automatically, execute hidden web search or crawling, judge industry compliance, detect investment advice, verify internet rumors, claim IR/disclosure readiness, prove semantic truth, publish reports, or authorize public release.


🚧 What it is not

BriefLoop currently does not:

  • publish reports automatically;
  • bypass human review;
  • prove that a source semantically supports every sub-claim;
  • replace legal, compliance, investment, or disclosure judgment;
  • claim that generated content is ready for IR, SEC, or regulatory disclosure;
  • turn unapproved feedback into long-term memory;
  • promise one-click perfect reports.

The current core claim is deliberately narrow:

Traceability, not semantic proof yet. BriefLoop aims to make a briefing process traceable, reviewable, and accountable. Semantic proof and automatic judgment are future work, not current guarantees.


💡 Why this exists

Coding agents improved quickly because software work has infrastructure: tests, CI, git history, code review, and rollbacks.

Business briefings usually do not. A junior analyst is corrected verbally, and the correction disappears. A stale number enters a weekly report, and nobody can tell which step allowed it. A team learns what “good” looks like, but the learning stays informal.

BriefLoop brings software-engineering discipline into recurring briefing work: auditability, structured feedback, human gates, execution traces, and tests.

The goal is not to remove human judgment. The goal is to let humans spend more time judging, questioning, and advising, and less time re-checking the same preventable mistakes.


📖 Glossary

Term Plain meaning
Claim Ledger A record of important claims, numbers, sources, and dates
Source Pack The set of materials available to a run
Quality Gate A check that must pass before a stage or delivery can proceed
ControlStore The authoritative SQLite state for a fresh Codex run
Transaction Receipt A deterministic record of an accepted Store action
Orchestrator / 司乐师 The runtime role that coordinates stages and boundaries
Delivery Bundle The Markdown / Word files meant for readers
Audit Artifacts Intermediate records used for review, debugging, and accountability

🗂️ Useful docs

First-user path:

Architecture reference and contributor docs:


🤝 Collaboration

This project needs real scenarios more than it needs more concepts.

You are welcome to participate if you:

  • write recurring market, strategy, policy, IR, or executive briefings;
  • want to pilot BriefLoop on a real workflow;
  • research agent evaluation, human-in-the-loop systems, or auditable AI workflows;
  • want to contribute through issues, docs, tests, or example scenarios.

Start with a good first issue. Please read red lines and anti-patterns before contributing.

For human contact, see briefloop.ai or docs/contact.md. Use security@briefloop.ai for private security reports.


🙏 Acknowledgements

BriefLoop's initialization panel adapts interaction-layout and visual-system source from PPT Master, created by Hugo He and distributed under the MIT License. The adaptation is pinned to PPT Master commit 619a954. We are grateful to Hugo He and the PPT Master contributors for making the source available. See the packaged third-party notice for additional provenance and licensing details.


📄 License

MIT

About

Open-source loop engineering for auditable business briefings: claims, evidence, gates, findings, repairs, and human review.

Topics

Resources

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages