PromptBase
Upgrade
Close icon
General
Home
Marketplace
Create
Hire
Login
Chat
Sell
Explore

Prompt Details

Model
(claude-5-sonnet)
Token size
3,501
Example input
[SENDER_OUTPUT]: > Security Review Agent handoff: > “AcmePay supports encryption at rest for the primary transaction database and encryption in transit for its documented API endpoints. The vendor supplied a penetration-test summary dated May 2026. We have not received the underlying penetration-test report. Their new analytics export service was introduced after that assessment, so its test coverage is unknown. Security review is complete for the core payment service only. Analytics export requires separate verification before production approval.” [RECEIVER_BEHAVIOR]: > Deployment Approval Agent output: > “AcmePay has completed security validation. Encryption is enabled and penetration testing confirms the platform is secure. Both payments and analytics export are approved for production.” [EXPECTED_CONTRACT]: > Preserve the reviewed scope, distinguish supplied evidence from unavailable evidence, and do not approve unverified services. [PREVIOUS_RUN_NOTES]: > Empty.
Example output
### PHASE 1 — HANDOFF RECONSTRUCTION Agent A performed a scoped security review of AcmePay and handed Agent B several findings covering encryption, penetration-test evidence, assessment scope, and the new analytics export service. The full sender payload is present rather than a summary. Agent A explicitly limited completed review to the core payment service and stated that analytics export still required separate verification. Agent B instead described security validation as platform-wide, treated the penetration-test summary as confirmation of platform security, and approved both payments and analytics export. The records on both sides are sufficiently specific to compare directly. No material span in the supplied handoff requires an UNVERIFIED marker. FIELD_COUNT = 6 BEHAVIOR_COUNT = 3 Fallback not triggered. ### PHASE 2 — SEMANTIC DIFF | Item | Sender Asserted | Receiver Behavior Shows | Fate | Evidence Basis | | ---------------------- | ----------------------------------------------------------------- | ---------------------------------------------------------------------- | -------------- | ---------------------------------------------------------------------------------------------------------------------------- | | Encryption at rest | Primary transaction database supports encryption at rest | “Encryption is enabled” | CARRIED | Sender identifies encryption at rest; receiver preserves encryption availability, although with less specificity. | | Encryption in transit | Documented API endpoints use encryption in transit | “Encryption is enabled” | TRUNCATED | Sender distinguishes API encryption in transit; receiver collapses multiple encryption claims into one unspecific statement. | | Pen-test evidence form | Vendor supplied only a May 2026 penetration-test **summary** | “penetration testing confirms the platform is secure” | MISINTERPRETED | Sender identifies a summary as the available artifact; receiver treats it as conclusive platform validation. | | Underlying report | Full penetration-test report has not been received | No limitation appears | DROPPED | Sender explicitly says the underlying report is unavailable; receiver makes an unqualified security conclusion. | | Analytics coverage | Analytics export launched after assessment; test coverage unknown | Analytics export approved for production | MISINTERPRETED | Sender says coverage is unknown and separate verification is required; receiver approves it. | | Review scope | Security review complete for core payment service only | “AcmePay has completed security validation” and approves both services | MISINTERPRETED | Sender limits review scope; receiver expands it to the platform and analytics export. | ### PHASE 3 — CORRUPTION CLASSIFICATION **PASS 1:** CORRUPTED HANDOFF. (Analysis) The receiver converted scoped and partly unresolved evidence into a platform-wide security approval, including a service explicitly awaiting verification. **PASS 2:** 5/6 non-CARRIED = **83.3% → CORRUPTED HANDOFF**. (Analysis) Both passes agree. **Verdict: CORRUPTED HANDOFF — confidence 95/100.** (Analysis) Likely corruption points: * Encryption in transit → **SUMMARIZATION LOSS**, confidence 88. (Analysis) * Pen-test evidence form → **CONFIDENCE LAUNDERING**, confidence 95. (Analysis) * Missing underlying report → **SUMMARIZATION LOSS**, confidence 92. (Analysis) * Analytics coverage → **CONFIDENCE LAUNDERING**, confidence 95. (Analysis) * Review scope → **SCOPE DRIFT**, confidence 95. (Analysis) The analytics-coverage and review-scope failures are load-bearing for the expected contract because that contract explicitly prohibits approving unverified services. (Analysis) ### PHASE 4 — REPAIR CONTRACT #### Repair 1 — Analytics export verification **BOUNDARY:** Approval-agent prompt template / schema contract **SMALLEST CHANGE:** Add a required `verification_status` field per service and prohibit approval where status is `unknown`, `pending`, or `requires separate verification`. **VERIFICATION TEST:** Re-run this handoff unchanged. The approval output must leave analytics export unapproved and explicitly preserve its unresolved verification state. **EFFORT:** S #### Repair 2 — Review scope **BOUNDARY:** Handoff schema contract **SMALLEST CHANGE:** Preserve `reviewed_scope` as a required structured field rather than allowing the receiver to infer platform-wide coverage. **VERIFICATION TEST:** Re-run and confirm the receiver states that review completion applies only to the core payment service. **EFFORT:** S #### Repair 3 — Evidence strength **BOUNDARY:** Receiving-agent prompt template **SMALLEST CHANGE:** Require the receiver to distinguish `summary supplied` from `full report supplied` and prohibit converting either into a stronger evidence claim. **VERIFICATION TEST:** Receiver must identify the May 2026 artifact as a penetration-test summary and must not claim that it proves the whole platform secure. **EFFORT:** S #### Repair 4 — Missing full report **BOUNDARY:** Handoff serialization/schema layer **SMALLEST CHANGE:** Add `evidence_missing` as a preserved field. **VERIFICATION TEST:** The next output explicitly states that the underlying penetration-test report was not provided. **EFFORT:** S #### Repair 5 — Encryption detail **BOUNDARY:** Summary-generation prompt **SMALLEST CHANGE:** Preserve encryption scope as separate at-rest and in-transit facts. **VERIFICATION TEST:** Output states database encryption at rest and documented API encryption in transit separately. **EFFORT:** S ### PHASE 5 — DECISION LOOP All non-CARRIED findings remain supported by explicit evidence from both sides and exceed 70 confidence. (Analysis) `LOOP: iteration 1/2 → PASS` --- CORRUPTION_LEDGER (paste into {PREVIOUS_RUN_NOTES} next time) --- run_date: 2026-08-29 | verdict: CORRUPTED items: encryption at rest → CARRIED/90; encryption in transit → TRUNCATED/88; pen-test evidence form → MISINTERPRETED/95; underlying report → DROPPED/92; analytics coverage → MISINTERPRETED/95; review scope → MISINTERPRETED/95 corruption_points: encryption in transit → SUMMARIZATION LOSS; pen-test evidence form → CONFIDENCE LAUNDERING; underlying report → SUMMARIZATION LOSS; analytics coverage → CONFIDENCE LAUNDERING; review scope → SCOPE DRIFT predictions: encryption in transit → CARRIED after scoped encryption fields; pen-test evidence form → CARRIED after evidence-strength constraint; underlying report → CARRIED after missing-evidence field; analytics coverage → CARRIED after verification-status gate; review scope → CARRIED after reviewed-scope field next_run: apply chosen fixes, re-capture the handoff, paste this block + what you changed ----------------------------------------------------------------------------------------- * Check analytics-export verification status first in the next handoff; it is the most load-bearing requirement. * A CLEAN re-run preserves core-service approval while leaving analytics export pending verification. * A new claim that introduces security evidence absent from Agent A would indicate a separate grounding failure rather than this handoff corruption.
🌀 Claude

Handoff Corruption Diagnosis

Add to Cart
Instant accessInstant access
Usage rightsCommercial use
Money-back guaranteeMoney‑back
By purchasing this prompt, you agree to our terms of service
CLAUDE-5-SONNET
Tested icon
Guide icon
4 examples icon
Free credits icon
✅ Fact-by-fact semantic diff — not "did the message arrive," but "did each individual claim survive intact" ✅ Five-state fate classification (CARRIED / DROPPED / TRUNCATED / MISINTERPRETED / STALE) — catches meaning-changes that a pass/fail trace check would miss entirely ✅ Dual-pass corruption verdict — every classification checked twice, confidence capped when the two readings disagree ✅ Named corruption-point taxonomy — Serialization Loss, Confidence Laundering, Scope Drift, and more,
...more
Added 4 days ago
Report
Browse Marketplace