OpenAI Build Week · Education track · July 2026

Scaffolding for Taste

The portfolio existed before the competition clock started. During Build Week, Codex with GPT-5.6 Sol helped turn that foundation into a more inspectable studio: new representations, guided evidence stories, reproduction briefs, and design rules that preserve the human critique behind the polish.

12 post-cutoff checkpoints 7 new reproduction guides 1 retained GPT-5.6 Sol thread 0 runtime AI calls
Follow the evidence Play the guided tour Watch the 2:45 demo Open the live portfolio Read the design memory
Eligibility line July 13 · 9:00 AM PT

Before 09:00

Foundation

  • Research portfolio
  • COGS 125 framing
  • Living heuristics
  • Reciprocal 3D room

After 09:00

Extension

  • Human/AI profile
  • Guided evidence stories
  • Paper Constellation
  • Case studies + QA
The line is the claim boundary, not decoration. Everything gray was already here; this page asks judges to evaluate only the extension.
Why

Fast AI-assisted production creates more options, but it does not make the judgment behind those options inspectable.

What

A research portfolio where experiments expose their origins, evidence, limits, reproduction paths, and reusable design lessons.

How

Codex accelerated bounded implementation passes; rendered critique and human decisions determined what survived.

What changed after the line?

The strongest Build Week story is not the number of commits. It is the sequence of representational decisions those commits preserve. The plot below uses horizontal position only for date; its three lanes name the kind of public artifact produced, not a score or ranking.

Selected evidence · date on the horizontal axis

The foundation existed before the cutoff.

The orange rule fixes the eligibility boundary. Equal-sized nodes are linked checkpoints, not measures of effort or merit.

Equal-sized checkpoint · lane = artifact role · x-position = authored date

  1. 01 · Draw the boundary

    Name the foundation before showing the extension.

    The portfolio, COGS 125 framing, living heuristics, first 2D/3D desk, activity workbench, and reciprocal cliff-room topology all predate 9:00 AM. The last named baseline, a765c5291, was committed before the window opened.

  2. 02 · Make the record readable

    Publish alternate representations without splitting the truth.

    d3f13be35 added the source-linked Human/AI research profile, publication catalog, and raw text alternatives. The canonical bibliography still owns bibliographic facts; the new routes reduce visual decoding work without making an AI-discovery claim.

  3. 03 · Guide the interpretation

    Let cadence, magnitude, and model evidence keep separate clocks.

    6b4b7bd59 added the scroll-led Build Rhythm story. f05f02b41 added the first case-study and reproduction layer, while b183e4b75 wrote the surviving constraints into design memory. The original workbench began on July 11; the guided teaching layer is the Build Week extension.

  4. 04 · Change the representation

    Build a desktop atlas, then stop shrinking it.

    855f1bce8 fixed the privacy-safe data contract; eeb0a5764 built Paper Constellation; and 6832a6a05 replaced the compressed mobile map with a chronological trail. The relationships and coordinates were human-reviewed; GPT-5.6 did not discover the research lineage.

  5. 05 · Make the process inspectable

    Turn playful experiments into things another person can study.

    Across f05f02b41, 1ec5d5f4f, 2070991b0, and the constellation work, seven experiments gained project pages and seven downloadable reproduction guides. The new contribution is inspectability, not a claim that every underlying interaction was invented during the week.

  6. 06 · Publish the proof

    Test the representation outside one attractive screenshot.

    739d2ba28 published paired desk evidence under a fixed capture rubric. 1ae1ead39 and c52de6932 closed responsive and cross-engine gaps. Post-cutoff desk work polished controls and evidence; it did not create the reciprocal room.

One record, two reading shapes

Paper Constellation is the clearest example of the Build Week method. The data contract stays fixed while the presentation changes around the reading task. Desktop supports spatial comparison; mobile prioritizes chronology and legible labels.

Desktop Paper Constellation atlas arranging accepted papers and anonymous future nodes across three research threads
Desktop atlas. Stable, reviewed coordinates support relationship tracing. Node position is not prestige.
Mobile Paper Constellation trail preserving three thread rails, accepted papers, and anonymous future nodes in chronological order
Mobile trail. The same record becomes a chronological path instead of an illegible miniature atlas.

Seven experiments gained receipts

The case-study layer makes a previously playful portfolio useful as a teaching artifact. Each page now exposes the same four questions: where did this begin, what changed, what evidence survives, and how could someone reproduce the pattern?

The loop GPT-5.6 accelerated

GPT-5.6 Sol was useful because it made bounded implementation and re-review faster. It did not supply the project’s research meaning or make the acceptance decision.

  1. 01Brief

    Name what a rushed visitor should understand.

  2. 02Codex pass

    Implement one bounded change with GPT-5.6 Sol.

  3. 03Rendered evidence

    Compare fixed routes, viewports, themes, and states.

  4. 04Human decision

    Keep, revise, or revert based on meaning and access.

  5. 05Design memory

    Write the surviving critique as a reusable rule.

Codex + GPT-5.6 accelerated

The iteration surface

  • Tracing routes and implementation history
  • Building bounded visual and responsive passes
  • Writing focused source and browser checks
  • Comparing failures across fixed states

Human critique decided

What deserved to survive

  • The research claim and eligibility boundary
  • Which visual relationships were defensible
  • Which directions to simplify or reject
  • Which lesson belonged in shared design memory

Reuse the design memory

The durable output is not a model label or a single visual effect. It is the human-authored checklist that future students, collaborators, and agents can inspect, challenge, and adapt.

Living design memory

Website design heuristics

A living markdown checklist for critiquing portfolio pages with clarity, restraint, and reflective practice.

Download Source 
Loading heuristics...

Exact evidence ledger and Codex provenance

Primary retained build thread: 019f652f-7154-7822-ad1c-daa5a066134b. The retained thread ran GPT-5.6 Sol at ultra reasoning effort. The model label records provenance, not isolated causality.

  1. a765c5291Named baseline: reciprocal cliff-room scene committed before the 9:00 AM eligibility line.
  2. d3f13be35Added the source-linked Human/AI profile, publication catalog, and raw alternatives.
  3. b183 · 6b4 · f05Recorded durable rules, added the Build Rhythm teaching story, and published the first experiment case studies and reproduction briefs.
  4. 855 · eebFixed the privacy-safe Paper Constellation contract and built the desktop atlas.
  5. 6832a6a05Recomposed the constellation as a chronological mobile trail.
  6. 739 · 1ec · 207Published paired desk evidence and completed provenance-rich experiment stories.
  7. 1ae · c52Closed responsive storytelling gaps and stabilized cross-engine visual checkpoints.

What comes next

The next step is a course-ready critique exercise: fixed viewports, a shared brief, comparable evidence, and a worksheet for deciding which observations should become durable design memory. Any claim about learning waits for an actual evaluation.

Credits

This page documents work completed for OpenAI Build Week with Codex and GPT-5.6. The site is a customized fork of the open-source al-folio academic theme. No OpenAI logo, third-party album artwork, or runtime OpenAI API is used on this page.