Proof at the Point of Creation

Executive Summary

As AI-generated content becomes indistinguishable from human work, the question of authorship can no longer be answered by looking at outputs alone. More and more, the cost of that uncertainty reaches well beyond the classroom and into our everyday lives.

Workings is a local-first provenance and proof system that cryptographically verifies the creation process of digital work without exporting the underlying content. It captures rich creation-time signals on the user's device, converts them into irreversible cryptographic commitments, anchored in time and independently verifiable by third parties.

Unlike AI Detectors, Workings does not rely on post-hoc analysis of the final artefact to infer authorship. Instead, it proves creation conditions at the process layer, and locks those commitments to an immutable timestamp.

A planned zero-knowledge proof layer will extend this further, verifying creation conditions without disclosing any underlying evidence at all.

Core properties

• –On-device capture and processing only
• –Verifiable AI-assistance disclosure - declared AI use is recorded and attested, not hidden
• –No raw content or artefacts leave the device unless the user proactively shares
• –Independently verifiable commitments over time-ordered evidence
• –Immutable anchoring for future-proof timestamps
• –Modular human-signal architecture

Workings functions as an origin layer for digital work - aiming to do for authorship what SSL did for web traffic, but for provenance.

Key elements of the Workings system are the subject of pending patent applications, covering aspects of on-device capture, cryptographic commitment structures, and zero-knowledge proof-based verification of the creation process.

1. Problem Statement: Output-Level Trust Is Broken

Generative AI systems now produce text, images, code, and designs that are statistically indistinguishable from human outputs. This collapses the reliability of any approach that inspects only the final artefact.

The limits of AI detection

• –Adversarial drift: Detectors lag model releases by weeks or months.
• –False positives: Genuine human work is routinely misclassified.
• –Non-falsifiability: There is no durable audit trail, only probabilistic scores.

Detection is a guess. The reliable path is to establish provenance during creation.

However, the necessary process-level evidence creates a second problem: the obvious way to capture it is surveillance. Cloud monitoring and proctoring tools demand that creators export their keystrokes, screens, and drafts to a third party for analysis. This is a privacy cost most individuals and institutions rightly refuse.

Trustworthy provenance therefore requires both: high-fidelity evidence of the creation process, and a guarantee that the evidence remains under the creator's control.

Workings resolves this by capturing rich process evidence entirely on-device and sealing it cryptographically in time. The summary or proof can leave while the data stays private.

2. Technical Design Goals

Workings is designed to satisfy five hard constraints simultaneously:

• –Strong user privacy - zero exfiltration of sensitive content
• –Cryptographic verifiability - mathematical proofs, not heuristics
• –Future resistance - validity does not decay as AI improves
• –Low operational friction - lightweight, offline tolerant
• –Regulatory safety - minimal data surface area

3. System Architecture (High Level)

Workings consists of five primary layers:

1. 1.Secure Local Capture
2. 2.Commitment & Hashing
3. 3.Signal Analysis
4. 4.Zero-Knowledge Proof
5. 5.Immutable Anchoring

Each layer is independently evolvable and communicates via narrow, cryptographically defined interfaces.

4. Secure Local Capture Layer

Purpose

Capture and retain rich local evidence of human-led creation at the time it occurs, while ensuring that no sensitive content is disclosed beyond the user's device.

Local Evidence Capture

Workings captures a user-configurable set of creation-time interaction evidence (including screenshots and input events) sufficient to support verification of continuity and authorship. All evidence capture occurs strictly on-device. Workings supports both explicit, user-initiated sessions and longer-form, document-based capture, where an uploaded file may be used locally to scope the relevant evidence window.

Captured evidence is never transmitted off-device by default. Instead, its integrity is fixed in time through cryptographic commitments, enabling tamper-resistant verification without requiring disclosure of the underlying content. Evidence held on the device is encrypted at rest and protected behind the user's credentials, with automatic locking when the device is idle.

AI-Assistance Attestation

Workings treats AI use as a spectrum. What counts as legitimate is rapidly changing over time and can be judged differently based on the setting or audience. Where AI tools are part of the workflow, Workings recognises when a known AI tool is used during a session and records that interaction - when, which tool, for how long - as part of the same on-device evidence stream as everything else.

At proof time, a signed attestation of that AI assistance is included in the report, cryptographically bound to the rest of the evidence. The result is that a proof doesn't just assert “a human made this” - it asserts “this was made under these declared conditions, including this AI assistance.” In this way users can be credited for legitimate AI use.

Privacy & Lossiness

The specific evidence types, sampling strategies, and sources of randomness are intentionally non-public and evolve over time. While rich evidence may be retained locally under user control, all artefacts that leave the device (hashes, commitments, proofs) are deliberately lossy and irreversible. They contain sufficient entropy to support cryptographic proof, but not enough information to reconstruct the underlying work or user behaviour. Even the link between evidence and the public anchor can be permanently destroyed, by the user, at any time.

5. Commitment & Hashing Layer

Each captured element is immediately transformed into a unique cryptographic commitment. Commitments are assembled into a tamper-evident, time-anchored structure that enforces chronological integrity across the session.

Security properties

• –Tamper evidence: Any modification of specific files invalidates the affected commitments and verification path (without invalidating unrelated evidence)
• –Fixed order: Re-ordering is detectable
• –One-way: Original data cannot be reconstructed

This produces a compact, verifiable commitment log for each session.

6. Human-Signal Analysis Layer

Because the record is genuine and time-anchored, the process it captures is legible. In many cases the human nature of the work is self-evident from the evidence itself, read by the person reviewing the report or timelapse.

A signal-analysis layer in the Workings architecture will turn that reading into an automatic, private assertion: evaluating the committed evidence stream against a set of human-plausibility constraints to test whether observed interaction patterns are consistent with recognised classes of human-led creation. The specific signals, thresholds, and weighting are deliberately non-public.

By design:

• –Raw signals are not exported.
• –Derived metrics are not exposed.
• –Only proof-relevant statements are asserted.

7. Zero-Knowledge Proof Layer

The Workings architecture includes a zero-knowledge proof layer. Once complete, it will attest to the following without revealing inputs:

• –The existence of a valid, untampered commitment set
• –That the commitments span a specific time window
• –That required human-plausibility constraints are satisfied

Proof construction details, zk circuit composition, and constraint definitions are deliberately abstracted and subject to change to preserve system integrity.

Why zero-knowledge

• –Verifiers learn only that conditions were met
• –No behavioural data, content, or identity is disclosed
• –Proof size remains compact and portable

Verification is deterministic given the commitments, enabling reproducible independent checks.

8. Immutable Anchoring Layer

To lock commitments in time, Workings anchors mathematical summaries of evidence commitments to an immutable public ledger. Proofs derive authority by referencing these anchors rather than being anchored independently.

Once anchored, a proof only needs to outrun today's AI, not tomorrow's.

A worked example

Seal your evidence or proof with Workings today, and its anchor is written to the public ledger today. Years from now, an AGI may fabricate a flawless record of the same work being made, but it cannot anchor that record in the past. Your work carries today's timestamp; any forgery, however perfect, can only ever carry a later one.

9. Verification Flow

A verifying audience:

1. 1.Receives a proof receipt (URL or file)
2. 2.Recomputes verification checks locally
3. 3.Confirms anchor existence and timestamp

Verification of evidence or reports never requires trusting Workings. Every inclusion path resolves against the public anchor, and producing a false path that resolves to it is computationally infeasible, for Workings as much as for anyone else.

Initially, retrieving that verification path relies on Workings' hosted records. As proof receipts become self-contained, carrying the inclusion path themselves, verification can be performed against the public ledger alone. Workings touches only the capture and sealing of evidence, not its later verification.

What a verifier reads is not a score but an account of the work taking shape. Depending on what the creator discloses, that may be the process evidence itself, drafts forming and edits accumulating, or an attested summary derived from it without exposing the underlying frames. Either way, because the record is tamper-proof and time-anchored, what they read is grounded in a genuine process, and human involvement is, in many cases, self-evident from it. Workings proves the evidence is real; the creator chooses how much of it to show.

A worked example

A designer produces a campaign concept over an afternoon, with Workings running. It captures the process on their device and seals it in time.

When the client asks how the work was made, they load the concept into Workings locally, generating a report: the timeline, the tools they used, and any AI they had switched on monitoring for. Anything that arrived without the work behind it would show as a step change. It's redactable, with no screenshots unless they include them. They send the report as a PDF to the client.

The client opens it in a browser, with no account and no software, and verifies it is a real report against the public anchor. They learn how the work was created, and never see anything the designer didn't want to share.

10. Security & Threat Model

Addressed threats

• –After-the-fact content fabrication
• –Replay or session splicing
• –Cloud leakage of sensitive data
• –Detector gaming via prompt engineering

Out of scope

• –Identity attribution: Workings proves process, not which specific human.
• –Device compromise: no system fully defends a device already controlled by an attacker.

Workings is designed to raise the cost of forgery beyond practical levels without introducing surveillance risk.

11. Regulatory & Compliance Posture

Because Workings:

• –Processes data locally
• –Encrypts and access-protects all evidence held on the device
• –Stores only pseudonymised commitments
• –Avoids automated decision-making
• –Lets users permanently sever the link between their evidence and any public anchor

…it typically falls below high-risk regulatory thresholds (EU AI Act, GDPR, AU Privacy Act, US sectoral regimes).

12. Positioning Summary

Workings is not an AI detector, watermark, or monitoring tool.

AI detectors guess about outputs. Watermarks travel with files and can be stripped. Monitoring tools surveil people and export their data. Workings does none of these.

It is a cryptographic origin protocol - aiming to do for authorship what SSL did for web traffic: make verifiable trust a default, invisible layer.

It answers one question reliably:

How was this work created?

Not a guess about the output. Instead, a cryptographic proof of the process: verifiable creation conditions, an anchored time window, and declared AI assistance - all without exporting the underlying data.