Inspiration
What it does
How we built it
Inspiration
Weekly notes mix facts, feelings, copied messages, uncertain dates, and later corrections. A fluent AI can easily smooth those gaps into invented certainty. Mind Gym Journal Builder was created around one rule: truth before form.
The goal is not only to generate a polished journal. The goal is to preserve what the user actually wrote, make uncertainty visible, and keep the human responsible for every final decision.
What it does
Mind Gym Journal Builder turns one week of raw notes into a reviewed and traceable journal through four stages:
- Source — the original text is stored exactly as written, made read-only, and protected with a SHA-256 checksum.
- Chronology and issues — GPT-5.6 proposes structured chronology records supported by exact excerpts from the source. Contradictions, unclear dates, missing context, and unsupported interpretations remain visible.
- Human review and Journal — the user confirms, corrects, rejects, excludes, or preserves uncertainty for every relevant item. Corrections are stored separately as immutable K records. Only reviewed material may enter the Journal.
- Approval and export — explicit approval creates an immutable final snapshot and unlocks a Markdown export with endnotes tracing each paragraph back through chronology, optional corrections, and the original source.
AI proposes. Evidence and the human decide.
How we built it
The application is a local Streamlit and Python prototype with SQLite storage. Its six-table schema and 21 database triggers protect the original source, AI analysis, review decisions, corrections, and approved Journal against late modification.
The live workflow uses the official OpenAI Responses API with structured output validation. Before any source is transmitted, the application displays a Slovak privacy warning and requires explicit consent. Returned records must contain valid source IDs and exact supporting excerpts before they can be stored.
The public repository also contains a deterministic offline demo using fictional data. It exercises the same production schemas, review gates, provenance links, approval, and export without an API key or network request.
How we used Codex
Codex helped build the project milestone by milestone: repository structure, SQLite persistence, ingestion, structured analysis, review workflow, Journal generation, approval, export, integration hardening, automated tests, documentation, and privacy audits.
Each milestone had explicit acceptance criteria and was completed before the next one began. Codex also helped create the fictional public demo, verify that private journal material was removed, and prepare the README and submission materials.
How we used GPT-5.6
In live mode, GPT-5.6 proposes the structured chronology and the reviewed Journal draft. It is never treated as the authority for events. Exact source evidence, schema validation, review gates, and human approval constrain every model-generated result.
A reference Journal may guide tone, but it cannot introduce new events. Rejected or excluded material is never sent into Journal generation, and unresolved material is omitted unless the user explicitly approves uncertain wording.
Challenges we ran into
The hardest challenge was making traceability a real system property instead of a promise in the interface. Validation in Python was not enough, so important invariants were also enforced directly through SQLite triggers.
We also had to preserve the exact original source across database close and reopen, handle safe migrations, reject malformed model output without partial writes, and create a public demonstration without exposing personal journal content.
Accomplishments that we are proud of
- Complete source-to-export workflow with human review gates
- Immutable source, corrections, decisions, and approved snapshot
- Paragraph-to-evidence endnotes
- Fictional offline demo requiring no API key
- 89 automated tests passing
- SQLite integrity and foreign-key checks passing
- Privacy-safe public repository and reproducible setup instructions
What we learned
Human-in-the-loop AI requires more than a confirmation button. It requires explicit states, immutable evidence, reversible decisions before approval, visible uncertainty, and technical barriers against silently rewriting history.
We also learned how effective milestone-based development with Codex can be when every stage has narrow scope and testable acceptance criteria.
What's next
Future versions could offer faster structured daily input, optional mood and craving tracking, easier weekly review, and a more accessible interface for people who do not want to write long daily notes. Deployment and additional export formats would come only after preserving the same privacy and provenance guarantees.
Log in or sign up for Devpost to join the conversation.