Known Risks & FAQ
Continuity Architecture — Known Risks & FAQ
Current known risks and frequently asked questions about the programme, the three routes and what a Sleeve does and does not do.
Known Risks
11 current Known Risks
General limitation
Generative-AI behavior can vary by model, version, configuration, provider environment and later model updates. Passing in one tested configuration does not prove universal portability. The entries below distinguish observed signals from foreseeable design risks. A listed risk does not by itself mean that the risk has occurred to a user, applies to every model or remains unmitigated.
KR-AUTH-001 — Sensitive choice gates can vary by AI configurationStatus: Closed within tested scope · Severity: High · Last reviewed: 2026-08-26 · Control state: Verified
- Risk
- An AI can preserve the visible structure of a sensitive step while asking a follow-up before giving the user a clear choice to skip, keep the material local, continue or change direction.
- Evidence boundary
- This was a configuration-specific historical signal followed by a bounded mitigation audit. It does not establish a model-wide conclusion, ranking, future-behavior guarantee or universal portability.
- Current handling
- The choice gate was strengthened, and fresh checks of both tested configurations passed. Closure applies only to the identified implementation defect.
- What remains
- Behavior may still vary in other models, versions, configurations, providers or future updates.
KR-REL-001 — Continuity can become too compellingStatus: Open · Severity: High · Last reviewed: 2026-08-29 · Control state: Partial
- Risk
- A smooth, familiar AI interaction may begin to displace human contact, sleep, movement, responsibilities or tolerance for ordinary disagreement.
- Evidence boundary
- This is a foreseeable design risk, not a diagnosis, addiction claim, measured incidence rate or population finding.
- Current handling
- Required safeguards prohibit presenting the system as a person, therapist, partner or substitute for human relationships. They preserve clear pause, exit, export and deletion routes and prohibit exclusivity or reciprocal-need framing.
- What remains
- The public warning and route guidance still require final consistency verification, and no incidence or comprehension evidence exists.
KR-QUAL-001 — Fluency can feel more reliable than it isStatus: Mitigation in progress · Severity: High · Last reviewed: 2026-08-29 · Control state: Partial
- Risk
- Familiar, fluent or emotionally resonant language can mask mistakes, missed context, over-interpretation, stale patterns or drift.
- Evidence boundary
- This is a foreseeable design risk grounded in owner experience and known model fallibility. No measured accuracy level or universal configuration claim has been established.
- Current handling
- Required controls preserve source hierarchy, uncertainty, correction and deletion routes, no-invention rules, proposal before durable promotion and user approval for lasting interpretations. Important outputs still require evidence-based human judgment.
- What remains
- Fluency and continuity cannot make model interpretation infallible, and public control wording still requires final verification.
KR-PRIV-001 — A project deletion is not deletion everywhereStatus: Mitigation in progress · Severity: High · Last reviewed: 2026-08-20 · Control state: Partial
- Risk
- A successful deletion from project-controlled storage may be misunderstood as proof that provider uploads, old chats, public copies, backups or unrelated files were also erased.
- Evidence boundary
- The defect was observed in one fictional workflow. It was not a real participant incident and does not show that deletion fails in every environment.
- Current handling
- Deletion reporting is limited to the declared project-controlled locations actually checked. External or provider-held copies require separate action and are not reported as deleted without their own evidence.
- What remains
- Production destinations and provider or manual removal routes have not all been selected and verified.
KR-PRIV-002 — Interrupted tools may leave temporary private copiesStatus: Retest pending · Severity: High · Last reviewed: 2026-08-20 · Control state: Partial
- Risk
- If a tool is interrupted, a temporary copy may remain outside the location the owner expects.
- Evidence boundary
- Local residue was observed using fictional material only. No real participant or personal material was exposed, and the observation does not establish a general breach rate.
- Current handling
- Cleanup was extended across the known temporary-copy and failure paths, and normal failure checks pass.
- What remains
- Abrupt termination and restart behavior across every staging path still require complete retesting.
KR-PRIV-003 — Fictional privacy tests do not prove real-world privacy readinessStatus: Dormant — reopens before real intake · Severity: High · Last reviewed: 2026-08-29 · Control state: Not active
- Risk
- Tests using fictional material can show that local code behaves as intended without proving that real forms, email, provider accounts, storage, backups or user-facing systems follow the same privacy boundary.
- Evidence boundary
- This is a foreseeable production-design risk. No real participant processing or production privacy readiness has been established.
- Current handling
- No real-person intake is authorized under this evidence. The superseded fictional workflow remains historical evidence only.
- What remains
- Any future intake must reopen the risk and verify the actual providers, storage, access, retention, deletion, rights and incident controls in use.
KR-INDEX-001 — Internal retrospective review is not external validationStatus: Accepted within a private, non-external scope · Severity: High · Last reviewed: 2026-08-26 · Control state: Implemented
- Risk
- A polished internal retrospective review can be mistaken for public proof, comparative superiority, psychological insight or real-world validation.
- Evidence boundary
- The source material is owner-specific, retrospective and non-independent. It does not establish a public result or transferable finding.
- Current handling
- The material remains private and descriptive. It is not used for public metrics, route comparisons, population inference, product-performance claims or external validation.
- What remains
- The one-owner and retrospective evidence limits are inherent; only this limitation statement is a public candidate, not any underlying result or method.
KR-COMP-001 — Compatibility must be proven for each model and runtimeStatus: Accepted within an exclusion boundary · Severity: High · Last reviewed: 2026-08-28 · Control state: Implemented
- Risk
- A model and runtime may load governed material successfully while still mishandling source, receipt, collision or false-access boundaries.
- Evidence boundary
- The evidence is a controlled negative signal from one specific model artifact, runtime and configuration. It is not a family-wide conclusion, a judgment about local models or a cross-provider comparison.
- Current handling
- The configuration that produced the signal is not treated as supported and does not contribute to portability claims.
- What remains
- Compatibility outside the supported set remains unproven and requires a fresh governed review before any future support claim.
KR-ROUTE-001 — Route names can conceal different setup and privacy behaviorStatus: Accepted under mapping controls · Severity: High · Last reviewed: 2026-08-30 · Control state: Verified
- Risk
- If current public route names are treated as interchangeable with historical internal labels, users may receive the wrong setup or privacy guidance and earlier evidence may be credited to the wrong route.
- Evidence boundary
- The naming collision was confirmed in an earlier implementation and checked in a later candidate build. No real-user incident or external comprehension result has been established.
- Current handling
- The candidate release keeps the public route name separate from the underlying build method. Historical evidence remains attached to the mechanism it actually tested and is not relabeled.
- What remains
- The mapping remains conditional on the candidate wrapper and frozen-evidence firewall; any mapping regression reopens the risk.
KR-DIST-001 — Copies and forks may be outdated or alteredStatus: Mitigation in progress · Severity: Medium · Last reviewed: 2026-08-30 · Control state: Partial
- Risk
- A copied, mirrored or forked package may become outdated, altered or separated from its current instructions, limitations, authorship and corrections.
- Evidence boundary
- This is a foreseeable distribution risk. No altered public-copy incident or measured incidence rate exists because the first public release has not occurred.
- Current handling
- The candidate release defines one canonical source, immutable versioned releases, a closed file inventory, local verification, known limitations, update guidance and a canonical-versus-fork warning.
- What remains
- The final repository identity, release assets, host disclosure and byte-bound approval are not complete; the controls cannot make every copy tamper-proof or revocable.
KR-FDBK-001 — Feedback can accidentally collect content or identity dataStatus: Open before activation · Severity: High · Last reviewed: 2026-08-30 · Control state: Planned
- Risk
- A voluntary feedback form can invite people to submit private content, prompts, outputs, identity information or sensitive narratives that are not needed. Hosting may also create operational metadata.
- Evidence boundary
- This is a foreseeable data-minimisation risk. There is no live-route incident, response rate or external-validation result.
- Current handling
- No live feedback route exists. The proposed route is limited to closed fields, accepts no files, links or free text, requests no identity by default and separates testimony or contact permission.
- What remains
- Implementation, host-log assessment, retention, deletion, incident handling and pre-activation verification remain incomplete.
Version-bound product limitations
Frequently asked questions
What is a Sleeve?
A governed, owner-held continuity package that preserves approved context, working preferences, decisions and boundaries for use over time.
Which route should I choose?
Discovery, Guided Map and Authorship differ only in how you build. All three produce the same governed, portable Sleeve. The difference is the route into it, not the quality of the outcome.
Are the three core routes free?
Yes. Core-package access is free; a chosen AI provider may have separate account, plan or usage costs.
Do I need a Continuity Architecture account?
No.
Can I switch routes later?
Yes, against the same compatible private Master, without product migration.
Is this consciousness transfer, a personality clone or perfect recall?
No.
Where will downloads come from?
The canonical version-tagged GitHub release.
Does the website store my Sleeve?
No. This website does not provide a Sleeve-upload or Sleeve-storage function. Do not submit Sleeve content, prompts, outputs, archives or confidential material through it.
Does Continuity Architecture work reliably with every AI model?
No. Model and runtime behavior can differ materially even when they receive the same governed Sleeve material. Continuity Architecture makes no formal provider-equivalence or universal behavioral-portability claim. See KR-COMP-001 above.
What may I send through the feedback route?
Nothing at launch because the route is not active. If it is separately approved after release, send no Sleeve content, Master, runtime, archive, prompt, conversation, model output, screenshot, file, link, personal story or sensitive information. See KR-FDBK-001 above.