ReasonWeave

Make reasoning visible.

Comment 2

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for ReasonWeave

Inspiration

Generative AI makes polished answers abundant. In education, that can remove the intellectual work that makes a question worth exploring: choosing a method, committing to a first model, evaluating sources, making something, and revising a belief.

ReasonWeave is for independent high-school and college learners age 13 and older. It is not a chatbot, homework writer, grading tool, or learner-profiling system. It is a finite studio for practicing agency around one real curiosity.

What it does

Every quest follows one bounded loop:

Spark → Choose → Predict → Investigate → Create → Reflect → Branch

The learner asks a question, chooses one of exactly three methodologically different routes, and commits to a prediction before evidence appears. In configured live mode, GPT-5.6 and OpenAI web search create a source-backed Evidence Lens that keeps Evidence, Inference, and an Open Question visibly distinct.

The learner then chooses one sourced finding and records what its sources establish, where their support stops, and whether the finding supports, challenges, or complicates the prediction. They turn that same finding into a concrete design choice, carry a short creation-anchor phrase into what they make, and reflect with “I used to think…,” “Now I think…,” and “I still wonder…”. Finally, they choose one of exactly three stronger questions as My next question.

That choice completes a finite Curiosity Map and unlocks a portable Discovery Card. It never starts an automatic recursive quest. The exported trace preserves the learner's prediction, evidence judgment, source boundary, design move, creation, reflection, and next question.

Why it matters for Education

ReasonWeave does not try to imitate a teacher or keep a student talking forever. Learner agency is structural: prediction precedes explanation; sourced evidence is not conflated with inference; the learner must judge one finding; the design must visibly follow from that evidence; reflection records a before-and-after model; and branching stops after three questions.

The result is not another AI answer. It is a portable trace of what the learner chose, thought, evaluated, made, changed, and still wonders. We do not claim that this prototype improves grades, retention, or curiosity. Its demonstrated behavior is inspectable.

How we built it

ReasonWeave uses Next.js, React, TypeScript, Zod, the OpenAI Responses API, Structured Outputs, OpenAI web search, and OpenAI moderation. GPT-5.6 has four finite roles in configured live mode:

  • propose exactly three diverse investigation routes;
  • turn one route into a safe, browser-completable quest;
  • synthesize web-grounded evidence with SDK-returned citations; and
  • provide reflection feedback tied to the learner's actual reasoning, followed by exactly three next questions.

Every model result is validated again by application schemas and stage-specific rules. Application code—not the model—owns route count, prediction order, citation associations, map size, safe activity constraints, state transitions, timeout recovery, and seeded/live provenance. Evidence can be admitted only when a completed web-search call returns its URL. Unsourced reasoning must be labeled Inference or Open Question.

The public deployment intentionally runs seeded-only: no API key is deployed, live generation is disabled, and the complete pre-generated journey is labeled visibly. A bounded credentialed local evaluation ran on two synthetic topics with a test-only GPT-5.6 Terra override: 2/2 fixtures and 126/126 checks passed, and the NOAA, CDC, PubMed, and NINDS sources returned for those topics were manually reviewed. We do not generalize that narrow result to deployed live behavior, all ten topics, moderation calibration, or educational efficacy.

Design

The interface is an editorial science studio rather than a chat transcript. Each stage has one clear job, a growing map keeps the finite journey visible, and the final reveal makes the learner's reasoning—not an AI avatar—the central visual artifact. The map has a text outline, the route and quest controls are keyboard operable, and the final Discovery Card can be copied or downloaded as Markdown.

How Codex contributed

Codex was the primary engineering and technical-operations environment. It helped translate the learning idea into enforceable product rules, implement the vertical slice, build deterministic ten-topic fixtures, create browser and accessibility tests, debug the restored-session focus handoff, define the server-only provider boundary, validate citation allowlisting, and prepare the release evidence.

The key product decision Codex helped surface was simple: GPT-5.6 may expand the space of inquiry, but software must preserve the learner's agency and the integrity of the trace. The finished application checkpoint passes 46 test files / 634 unit tests, 145/145 deterministic fixture checks, a production build and bundle budgets, a 28-pass browser matrix with deliberate project-scoped skips, and a dedicated no-key judge path.

Challenges

  • Preserving learner agency while still providing useful AI help.
  • Associating evidence only with URLs actually returned by web search.
  • Making generated content bounded, testable, and safe enough for a reliable demo.
  • Building a finite visual map that remains keyboard accessible and recoverable without SVG.
  • Keeping a polished seeded path without ever presenting it as a live model call.

Accomplishments we're proud of

  • The prediction gate is an application invariant, not a prompt suggestion.
  • The Evidence Lens keeps evidence, inference, and uncertainty visibly distinct.
  • The learner must explain how a selected finding shapes a design choice.
  • The final map is generated from actual session state and has an accessible text outline.
  • The complete fallback remains useful while being visibly honest about its provenance.

What we learned

The difficult part of educational AI is not generating more content. It is deciding which parts the model must never take away from the learner. Once those boundaries became state-machine rules—predict first, inspect sources, judge evidence, create, reflect, then stop—the experience became both more educational and more testable.

What's next

Next steps are broader topic and moderation calibration, consent-safe learner and educator walkthroughs, and deployment controls for any future live-model release. Accounts, grades, LMS features, analytics, and learner profiling remain intentional non-goals.

Built with

Codex, GPT-5.6, OpenAI Responses API, OpenAI web search, OpenAI moderation, Structured Outputs, Next.js, React, TypeScript, Zod, Vitest, Playwright, and axe-core.

Built With

Share this project:

Updates