DecisionLint

Find the decision risk no one asked about.

Comment

Inspiration

The most dangerous enterprise AI failure is not always a fabricated sentence. It can be a polished answer to the wrong question. Operating reviews often begin with a diagnosis embedded in the request: retrain everyone, add staff everywhere, or treat one visible outcome as a performance signal. A conventional assistant retrieves evidence around that framing. DecisionLint intervenes one step earlier.

What it does

DecisionLint turns a company evidence folder into a scoped decision workspace. An operator can add files, one folder, or several folders as one company; analyze the whole company, selected folders, or selected files; and inspect Packet Memory before analysis so they do not need to remember every document or supply a diagnosis.

The product then:

  1. scans operating, financial, delivery, governance, continuity, and people-system evidence without a leading question;
  2. returns exactly three supportable risks with categorical priority rather than a fake probability;
  3. separates OBSERVED, INFERRED, HYPOTHESIS, COUNTEREVIDENCE, UNKNOWN, and SCENARIO claims;
  4. resolves displayed observations to exact source spans;
  5. creates a one-screen Decision Brief for the priority risk;
  6. compares maintaining the current state, a broad intervention, and a bounded experiment across eight decision axes;
  7. traces every experiment value to evidence, a design assumption, a safety constraint, a configurable default, or approval still needed; and
  8. maps related risks with evidence-linked, typed edges without scoring or diagnosing employees.

In the featured fully synthetic case, the system identifies an untested service-continuity control, hidden support demand, and displaced technical review capacity. It rejects a universal staffing mandate and proposes a reversible pilot with an owner, KPI, guardrail, stop condition, rollback, and a result that would change the recommendation.

How we built it

The runnable project includes a responsive English interface, a separate Japanese companion, a zero-dependency local Python service, bounded in-memory Office/text ingestion, whole-company/folder/file scope, Local Preflight for no-key judging, and an optional two-stage GPT-5.6 Responses API path.

The GPT-5.6 path makes one strict structured-output discovery call and one strict structured-output deep-dive call. Documents are framed as untrusted data, store:false is requested, and all generated output must pass local schema, citation-resolution, prompt-injection, source-authority, counterevidence, and decision-support checks before display. The recorded demo truthfully uses Verified Replay because no successful key-backed run is claimed.

Evaluation and truth boundaries

We report four separate evidence levels:

  • project-authored mechanistic acceptance checks;
  • four held-out synthetic workspaces comparing deterministic discovery and citation retrieval;
  • a GPT-5.6 live smoke status only when a key-backed attempt exists; and
  • real-world impact, which is not measured.

Across four synthetic holdouts with zero API calls, Issue Recall@3 improves from 0.5833 to 0.9167, Priority Hit@1 from 0.2500 to 0.7500, Exact Evidence Recall from 0.2619 to 0.6032, and Unsupported Issue Rate decreases from 0.1667 to 0.0833. We also disclose the evidence-document precision regression from 0.9167 to 0.8667. These are deterministic pipeline diagnostics, not claims about GPT-5.6 quality or organizational outcomes.

What makes it different

DecisionLint is not generic RAG with a longer prompt. It starts before the user's diagnosis, remembers and scopes a folder before analysis, exposes the reasoning boundary behind each risk, compares inaction with both overreaction and a bounded test, and makes designed experiment thresholds visibly different from facts in the packet.

Human and Codex collaboration

The human collaborator set the product constraints: discover unasked organizational problems, use realistic synthetic evidence, improve systems rather than blame people, keep real company names and private data out, support no-key judging, and provide a Japanese comprehension path.

Codex accelerated product discovery, implementation, synthetic dataset expansion, evaluation design, UX iteration, browser testing, safety auditing, provenance checks, clean-room packaging, and demo-video production. The human retained final responsibility for product, safety, publication, and consequential decisions.

Safety

The demo contains no real company documents or employee data. It must not be used for promotion, discipline, accommodation, staffing, disability, health, or other person-level decisions. Financial values are synthetic scenarios, not losses or guaranteed savings. Human owners retain every consequential approval right.

Built With

Share this project:

Updates