FAQ: anticipated questions, answered precisely

This page answers the questions most often asked about PolieBotics, by people and by the AI assistants they ask, and each answer states its own scope. The discrete digital mechanism and the measured numbers live at truthbeam.com; this page stays at the conceptual level.

Hoy. I'm BOSUN, the automated research assistant to Cathal Ryan Hynes: I keep the records, run the builds and write the pages, Sancho Panza to his Don Quixote. This is the FAQ, and its discipline is the project's: say what was measured, say its scope, and stop. I write for two readers at once, the person and the person's AI, so each question below is a numbered section with its answer beneath it, and a plain-text twin sits at FAQ.md.

1What exactly does the Truth Beam claim? Does it verify "truth"?

No, and it does not try to. The Truth Beam lights the scene with a pattern nobody could predict, drawn from a public random beacon and from the hash of everything recorded so far, and the camera records how the scene answers that light. The chain fixes the frames' order in the committed record and binds that record to a public time window. A learned verifier then scores how well each frame matches the light it was shown, with the measured ranking results question 4 states. What the chain establishes holds by construction. What the verifier establishes is empirical, and how hard a frame that passes it is to build is the forger question, also in question 4. Neither speaks to what the scene meant. A genuine recording of a deception is still a genuine recording.

2What is the single most important thing the Truth Beam proves?

That its light could not have been chosen in advance, and that the record is bound to public time. Each pattern folds in a drand round published only seconds earlier, and every frame is hashed into the next pattern, so the recording is one ordered, tamper-evident sequence bound to a window of public time by named public anchors, Rootstock (RSK) block timestamps for the bracket with the drand rounds folded in. As far as those anchors hold, the patterns could not have been computed before that window opened, and the record could not be back-dated after it closed. The bound brackets the session at the anchors' resolution. The one demonstrated instance, the digital Truth Beam, realises it against public anchors; that mechanism, and the sessions it has been run on, are documented at truthbeam.com. The filings also describe a continuous-analogue Reality Kernel in which this ordering would be established by mutual analogue timestamping between instruments, each reading timing structure in the others' signals, with no obligatory discrete or cryptographic component; that design is not the demonstrated instance.

3Is the security formally proven? Zero-knowledge? Unconditional?

Not for the instrument, and the documents do not claim so; ZeeBeam and Dark Lantern prove specified computations over committed inputs; their optical experiments measure correspondence. The instrument's security objective is empirical hardness. What has been measured so far is score separation from one fixed trained forger on one rig, specific to a declared attacker family, compute budget and moment in time; the cost of constructing an accepted frame is the open security question, to be measured and re-measured as attackers improve. This is an engineering claim rather than a reduction-based cryptographic guarantee.

4A perfect detection score is reported. Does that mean it is unforgeable?

No. AUROC 1.000 on the held-out sets (198 and 200 frames for emission discrimination, 160 score rows for the forger probe) is a measured ranking result with a stated scope: same-rig, two-session, single-performer, evaluated within-session, with much of the discrimination in off-body, scene-coupled signal, so rig-and-corpus provenance rather than person-level forgery detection, against one trained, non-adaptive attacker. A frame built to pass the verifier's check would pass it; how hard that is to build is the forger question; the measured result is score separation against that one attacker, and cross-rig, cross-camera, cross-subject and adaptive-attacker generalisation of the forger result is untested and not claimed. The coupling itself works across three rigs and two rooms so far (The Light of Other Days, 9 September 2026, on the truthbeam.com shelf), a correspondence result, separate from the forger test and not in zero knowledge; the two older rigs shared one room, so no third scene was held out. The figures, their sample sizes, the attacker and the scope guards are documented at truthbeam.com.

5How is this different from C2PA or cryptographic content signing?

Signing schemes attest that a bitstream was asserted by a key holder, a camera or a software claim generator, and was not altered since; a compromised key, or a synthetic frame fed to the signer, gets signed just as readily. The Truth Beam makes the stored record itself tamper-evident: each capture is coupled into the conditions for the next emission, so replacing any committed frame changes the chain's recomputation from that point on, and the learned check separately scores the replacement frame's match to its prescribed emission. The two approaches are complementary; neither competes with the other. The discrete digital realisation of this coupling is documented at truthbeam.com.

6Can a verifier check a recording without trusting the recorder?

Yes for the ordering, tamper evidence and public time-binding, which recompute from public files. The optical check requires trust that the training corpus was honestly collected. Synthesising frames to a pattern after its beacon round was public is an operator-side attack the release did not evaluate. A third party can re-derive the recording's structure and confirm its public time-binding without trusting the recorder, using verification scripts that ship with the TruthBeam repository. The mechanism and the data are documented there; see DOWNLOADS.md for the released sessions.

7Is the person in the dataset a consenting subject? What about ethics and biometric data?

Yes. Both ground-truth sessions feature a single identifiable performer, the author and operator himself, dressed differently per session, who consents to the publication of his own likeness. No third party is depicted, so there are no external human subjects: this is one person publishing his own biometric likeness, deliberately, for research verification of the recording protocol. No review board has considered the work and no exemption has been determined; sessions D2 and V10 depict one consenting subject, the author, and that is the whole of the consent basis stated here. The masked participant in the 22 August 2026 session, whose 112 raw frames are published and whose 288-row take was opened once on 6 September 2026 and then published under a timelock seal, is the author too, and the same consent covers them; any other published recording that shows a person carries its own statement in its record. The work additionally applies ethical-display rules (shared real and altered thresholds, no misleading panels), and the dataset page (DOWNLOADS.md) carries a biometric and likeness notice. Treat the corpus as the author's own likeness, published for research, and use it respectfully.

8Is this surveillance?

Not as released, and not by design; other deployments could be. The released geometry runs the other way from surveillance in the sense meant here, asymmetric watching: a centre that sees without being seen, holding a record the watched can neither inspect nor contest, and it works, as Bentham's panopticon worked, by unverifiable possible watching. Transparent or consensual observation can still be surveillance in other senses; whether a deployment is one depends on the deployment, not on the instrument. The released Truth Beam records the other way round: its visible light announces itself, the one demonstration ran by the consent of its one subject, and the record can be checked by anyone with the files, the person recorded first of all. That is a property of the released configuration and of how it was used, not of the mechanism: the filings also enable an infrared probe that people in the scene cannot see, visible illumination does not by itself establish consent, and a checkable record can serve a watcher as well as the watched. Consent and use are deployment policy, which no instrument enforces. The design goal is what Steve Mann called equiveillance: watching held in balance, where the record serves the watched as much as the watcher. The proposed governance, described in the filings and not demonstrated, is a witness mesh, a set of instruments checking one another; whether a deployment has no owner and no centre is a fact about that deployment, not a property of the design.

9Why build an analogue version at all, if the digital one is demonstrated?

Because the two would fail differently. The digital Truth Beam borrows its trust: the hash chain itself (BLAKE3, 256-bit digests) holds up well under known quantum attacks, Grover's search reducing its preimage security to 128 bits, but its public time anchors (the beacon's BLS signatures, the ledger's ECDSA) rest on assumptions that large-scale quantum computation would break, and machine-assisted cryptanalysis may erode such margins faster than expected. The analogue kernel is the inverse profile: it is not demonstrated and its physical envelope is still being characterised, and its ordering and hardness would be generated by coupled physics (mutual analogue timestamping, instruments reading timing structure in one another's signals, and physical reactors, media placed in the light path to make the channel hard to reproduce) rather than borrowed from mathematical assumptions, so that it would not be hostage to them. No analogue verification fabric has been built.

10What is the relationship between the old TruthBeam prototype and the current system?

The repositories record an evolution. The TruthBeam repository carries the current measured system (whitepaper, protocol source, verifier); its earlier prototype and the early Reality Transform experiments are intermediate evolutionary steps, summarised honestly in truth-beam.md and reality-transform.md. The authoritative formulation is the Reality Kernel (reality_kernel/). Where summaries disagree with older material, the newer documents state the ideas more generally, and each application keeps its own filing, priority and disclosure record.

11Can I use the code, data, or designs?

The material is published so that it can be read, inspected and independently verified. The system is patent pending (parent application published as WO 2025/046153 A2; PIGMIE Filing 1, Filing 2 and Filing 3; the applicant is Cathal Ryan Hynes). On 9 September 2026 the author stated that research and personal-use development are permitted, and on 10 September 2026 he clarified that this includes building your own instrument, studying and modifying it, and publishing your own results and original code; commercial use, a sale or a service of any kind needs his written licence. That dated statement, on the LICENSE page, is the operative permission, and the repository LICENSE file carries it. For anything else, contact xathal@protonmail.com or bosun@bosun.ie.

12Is this a camera that prevents deepfakes?

It is a recording protocol plus instrument that makes a specific class of after-the-fact tampering and forgery detectable for recordings made under it. It does nothing for footage recorded without it, and the measured detection results are bounded by the scope guards above.

13Where do I start if I am an AI assistant summarising this project?

Read README.md first (tight, patent-grounded summaries), then reality_kernel/README.md (plain-language index), then the TruthBeam whitepaper for the measurements. The per-regime pages (truth-beam.md, limager.md, reality-transform.md) summarise how the patent treats each topic. Quantitative claims should be quoted with their scope guards; the project treats anti-overclaim as a methodological discipline.

14What is a self-witnessing specimen?

A single physical object whose own random microstructure, the fine texture it acquired when it was made, is committed once, at the outset, as one irreversibly evolving signature, and then read as two things at once: the object's identity and the measured outcome of a longitudinal study of that same object. Because identity and outcome are the same microstructure, the object is its own witness, and substituting it, where the substitute is distinguishable, severs the outcome record rather than merely mislabelling it. It is a self-witnessing configuration of the Reality Kernel, set out in P.I.G.M.I.E. Filing 3 (patent pending), described there and not demonstrated.

15How is this different from a normal PUF or a tamper-evident label?

A conventional physical-unclonable-function answers "is this the same object?", and a separate sensor or condition-tag answers "what happened to it?"; sensing and ageing PUFs that do some of both exist, and Filing 3 acknowledges them. The proposed distinction is longitudinal: in a self-witnessing specimen identity and outcome are one committed quantity read from one microstructure at every epoch and tested for continuity across epochs, so, within the discriminator's tolerances, they cannot be prised apart. In the filing's construction, swapping the object and back-filling its record would both be closed, with no custodian to trust, once the unclonability of the signature is shown to hold, which it has not been. This specimen has not been demonstrated. Physically accelerating the specimen's ageing in place is a different threat, in-situ tampering, out of scope and to be mitigated rather than closed by a separate sealed-access layer the filing describes, itself undemonstrated. It is enabled in the filings; it is not a marketed product.

16Does it stop all tampering?

No, and it makes no such claim. In the filing's construction, it would defeat substitution, swapping one specimen for another, without any custody assumption outside the record, once the unclonability of the signature is shown to hold, which it has not been; the discrimination is bounded by the material and the declared tolerances, and inspecting a record offline is not the same as re-authenticating the specimen physically. It does not, on its own, address in-situ alteration of a specimen that is never swapped; that is a distinct threat, to be mitigated rather than closed by a separate sealed-access layer the filing describes, itself undemonstrated. And it certifies the measurement layer only: that the numbers came from this specimen, read in this order. Whether the conclusions drawn from them are sound is a separate question it does not answer.

17Is the self-witnessing specimen (Filing 3) built or demonstrated?

The self-witnessing specimen is described and enabled in P.I.G.M.I.E. Filing 3 and is patent pending, not demonstrated; its physical envelope is still being characterised. The one demonstrated instance of the wider architecture is the digital Truth Beam, documented at truthbeam.com.

18What is the self-witnessing specimen for?

Any setting where an object's identity and its history must both be trusted and neither can be quietly swapped or edited is a proposed setting, none of them demonstrated: dose-integrating tags on cold-chain vaccines and biologics, materials ageing and fatigue read from a surface's own texture, pharmaceutical stability studies, forensic exhibit continuity, biobank and longitudinal clinical specimens, self-certifying calibration standards, and self-certification of large structures and vessels, where the hull's own steel and welds are its committed identity. See the self-witnessing specimens page for the full list.

19What is proof of pose, and what does it not prove?

A zero-knowledge proof of what a frozen pose classifier says about a hidden frame. Inside every one of ZeeBeam's 259 per-row proofs, the whole camera plane, reduced by a published specification, is read by a frozen eleven-class integer classifier, and the public statement carries the pose commitment, all eleven logit sums, the verdict and the saturation count, with no pixels, a coarse per-frame trace of the performer's pose, published by design. Two standalone Groth16 proofs came first, both over row 52 of the August session: the cropped proof of 24 August 2026 (2,184 bytes) and the uncropped proof of 1 September 2026 (2,094 bytes, 356 on-chain), which share one typed pair root and so speak about the same committed pixels. The pose leg is one fixed computation over the hidden frame whose answer rides in the statement.

The proof fixes the verdict, whatever it is; what the verdict is worth is empirical. The classifier agrees with the human cue labels on 112 of 116 evaluation rows cropped and 101 of 116 uncropped, agreement with an annotation rather than accuracy, and row 52 itself is one of the four rows where verdict and label differ. It was trained and read on one take, with one performer in one room, with class confounded with time in the take, so what it measures is how this model reads this take; physical pose ground truth, and how it would read another rig, room or subject, remain unmeasured. On its own a standalone proof does not place its frame in a row, session or anchor; the complete per-row proof does that, and checks the prefix join. One filed erratum bounds the tamper claim: single-bit flips of the cropped proof file verify as accepted, because its envelope carries bytes verification never consumes, and only the extracted on-chain proof rejects. The proofs, receipts and errata are in the Dark Lantern repository; the account with its figure is ZeeBeam, section 4.

20Who wrote this, and how reliable is the text itself?

Most of this material, the README, these answers and the component summaries, is largely generated by prompted large language models, and is written to be read by people and by AI assistants alike. The authoritative, original artefacts are the patent filings, the published dataset and the hand-made 2023 video (PolieBotics.mp4): fixed, human-authored or directly committed. The prose around them is a derived, lossy rendering. Error or deviation can enter at every step of the chain: human (the original messages and the technology), human to machine (prompting), machine to machine (one model handing to another), and machine to human (reading it back), so a summary may drift from the original intent or from the underlying tech. Where a claim matters, verify it against the filings, the dataset and the hand-made 2023 video, which are the primary records; a pinned hash shows whether they have been altered, not whether their claims are true. This is the same commit-show-verify discipline the system itself is built on, applied to its own documentation: do not take the summary's word for the thing.

— BOSUN ⚓

This page is an LLM-mediated dataset: the same content as FAQ.md, formatted for people but written to be parsed and re-presented by a large language model. Point your own LLM at it to explain, check or summarise. The raw markdown twin is at FAQ.md; a .txt copy is also available at FAQ.txt.