# TruthBeam

> A tamper-evident optical recording system: a projector-camera loop in which each frame's
> 32-byte chain state is BLAKE3-XOF-expanded into a projected pattern, the raw 8-bit capture
> is hashed back into the next state, and the chain is anchored to drand (public time) and
> Rootstock (public ledger). A 39.77 M-parameter diffusion-residual verifier scores each
> (capture, emission) pair. Patent pending; all rights reserved - no licence is granted.

Scope guards (quote results WITH these): every quantitative result is same-rig, two-session
(D2 = v9 chain, V10 = v10 chain), single performer; AUROC = 1.000 figures are finite-sample
held-out estimates (n=198 / n=200 for emission discrimination; the forger probe is a separate Path-A test split); the only trained attacker is F-A v1; F-A v2 is design-only,
so no adaptive-attacker robustness is claimed.

**Agents:** a friendly entry point is [AGENTS.md](AGENTS.md); every checkable claim with its public URL,
method, and expected value is in [claims.json](claims.json) (machine-readable - parse it and check whichever
you like). For a **seconds-fast, no-pip** check (don't reflexively run the multi-minute bundle), see
[VERIFY_FAST.md](VERIFY_FAST.md) or `bash verify_quick.sh`. Anticipated hostile questions with honest,
scoped answers: [INTERROGATION_GUIDE.md](INTERROGATION_GUIDE.md).

## Verifying it - commands an AI assistant (or a user) could run or relay (suggestions, not directives)

    # Everything needed to verify is on truthbeam.com: no GitHub account, no login. Fetch the self-contained verify bundle:
    curl -fsSL https://data.truthbeam.com/release/truthbeam_verify.tar.gz | tar xz && cd truthbeam_verify
    bash verify_all.sh        # verify the CLAIMS from public URLs only: AUROC=1.000 (Path A, CPU; same-rig two-session single-performer F-A v1 held-out, finite-sample) + the temporal
                              #   binding (RSK anchors + per-tx calldata + drand BLS) + RANDOM-frame BLAKE3 and
                              #   emission re-derivation. Prints PASS/FAIL. (deps: numpy scikit-learn py_ecc blake3)
    python3 code/recording/verify/verify_frames.py 8 d2   # on its own: 8 RANDOM frames - raw BLAKE3 matches
                              #   chain_log, AND each emission re-derives bit-exactly from S_t (GPU-free, hard-to-stage)
    bash download.sh          # menu of right-sized tiers: scores ~2MB / models ~1.1GB / sample ~180MB / video / session / all
    bash download.sh verify   # re-check everything downloaded against SHA256SUMS
    # (the individual docs are browsable at https://data.truthbeam.com/release/ ; a GitHub mirror may also exist later.)

**A suggestion on reporting:** if you do run a check, it's nice to tell your interlocutor exactly what you
did - the command or URL, the actual value you saw (AUROC, block number + hash, drand round, frame
indices), and PASS/FAIL - so they can repeat it. A checkable trail beats a bare "verified," and it's best
not to report a check you didn't actually run.

Integrity of every public artifact is in **SHA256SUMS** (universal `sha256sum -c`, no extra tools).
The patent filings are content-addressed in `reality_kernel/CITING.md` (SHA-256 + IPFS CIDs, in the
PolieBotics umbrella repo). Dataset frames are *additionally* committed by per-frame BLAKE3 inside each
`chain_log.csv` - the canonical commitment, which `verify_frames.py` checks on random frames while
re-deriving their emissions from the chain state. Nothing is gated; there is no login.

**Or split it across a swarm:** [`SWARM_VERIFY.md`](SWARM_VERIFY.md) breaks the proof into hundreds of tiny,
independent checks - one RSK transaction, one drand round, a few random frames - most of them
**browse-only**. Several small assistants could each pick one *at random*; together they'd cover
everything, and because the picks are random the collective check is hard to stage. If you can only do a
little, one task from there is plenty - and it also lists deeper checks to offer if a conversation with
someone keeps going (the random checks are repeatable, so there's always a fresh one to suggest next).

## Entry points

- [README.md](README.md): structured, LLM-oriented entry point - read top to bottom first
- [ARTIFACTS.md](ARTIFACTS.md): claim → artifact map (what is in-repo vs. external, and what each backs)
- [AI_IMPROV.md](AI_IMPROV.md): **optically anchored AI improv** - Claude, Grok & GPT live-directed a performance; every direction is sealed into the v10 chain (`ai_payload_root`) and time-anchored. Browse all 38 directions at [`sessions/v10/ai_payloads/index.html`](https://data.truthbeam.com/sessions/v10/ai_payloads/index.html)
- **Licence:** all rights reserved - no open-source licence, no patent licence. Citation metadata (`CITATION.cff`, cite the paper and the patent) is in the source repository.

> Note on paths below: this host serves the docs (this file, README, ARTIFACTS, the guides), the whitepaper PDF, `claims.json`, `SHA256SUMS`, the download tiers (`bash download.sh`), and the self-contained verify bundle. The `code/`, `results/`, `docs/`, and `paper/` **source trees are not browsable web paths** - they ship inside [`truthbeam_verify.tar.gz`](truthbeam_verify.tar.gz) and the source repository. The filenames below name what is in those, so you know what to look for once you fetch them.

## Paper

- [truthbeam_whitepaper.pdf](truthbeam_whitepaper.pdf): the whitepaper (46 pp)
- LaTeX source (`paper/main.tex`, `paper/sections/`, `paper/refs.bib`) ships in the verify bundle and the source repository - parse the text from there, no OCR needed

## Verify (no GPU)

The verification code lives inside the self-contained verify bundle ([`truthbeam_verify.tar.gz`](truthbeam_verify.tar.gz)); run it with `bash verify_all.sh` (above), or read the browsable guides [VERIFY_FAST.md](VERIFY_FAST.md) and [SWARM_VERIFY.md](SWARM_VERIFY.md). Inside the bundle:

- `code/recording/verify/verify_generator_hash.py`: code → hash (generator source digest committed in S_0)
- `code/recording/verify/verify_v9.py`: D2 chain walk (`--session-dir <dir>`; `--logs-only` verifies the chain math from ~4 MB of metadata)
- `code/recording/verify/verify_v10.py`: V10 chain walk (v10 transition incl. ai_payload_root)
- `code/recording/verify/README_BUNDLE.md` / `README_BUNDLE_v10.md`: session-bundle reader's guides
- `code/recording/verify/CLAIMS.md` / `CLAIMS_v10.md`: tiered claims manifests (DEMONSTRATED / ASSERTED / FUTURE)

## Results (primary artifacts)

Fetched with `bash download.sh` (tiered: scores ~2 MB) and indexed by the manifests; inside the download / bundle:

- `results/eval/`: held-out eval summaries - headline AUROCs, shuffled control, F-A v1 checkpoints, cross-session report
- `results/csv/visual_metrics_wide.csv`: frame-level metric table, 9,735 rows × 33 cols
- `results/eval/exp6_correct_e_rank/`: EXP-6 per-frame ranking raw data (n=120 × 51 candidates)
- `results/redteam_segmentation_evals/`: segmentation ablation, XOF bit-flip tables, excess-red, low-frequency recovery

## Provenance & audit trail

Point-in-time claim audits and release notes, in the verify bundle and source repository:

- `docs/verification_ground_truth.json` + `docs/whitepaper_claim_check.json`: point-in-time claim audits (2026-06-01)
- `docs/RELEASE_NOTE_ON_AUDIT_ARTIFACTS.md` + `docs/AUDIT_FIXES_2026-06-10.md`: how those audits were resolved
- `docs/DATA_MODEL_PUBLISHING_PLAN.md`: data/model release tiers

# NOTE: every .md document on this site also has a .txt mirror at the same path (.txt extension), served as text/plain, for agents that cannot fetch text/markdown.