Real products.
Room to explore.

What the recordings show

Product screenshots and videos were captured in October 2026. They show working interfaces, with captions describing the actions. Personal information is hidden where necessary; gallery images are selected safe-for-work artwork. Edited recordings disclose cuts or shortened pauses.

The homepage tour combines original-speed excerpts with new recordings of the working Studio and World. Its text alternative explains each scene. The World sequence is an illustration using a recorded answer; the Mac sequence is the actual exchange.

The teapot comparison uses two completed image-generation jobs through PanelForge. The first generated a new original; the second used it as the sole reference. The gallery retains their relationship, and the original file stayed unchanged. The screenshots review completed jobs in two separate conversations; they do not record inference. The optional motion study uses those two images in one actual WAN 2.7 render through an isolated PanelForge workspace. Its public preview omits the returned audio while preserving the video frames, timing and AI provenance metadata.

The Code History walkthrough imports four hand-authored sessions into an isolated copy of the real app. Its transcript capture includes an unreleased correction that hides invalid archive edit controls. Source archives remain intact; view preferences and login state are separate. No fresh model response or analysis was generated for those sessions.

The Workspace hierarchy film moves a fictional project in an isolated copy of PanelForge, then reloads it. It includes an unreleased fix that loads the destination picker. Saved-row and image-file checks confirm this fixture retained its contents. A small mask hides the image identifier; the static Forge branding remains. The phone capture shows the saved tree after the move.

The Grounded Analysis feature uses one actual model-analysis job on an authored fictional notebook conversation. Its recorded name checks rejected an uncovered submission, then accepted the model’s repair. Two unchanged saved statements are compared with the source after the run; one loses a project qualification. The film reviews the actual app after analysis, with identifier masks and fallback fonts. Its links identify a conversation, not an exact message.

The native chat recordings use an iPhone Simulator. The Attention recording uses the released Mac app. Pencil Page uses synthetic handwriting with actual model feedback. Conference Notes uses an isolated copy of the real application, fictional attendees and an actual analysis. Its recording reviews that analysis after it completed.

What you can try here

The native tool-boundary feature presents two recorded checks of the actual patched Codex app-server. Scripted local upstream messages drive a real fixture read, a same-thread app-server restart/resume and an excluded patch. A separate fresh thread grants the patch tool and creates its expected file; it does not widen the protected grant. The model/account catalog is synthetic, no real model inference occurs, and the shell grant is not a read-only filesystem permission. Gateway/MCP scopes are a separate boundary.

Forge World is an illustrated map of the projects. Its screens use real captures; its buildings, paths and movement are a designed way to explore the work. The Carry the answer scene explains the recorded Attention task with a moving answer slip and a fixed outcome. It sends no question or answer.

The Small Edition Press makes real 1080-square PNGs using the Studio’s adapted Kitforge renderer. Its miniature press and drying rack illustrate the process; the draft, finished pixels and downloads are real. Each of the two prints keeps its own pixels until you explicitly replace it. Words and prints stay in the current page session.

The Memory observatory searches three exact historical records in the same reviewed sample used by the Memory Lab. Visitors can inspect available source wording and download selected records as JSON. The desk is an illustration; selection does not verify a claim or establish that similar names should merge. Missing public excerpts do not mean a record has no source.

The Connections desk turns four recorded gateway results into locally generated SVG diagrams. It keeps each case’s status, observed forwarding count and available evidence together. Its scene illustrates separate questions; it does not trace live middleware or send a request. Role names are aliases for fictional supporting services.

The Memory Lab uses a reviewed technical sample from the knowledge graph. Its controls change that local sample. Source excerpts preserve the supporting wording, including historical names. The workbench diagrams explain configuration and workflow relationships. The Workbench, Recovery and Gateway labs replay actual controlled executions, with original output and downloadable receipts. Their controls move through saved results.

The scheduled-response example records an actual CLI closing and a new CLI resuming the same test session without another explicit query. The scheduled response matched the requested fixed text. Its separate classifier examples use synthetic transcripts; mocked engine tests and the live CLI run have different scopes. The live run did not exercise engine startup or delivery callbacks.

The three-engine example presents saved output from actual translation code given synthetic upstream messages. Its command is inert fixture text. Selecting an engine reveals a recorded mapping; it does not start an agent. The source notes identify the mocked transport used for Claude and the limits of the focused checks.

The train-platform cards are typeset presentations of actual tool responses from an offline run with historical synthetic travel data. Selecting a train reveals a saved result. The examples are labelled with their sample date and do not provide current travel information.

The Branch Lab follows a reference through the actual PanelForge interface and fork service in an isolated fixture. Answers are seeded rather than generated. Its comparison recording includes a locally committed, unreleased fix; the capture notes and receipt identify it.

The Pencil Page Timing Lab selects recorded output from the unmodified Swift scheduler. Drawing changes, check times and request outcomes are fictional inputs. Its separate native recording shows a manual review with real model feedback.

The native Question Lab uses real Claude inference through an isolated iPhone Simulator and a local API compatibility shim. An automated participant answers its built-in question. This capture includes an unreleased Claude-only correction, and the resulting plan remains unexecuted text. It is separate from Attention’s cross-device question flow.

The Attention Lab follows an actual Codex terminal task through the shared Attention MCP and a durable native Mac card. An automated participant returns a layout choice, and the same task builds a real downloadable page. Its film separates the live exchange from the later app-update check; only Mac arrival was observed.

The recorded voice turn uses actual PanelForge voice components and an isolated server pipeline with fictional file-backed stock speech, real Groq transcription, one Claude Engine model turn and OpenAI generated speech. Its two saved clips have native opt-in controls; playing them opens no microphone or provider connection. The original capture was process-muted and shows one chunk progressing to completion, not human audibility or the complete ordinary application UI.

The separate sound example uses PanelForge’s actual audio runtime and ordered player with three synthetic WAV notes. Its local sample events demonstrate quiet startup and ordered playback after a deliberate Listen action. It connects to no microphone or speech provider; separate PCM-stream findings are disclosed in its receipt.

The Listening Lab retains Saturation Listening’s actual waveform and timed-text components. Its original Danish story uses clearly labelled AI-generated Cedar narration, reviewed timing estimates and real waveform data. The recordings use the same fixture; the portrait recording is browser touch emulation.

The separate listening-plan check records the actual local app with an unreleased callback-scope correction. Its seeded six-day history, single authored question and duration metadata are fictional. Mark complete saves a chapter choice and changes the next plan; no listening or learning is measured.

The Archive Lab replays an actual isolated Photo Archive index rebuild using three synthetic PNG originals. A separate browser check calculates their hashes and can demonstrate a mismatch by changing a temporary memory copy. The original files remain available unchanged.

The Curation Lab replays actual PanelForge endpoint execution against a disposable database and fictional knowledge triples. It shows type-scoped name changes, retained source links and an immediate audited inverse. The interface is a portfolio exhibit; it does not make live graph changes.

These interactive examples run within the portfolio. They do not connect to private records or operate the underlying services.

What the history establishes

The Model Policy feature shows actual saved output from the unchanged tagged Python hook with fictional configuration. The same explicit request produces empty stdout without a default role and a JSON denial when a synthetic role changes the model. No Codex, child agent, model or provider process was involved; installed-hook integration and the separate Metaclaude guard were not tested.

The steering feature presents saved events from one real Claude Engine stream. A correction queued during a deliberately delayed fictional-file read, and the same stream returned that code in the changed format. The static sequence preserves the observed order and timing; a private allow-once hook and namespace restricted the read. No live service starts here, and this check does not establish PanelForge UI or database persistence.

The browser handoff uses the actual ownership hook, extension popup and bridge with two same-address tabs. A test session endpoint records the selected request; no agent reads the page or replies. The maintained-profile film uses a fictional profile in isolated PanelForge: one manual edit and a restore create new versions while preserving history. No model or automatic refresh runs in that recording.

The FindThat feature shows a real read-only message search inside one already-known technical review. Its selected finding leads to historical source correction and independent reread records, plus an exact tagged source excerpt. It does not show whole-archive discovery, a runtime screenshot test, current installation or human approval. Original private records are omitted.

The release explorer connects selected decisions with repository tags and source commits. Tags establish checkpoints in the code’s history. A dated capability or historical session does not, by itself, establish what is enabled in production today.

Case studies separate the source history from the workflow demonstrated in their recordings. Evidence drawers preserve useful detail without putting every technical qualification in the main story.

Files you can take away

The six-page selected-work PDF brings together reviewed product captures and writing from this portfolio. Its links lead back to the relevant stories and demonstrations. The PDF contains searchable text and embedded fonts.

Studio Lab adapts Kitforge’s actual rendering function so you can create and export a graphic here. Its sample PNG and MP4 files are actual browser downloads. The separate Kitforge sample is an actual export from the complete application. The conference PowerPoint was generated by the application and retains editable objects. Its companion PDF is rendered from that file. Fictional notes remain clearly labelled, and personal template-author metadata has been removed.

A few well-made building blocks

This site uses Astro, Three.js, React, DM Sans and Fraunces. Fonts are served with the site. Distributed library and font notices are available below.