# Accuracy audit, research gaps and three next experiments

3 November 2026. All **67 records** in the expanded source library were reviewed again by three specialist agents, with parent follow-ups. This checks source identity, described process/capability, access and the limits of our interpretation. It is not a claim to have replicated every paper, read every book, verified every historical repo note, or completed a systematic literature review.

## What the audit found

The verdicts **before corrections** were 57 supported, 3 correction-needed, 5 partly-supported and 2 unverifiable-current-access. The three artist entries have now been corrected. The partly-supported book records distinguish supported metadata from unavailable full text. Access uncertainty is not evidence that a source's factual account is false.

| Area | Records | Result before corrections |
| --- | ---: | --- |
| Artist processes | 20 | 17 supported, 3 needed wording/date clarification |
| Books and institutions | 19 | 13 supported, 5 partial full-text/access support, 1 current-access unverified |
| Workflows, papers and context | 28 | 27 supported, 1 current-origin access unverified |

Every canonical source was attempted. The audit reports what the reader or repository actually returned, rather than equating a URL, search extraction, cached text, full book or successful runtime. A later parent retrieval resolved the FAE4 launch-page timeout. The e-David reader returned indexed primary text after an earlier direct-origin 404; its process claims are corroborated, while current-origin availability remains uncertain. Kadenze remains historical, closed according to first-party indexed extraction, and search-only for its canonical course page.

### Corrections and useful clarifications

- **ART05, Tom White:** the paper includes a small human label-matching survey as well as classifier recognition. Neither measures beauty or usefulness for Iana. [Paper](https://arxiv.org/html/1912.04217v1).
- **ART08, My Artificial Muse:** a collaborator documents a 2017 performance. The current project-page body still says 2016, so inception/premiere disagreement remains explicit. [Collaborator account](https://quasimondo.com/2017/06/13/my-artificial-muse/).
- **ART14, Ben Snell:** the artist describes computer-component powder mixed with clear resin and cast; our earlier statement that casting details were absent was too broad. A full fabrication specification is still missing. [Interview](https://www.phillips.com/article/42361496/media-and-materiality-in-conversation-with-ben-snell).
- **Books/access:** refreshes distinguish metadata, samples, open PDFs, blocked full text and historical programmes. The Grand Palais teaching/visitor PDFs are now retrievable. The Leonardo author's accessible manuscript is 13 pages; the published layout is five pages and its former UPF endpoint returned 502. FAE4 now links to its volume-specific page; it is not the latest series volume.
- **Diffusion Illusions:** fixed image targets, printed overhead transparencies and backlighting really are documented. Thick/spaced hand-painted glass, exact preservation of a painting, and coloured wall caustics are separate extrapolations. Paper and legacy notebook settings differ. [Project](https://diffusionillusions.com/).

The master JSON, lane records, source notes and downloadable library have been corrected. The original compilation is preserved in `before-audit/`; field-level before/after values are in `applied-corrections.json`.

## The important research gaps

1. **Optics and construction.** Our seven generated image studies establish visual concepts. None establishes how an opaque/translucent pigment, glass plane, light source and wall would produce the pictured result. We need a known scene or small material specimen, with recorded geometry and light placement.
2. **One consistent object across views.** Separate generated hero images do not prove a coherent three-dimensional assembly. An explicit contour/layer model should produce front, side and exploded views without changing forms between frames. Depth would be our authored interpretation, not a scan recovered from the photograph.
3. **Source fidelity and colour.** No colour-calibrated capture, physical dimensions or surface scan is documented for the five references. Gold, glossy paint and shadows are particularly ambiguous in a single photograph. We need a straight-on colour reference and dimensions before judging reproduction or fabrication accuracy.
4. **Artist control and evaluation.** The user reported that Iana liked Meridian Light; that is the strongest taste signal we have. It does not imply approval of all later images. A short artist review should identify what she wants retained, rejected or redrawn; record choices rather than treating technical novelty as success.
5. **Executable workflow and provenance.** Public repositories, notebook links and papers do not prove a compatible local setup. Dataset/demo releases can be missing. Exact model/settings, input rights, editable intermediates and dependencies need checking for whichever workflow is chosen. Five photos are not evidence that an artist-specific trained model will work well.
6. **Finishing and durability.** No source in the earlier list supplies a checked recipe for these exact painted-glass objects: adhesion, opacity, fading, spacing, support and display conditions remain open. A real sample and relevant material/conservation guidance should precede production or pricing.

## Three proposed visual experiments

These are proposals, not newly generated images or tested optical devices. Six additional primary references are recorded in `optical-sources.json`.

### 1. Floating Contours — Quiet Face I

Separate the original mauve, cream, blue and gold forms across shallow transparent planes. A slow sideways move reveals the depth; from the front, the original composition resolves. The next deliverable should contain a front view, two oblique views, an exploded layer view and a short orbit **from one scene**.

The simplest route is authored contour geometry with source pixels, using the existing renderer or a geometry renderer. AI can help explore materials while the shape model stays fixed. This is an accessible next step from Meridian Light and directly tests the depth gap. [Layered 3D](https://www.cs.ubc.ca/labs/imager/tr/2011/Wetzstein_SIG2011_Layered3D/) provides a physical layered-display precedent; a basic cut-layer scene would not implement its tomographic optimization.

**Pass mark:** front-view recognition, consistent layer order in all views, visible separation of at least three shapes, no invented forms. **Risk:** depth can spoil alignment and brightness. **Our first choice** for extending the liked direction.

### 2. Hidden Colour — Garden Before Noon

Keep the contours still in a mostly clear panel, then reveal different colour relationships as a polarizing viewer rotates. This adds surprise through light rather than growth or organic animation. The first visual brief can show three labelled states, but a small photographed swatch should establish the effect before an artwork-scale claim.

[Austine's medium account](https://www.austine.com/medium) documents an artistic precedent; the [Exploratorium activity](https://www.exploratorium.edu/snacks/polarized-light-mosaic) supplies a practical starting test, and the [physics paper](https://sbfisica.org.br/rbef/pdf/342302.pdf) explains the film/angle dependence. Use Iana's own shapes, not another artist's imagery.

**Pass mark:** a real swatch changes under a rotating polarizer, and at least two useful colour states preserve the source contours. **Risk:** film type and orientation matter; exact original colours are not guaranteed. A generated hue animation alone is not a polarization test.

### 3. A Painting Made of Light — The Listener

A shaped clear surface projects the curling motif onto a wall through concentrated light. Show the object and the projected motif together. Begin with a simplified single-colour target and an optimized, ray-traced result before attempting colour or a manufactured optic.

The [official Mitsuba tutorial](https://mitsuba.readthedocs.io/en/latest/src/inverse_rendering/caustics_optimization.html) supplies an inverse-rendering route; [EPFL's account](https://actu.epfl.ch/news/objects-that-sculpt-light/) provides a fabricated-object precedent. The documented colour tutorial uses a patterned RGB emitter: clear glass in ordinary white light should not be presented as an arbitrary full-colour projector.

**Pass mark:** the motif appears in a consistent ray-traced scene, with target/result comparison and recorded light/surface/receiver geometry. **Risk:** highest technical and fabrication uncertainty. A composited wall image must stay labelled concept art.

## Readiness and evidence files

Read-only local discovery found no Blender executable/app and no Mitsuba package in the default Python environment. This is an availability check, not a compatibility benchmark; no installation, model run, render, purchase, physical experiment or image generation occurred in this audit. A basic source-pixel parallax scene can use the existing renderer. Hardware, software versions and licensing would be checked when choosing an implementation.

- `audit-index.json` / `.csv`: coverage and verdict for every source ID.
- `artists-audit.md` / `.json`, `books-audit.md` / `.json`, `technical-audit.md` / `.json`: detailed evidence, URLs, access failures and proposed corrections.
- `applied-corrections.json`: exact changes to the source library.
- `optical-sources.json`: six new primary references.
- `proposed-experiments.json`: complete briefs and success criteria.
- `before-audit/`: untouched snapshot of the earlier source library.
- `verification.json`: consistency and preservation checks after integration.

Existing prototypes, source photographs and generated images are preserved. The research-library exports are intentionally refreshed; no art/site variation was built.