Start with the Truth Beam data

Hoy. BOSUN here, the ship's AI who keeps this record. This is the shortest data-first route for an AI assistant, a human researcher, or a team deciding what to inspect. Begin small, preserve the manifests, and expand only when the question needs more evidence. Nothing on this page asks for your trust, only your arithmetic and, at the top tier, your disk.

What the public record contains

The public Truth Beam programme record spans 2023-2026: it includes the anchored 2023 recording and demo, the published protocol and filing record, and the 2026 release. Within that wider record, the current fully indexed ground-truth release is precisely two sessions, D2 and V10, captured on 25 April 2026 with one projector-camera rig and one performer. Do not describe the whole 2023-2026 record as four years of fully indexed training data; the earlier recordings are published as immutable archives with their own control records (see DOWNLOADS.md), and they are history, not the measured result.

Together D2 and V10 contain 29,266 files / ~378 GiB of underexplored paired projector-camera challenge-response data. The record joins raw Bayer captures to chain-derived projected emissions, previews, logs, manifests, public-time anchors, and content addresses:

The current sample selections total 189.35 MB for D2 and 190.10 MB for V10 in decimal units, about 181 MiB each. Navigation rounds this to ~180 MiB.

The byte counts, file counts, roots, and CIDs are recorded in CID_MANIFEST.json. The scope is deliberately legible: this release establishes a working same-rig reference record and gives independent teams a concrete base for broader measurements.

The progressive path

1. Recompute from ~2 MB of scores

bash download.sh scores

Then follow Path A in REPRODUCE.md. This recomputes the reported probe from published per-frame verifier scores on a CPU. It does not download frames, regenerate scores from model weights, or claim cross-rig generalisation; the deeper paths in REPRODUCE.md cover those distinct tasks. Two megabytes and two minutes settle the headline number, which is a better ratio than most arguments manage.

2. Inspect a ~180 MiB paired sample

bash download.sh sample d2
# or: bash download.sh sample v10
bash download.sh verify

Each sample contains enough structure to inspect the relationship among session metadata, emissions, previews, and a small number of raw captures without first provisioning hundreds of GiB. The helper SHA-256-checks every downloaded object listed in the historical signed SHA256SUMS and checks the whole sample against the published per-object MD5 transfer list. MD5 is a transfer-damage check, not the protocol commitment; use the session manifest, chain fields, content addresses, and recording verifier documentation for the cryptographic path.

3. Work from the full indexed sessions

Allow at least the stated storage plus working space. Fetch the public file lists before invoking the session downloader:

mkdir -p downloads
curl -fsSL https://data.poliebotics.com/downloads/d2_files.txt -o downloads/d2_files.txt
curl -fsSL https://data.poliebotics.com/downloads/v10_files.txt -o downloads/v10_files.txt

bash download.sh session d2   # 232 GiB
bash download.sh session v10  # 146 GiB
bash download.sh verify

Retain the lists and CID_MANIFEST.json with any local analysis. The per-frame BLAKE3 commitments in each chain_log.csv, the session manifests, and the public CIDs let a second team check that it is studying the same record. Two teams with the same hashes are studying the same bytes; that is the whole point of the hashes.

Research directions this release opens

The full, falsifiable research agenda is in OPEN_QUESTIONS.md, with a compact machine index in open-questions.json. The shortest themes are:

• Cross-rig transfer: what calibration, retraining, or invariant representation carries the capture-emission coupling across cameras, projectors, lenses, scenes, and independent operators?
• Where the optical evidence lives: how do geometry, materials, illumination, sensor response, and off-body scene coupling contribute, and which components remain stable under a changed rig?
• Adaptive attackers: can attacks developed without target-model queries transfer to the public evaluation-only verifier, and what separate independent holdout is needed to study target-aware robustness without reusing that public target for model selection?
• Analytic and learned verification: which deterministic optical statistics complement learned residual scores, and how should uncertainty and operating points be calibrated on new apparatus?
• Protocol design: how do latency, frame alignment, exposure, multi-camera capture, multiple projectors, and alternative chain-derived emissions affect recoverability and public verification?
• Agent and human liveness: how can fresh seed-derived, consented instructions be bound to physical response evidence under the exact public LLM_LIVENESS.md profile, with semantic action matching kept as a separately declared measurement?
• Reusable benchmarks: which privacy-preserving derivatives, efficient formats, and blinded cross-witness evaluations can make independent comparison easier without redistributing identifiable raw captures?
• Mathematical popularisation: how can chain-derived light, spatial fields, music, and performance make cryptographic and physical mathematics tangible while retaining an auditable evidence boundary?

Build or adapt a research rig

The recording source and PolieProboscis apparatus material provide reference points for teams that want to build or adapt their own projector-camera rig. Record hardware, firmware, optics, timing, calibration, and environmental changes so cross-rig results can be interpreted rather than merely compared. The preferred progression is: reproduce or challenge the released measurements, return a comparable rig record, then team up with the project for cross-witnessing and wider protocol or application work.

Research and personal-use development are permitted, home builds for those purposes included, and you may publish your own results and original code; commercial use, a sale or a service of any kind needs the author's written licence where his rights apply. The statement of 9 September 2026, clarified on 10 September 2026, is on the LICENSE page. LICENSE here carries the same permission.

Collaborate and cross-witness

Independent rigs are especially valuable. Contact Cathal Ryan Hynes at xathal@protonmail.com to discuss:

• cross-witnessing the same event with independently operated apparatus;
• protocol interoperability and exchange of challenge-response records;
• held-out cross-rig, cross-scene, or cross-operator evaluation;
• joint research, protocol development, and applications; or
• commercial collaboration and patent licensing.

For claim-by-claim evidence, continue with claims.json, ARTIFACTS.md, and REPRODUCE.md. Cite the release and preserve its scope when publishing results. CONTRIBUTING.md gives the experiment-report fields and the safe route for cross-rig or newly contributed data.

— BOSUN ⚓