Assurance LedgerSkimSpeak reference case
DEMONSTRATION ONLY — SELF-ASSESSMENT — this record shows the Assurance Ledger product and is not an issued independent assessment of SkimSpeak.

Private implementation. Publicly checkable claims.

Trust the promise.
Verify the release.

A company grants temporary, read-only access to its private software. An auditor tests each bounded promise, seals the sensitive evidence, and publishes a signed record tied to the exact file customers can download.

Level 1 · Release check

SkimSpeak demonstration

partner-demo-2026-08-22

Assurance incomplete
1 / 6claims evaluated
Ed25519signed record
236served bytes
Expected artifact SHA-25648450e4c4472f8e870e955a544acb281085ca4b12f4c3f5603bb1214303b39ca
Download signed record

Run the public verifier to fetch the exact record bytes, validate signer authority against the verifier-owned trust registry and issuance policy, fetch the signed absolute endpoint, and hash the response in this browser.

The exact verification boundary

It stops at the served file.

The verifier establishes that the response bytes offered by the publisher are the bytes the auditor named. It does not inspect the downloader’s computer or claim what they later install or run.

HTTPS protects transport. A signer key obtained outside the publisher endpoint protects against a substituted record. Multi-vantage monitoring and a transparency log are the production controls against targeted endpoint equivocation. This local demo pins a publisher-controlled key and claims no independence.

The real reference product

SkimSpeak 0.1.11 application release

This text fixture proves the Ledger mechanics. It is not the app binary, and the current private-beta download does not inherit its result.

No issued record

The application stays gray until a current candidate is built with provenance, privately assessed, and bound to the exact bytes offered by its download endpoint.

Audit 00 / 3bounded remediations closed
Audit 13 / 3import controls closed

Application claims issued green: 0 / 6 → 0 / 6. Safer code is real progress; it is not yet release assurance.

  • Source commitment → isolated build provenance
  • Behavioral evidence → scoped claim rows
  • Candidate digest → observable publisher endpoint

Level 2 → Level 3

Promises, then evidence.

Each row is a falsifiable statement. Open it to see the decision rule, evidence grade, assumptions, controls, exclusions, and expiration.

IDENTITY-01The file served now is the file named by this recordartifact identityNot evaluated

The promise

At verification time, the bytes returned by the absolute endpoint named in this publisher-signed demonstration record have the exact declared byte length and SHA-256 digest.

Decision rule

served_bytes.length == 236 AND sha256(served_bytes) == declared_sha256

Recorded result: pass

Evidence grade

Vendor attested

Not evaluated in this demonstration.

Evidence

endpoint-byte-check

The open verifier fetches the endpoint and hashes the response locally.

What this conclusion relies on

REL-27

The verifier received the demonstration signer public key through a trustworthy channel and has not been given a substituted verifier.

Control: Pin the public key in open verifier code and distribute it out of band.
REL-08

The verifier's network view is representative of the artifact offered to intended recipients.

Control: Use HTTPS plus independent monitors from multiple networks to detect targeted equivocation.

Limits

  • No claim is made about installation, execution, or files on a recipient's computer.

Evidence valid through Nov 20, 2026 unless a listed freshness trigger occurs first.

BUILD-01The assessed source was bound to the named artifactprovenanceNot evaluated

The promise

The assessed private source commitment flowed through a declared provenance-producing build to the exact artifact digest in this record.

Decision rule

observed_source_commit == declared_commit AND build_output_sha256 == declared_sha256

Recorded result: not evaluated

Evidence grade

Stated reliance

Not evaluated in this demonstration.

Evidence

missing-build-provenance

No source-to-artifact provenance was issued for this in-memory text demonstration. A real application release must supply it.

What this conclusion relies on

REL-25

The publisher supplied the complete source and build configuration needed for the scoped release.

Control: Record repository, workflow, submodule, dependency-lock, and secret-free environment inventories before testing.
REL-04

The isolated builder executed the committed workflow without an undisclosed substitution.

Control: Pin build inputs and retain signed provenance; use a hardened hosted builder for a production assessment.

Limits

  • This demonstration artifact is not the SkimSpeak application binary.
  • The illustrative record is not an issued third-party build attestation.

Evidence valid through Nov 20, 2026 unless a listed freshness trigger occurs first.

CONSENT-01Summarization requires an explicit user actionuser controlNot evaluated

The promise

For the declared desktop interaction paths, provider-bound summarization begins only after the user invokes the configured shortcut and accepts the applicable provider disclosure.

Decision rule

all observed provider requests have a prior shortcut event and valid disclosure state

Recorded result: not evaluated

Evidence grade

Stated reliance

Not evaluated in this demonstration.

Evidence

consent-state-trace

The production behavioral trace and its commitment have not been issued for this demonstration.

What this conclusion relies on

REL-12

The declared interaction paths cover the release paths represented by this claim.

Control: Publish the path inventory and list excluded extensions, debug modes, and future features.

Limits

  • No universal claim is made about modified builds, developer tooling, or future interaction paths.

Evidence valid through Nov 20, 2026 unless a listed freshness trigger occurs first.

VERBATIM-01Verbatim mode does not call a summarization providerprivacyNot evaluated

The promise

Across a committed verbatim-mode test corpus and declared network boundary, a deterministic egress monitor observes zero summarization-provider requests.

Decision rule

count(provider_requests where mode == verbatim) == 0 for committed cases

Recorded result: not evaluated

Evidence grade

Stated reliance

Not evaluated in this demonstration.

Evidence

verbatim-egress-log

The decision rule is deterministic, but no artifact-bound egress run was issued for this demonstration.

What this conclusion relies on

REL-15

The audit environment observed all relevant outbound network paths.

Control: Deny unlisted egress and record traffic at the sandbox boundary.

Limits

  • Observed absence across a committed corpus is not a universal statement about every possible execution on every machine.

Evidence valid through Nov 20, 2026 unless a listed freshness trigger occurs first.

REDACTION-01Declared sensitive token classes are redacted before provider egressprivacyNot evaluated

The promise

For the public synthetic fixtures and declared token classes, the provider-bound payload contains the configured replacement tokens and no original sensitive values.

Decision rule

for every declared fixture: originals absent AND replacements present in captured provider payload

Recorded result: not evaluated

Evidence grade

Stated reliance

Not evaluated in this demonstration.

Evidence

redaction-fixture-results

The rule is deterministic, but no public vector set and artifact-bound payload trace were issued for this demonstration.

What this conclusion relies on

REL-19

The public token-class list accurately defines what this claim covers.

Control: Show covered and excluded classes; treat newly supported classes as a freshness trigger.

Limits

  • No claim is made that every possible identifier or contextual inference is removed.

Evidence valid through Nov 20, 2026 unless a listed freshness trigger occurs first.

SEMANTIC-01Summaries preserve declared facts on the committed evaluation setqualityNot evaluated

The promise

On a precommitted fictional evaluation set, a frozen evaluator profile scores summaries against a published preservation and prohibited-invention rubric.

Decision rule

fact_preservation_rate >= 0.90 AND prohibited_invention_count == 0 on the committed set

Recorded result: not evaluated

Evidence grade

Stated reliance

Not evaluated in this demonstration.

Evidence

frozen-evaluator-run

No evaluator run or numeric result is represented by this demonstration record. A production run would treat the model as a measuring instrument, not an oracle.

What this conclusion relies on

REL-21

The frozen evaluator and rubric are sufficiently calibrated for this bounded quality measurement.

Control: Publish model, weights, tokenizer, harness, prompt, scorer, runtime, disagreement rate, and human calibration commitments.
REL-19

The committed fictional set is representative of the usage described by the claim.

Control: Publish the sampling method, uncertainty, known failures, and scope limits.

Limits

  • The result is not a universal guarantee that every future summary is correct.
  • A frozen open model makes the measurement reproducible within its execution profile; it does not make the judgment logically true.

Evidence valid through Sep 21, 2026 unless a listed freshness trigger occurs first.

The commercial workflow

Open the audit room. Keep the IP inside.

SkimSpeak begins with publisher-authorized superuser access. The same verifier accepts narrower access adapters as the service matures.

  1. 01Authorize

    Time-boxed, least-privilege access with a signed scope and access receipt.

  2. 02Inspect

    Map claims to source, build, tests, runtime observations, and explicit counterexamples.

  3. 03Seal

    Keep source and sensitive traces private; export commitments, outcomes, limits, and reliance statements.

  4. 04Sign

    Bind the claim graph to the assessed source, build provenance, and exact served artifact digest.

  5. 05Monitor

    Date-based expiry is implemented. Automated dependency, workflow, key-revocation, and endpoint-drift monitoring is a production-stage control.

Reference adapter

Publisher superuser

Fastest way to prove the complete system with a cooperative reference customer.

Customer-controlled

Audit skill in CI

The customer runs an open, signed skill inside its boundary and exports only the minimized evidence package.

Managed service

Temporary GitHub App

Read-only Contents, Actions, Checks, Attestations, and metadata access; one-hour tokens are revoked after the audit and no production write access is required.

What the check means

Confidence without pretending certainty.

Current

All predeclared in-scope claims are current

The signed record and exact endpoint bytes verify, and no in-scope evidence row has expired.

Stale

It passed, but needs renewal

Date-based expiry is computed by this demo. Other signed freshness triggers are disclosed but not automatically monitored yet.

Failed

A required check failed

The signature, artifact identity, or a required claim did not match the signed statement.

This demonstration does not claim

  • No inspection of a downloader's machine, installed files, or running process.
  • No assurance for future releases or bytes not matching the declared digest.
  • No universal correctness, security, privacy, or availability guarantee.
  • No cryptographic verification of a third-party provider's retention or deletion behavior.
  • No independent assessment is represented by this demonstration record.