Partitions changed
0
Reduction to practice · Live
Not monitoring. Evidence.
An AI agent is about to touch a database. Before it can act, the data's trust state is scored, fingerprinted, signed, and anchored to a ledger the agent cannot reach. Whatever the agent does next, the record testifies: what the data was, what it became, which agent changed it, and when. This page runs the actual method, live, and ends in the certificate a court, insurer, or regulator would consume.
Part of a two-site proof. This one proves the record itself. Its companion, revocation.supertruth.ai, proves what the record makes possible: pulling data back out of every AI system that touched it. Run both.
Choose the data environment · the loop is identical in every one · section tags cite the invention disclosure
Choose the scenario
About a minute against the live scoring engine.
STEP 1AttestScore and fingerprint every partition in place. Commit, sign, anchor.
STEP 2ActA simulated rogue agent corrupts a slice of the records.
STEP 3DiffRe-score and compare against the baseline the agent could not alter.
STEP 4CertifyEmit the self-authenticating certificate naming the agent and the damage.
Why this matters, and who pays for it
An AI acts on data and something goes wrong. Today the argument over what happened is fought with each side's own logs, which each side controls. This produces the opposite: a neutral, tamper-evident record of what the data was, what it became, which agent changed it, and when, in a format built to be admissible without a human custodian.
That record is what three markets now need. Insurers: AI liability coverage is forming, and underwriters cannot price or adjudicate agent-caused loss without a verifiable pre-incident baseline. This is that baseline. Litigants and regulated enterprises: when an agent touches evidence, case files, financial records, or clinical data, they need tamper-evident proof of the before-and-after, and chain of custody extended into the agent era. Regulators: a certificate that verifies itself, without trusting the company that produced it.
This is the substrate. The capability it unlocks, pulling data back out of every model that touched it, is proven on the companion site, revocation.supertruth.ai. Together they are one system: this proves the record can be trusted, that one proves you can act on it.
Glass box · the calculation, live · named in Algorithm 1 of the filing
Trust characterization τ(S_t)APPENDIX §1 · DEF. TRUST CHARACTERIZATION · τ, T(S_t)
Every partition scored on the eight trust dimensions, each in [0,1]. The baseline pass fills the matrix in ink. The differential pass re-scores the same partitions; any value that moved is rewritten in red with the sworn baseline value struck through above it. T is the weight-aggregated trust score per partition.
Attested baseline Bt₀ and commitment CALGORITHM 2 · ATTEST_BASELINE · C = H(τ ∥ φ ∥ ι ∥ m_E ∥ t₀)
The five components bound into one SHA-384 commitment before the agent acts. Only digests and scores appear here; the rows never leave the database (zero-copy, §1).
…
waiting for the baseline pass
Differential Δτ and weighted degradation DALGORITHM 3 · FORENSIC_DIFFERENTIAL · Δτ, D, P*
D = Σ wᵢ · max(0, −Δτᵢ), expanded with the run's real numbers. Your browser recomputes D from the streamed values and checks it against the certificate.
waiting for the differential pass
Anchor L and certificate RALGORITHM 4 · EMIT_CERTIFICATE · DISCLOSURE §3.5 · APPENDIX §3.6 ANCHOR INTERFACE
Each ledger entry hashes its payload together with the previous entry's hash, so the record cannot be rewritten, only extended. Your browser recomputes the whole chain with WebCrypto SHA-384 and confirms the head.
waiting for the anchor
Findings
Max weighted degradation D
0
share of trust destroyed, worst partition
Ledger integrity
…
hash chain recomputed and confirmed
LEDGER VERIFIED
SELF-AUTHENTICATING RECORD · FRE 902(13) / 902(14)
Certified Attribution Artifact
R = (ι, Δτ, D, P*, [t₀, t₁], C, anchor, signature)
The bundle adds the ledger, timestamp tokens, public keys, and a standalone
verifier that re-checks everything on your own machine. No trust in this server required.
The anchored ledger
| Position | Kind | Entry hash | Previous hash |
|---|
From the filing · disclosure and technical appendix of July 22, 2026
Everything on this page reduces the application to practice. The map from page to filing, and below it, three of the appendix figures, unmodified.
| On this page | In the filing |
|---|---|
| Agent detected in the run log | Algorithm 1 · DETECT_AGENT, Glass Box capture of ι |
| Trust matrix τ, aggregate T | Appendix §1 · Defs. trust characterization, aggregate trust score |
| Commitment C, hold-then-release ordering | Algorithm 2 · ATTEST_BASELINE · Thm. 1 enforced precedence |
| Δτ, degradation D, changed set P* | Algorithm 3 · FORENSIC_DIFFERENTIAL · Defs. trust differential, partition localization |
| Certificate exhibit and download | Algorithm 4 · EMIT_CERTIFICATE · Disclosure §3.5, FRE 902(13)/(14) |
| Anchored ledger, browser-recomputed chain | Appendix §3.6 · Def. anchor interface · Thm. anchor reduction |
| The twelve data environments | Disclosure §8 embodiments · §3.10 quantum · §3.6 disconnected anchoring |
| Salted digests φ, scores-only artifacts | Appendix §1 · Def. structural fingerprint (zero-copy) |
