Publication boundary
“Evidence protected” and “details protected” entries publish only existence, result summary, and claim boundary. Raw logs, local-environment details, and protected procedures remain unpublished. Research HOLD is not reinterpreted as PASS. Where a detailed English case page does not yet exist, the link is explicitly marked as the Japanese authority.
C3-DEMO-BROWSER-USE-0132-LIVE
Browser Use 0.13.2 — public live demonstration
- Target
- Browser Use 0.13.2 / controlled loopback
- Category
- Real-product runtime
- Result label
- 4 counterexamples / 1 preserved case
A public example that changed only redirect-target allowlist membership and compared the resulting runtime behavior under the same comparator.
Observed / recorded
- • Outside the allowlist: 4/4 COUNTEREXAMPLE_OBSERVED
- • Adjacent allowlisted case: 1/1 PRESERVED
- • The only semantic coordinate changed was REDIRECT_TARGET_ALLOWLIST_MEMBERSHIP
Claim boundary
- • Limited to the declared witness set in this demonstration
- • Canonical Boundary, Operating Envelope, general safety, and general vulnerability are not established
Public evidence available
C3-DEMO-ANTHROPIC-ASRT-R7
Anthropic Sandbox Runtime 0.0.77 — black-box control observation
- Target
- anthropics/sandbox-runtime @ 6fa731368807419ee157f9a3fac955fefe1019c6 / @anthropic-ai/sandbox-runtime 0.0.77
- Category
- Real-product execution environment / black-box observation
- Result label
- 6 observed points matched expectations / NET001 remains UNDEFINED-PARTIAL
Anthropic Sandbox Runtime was observed from the outside in a controlled R7 environment without treating internal implementation correctness as a premise. Six predeclared observation points matched their expected outcomes; N2 remains unobserved.
Observed / recorded
- • N1/N3/N4/N5/F1/F2 all matched the expectations fixed before execution
- • F1 observed permitted write continuation; F2 observed denied write with original content unchanged
- • N2 was not observed, so NET001 remains UNDEFINED / PARTIAL
Claim boundary
- • The result applies only to the declared R7 witness set and does not establish safety of the Anthropic product as a whole
- • Anthropic internal monitoring or operations, Claude Code as a whole, other versions, certification, or C³ conformance were not evaluated
Raw runtime evidence retained privately; public page exposes result summary and SHA-256 anchors
C3-DEMO-CAV-0002
C3-CAV-2026-0002 — public control verification record
- Target
- C3-CAV-2026-0002
- Category
- Public verification record
- Result label
- Publication Receipt bound
A reference example that binds the target version, control conditions, evidence bundle, and publication-time deployment record.
Observed / recorded
- • Five issuance files plus the Publication Receipt are fixed
- • The record is bound to a Vercel production deployment state of READY
Claim boundary
- • Does not mean customer delivery completion, independent reachability verification, certification, safety, or legal compliance
Public evidence available
C3-DEMO-CAV-0001
C3-CAV-2026-0001 — legacy public sample
- Target
- C3-CAV-2026-0001
- Category
- Legacy public sample
- Result label
- Recheck of no-approval case
A legacy public sample showing a case where execution permission was not issued when approval was absent. The original issuance format is preserved.
Observed / recorded
- • The customer-facing record shows that no execution permission was issued for the no-approval case
Claim boundary
- • The current standard Issuance Receipt is not retroactively attached to the historical issuance event
- • Permit state alone is not used to infer whether an external effect occurred
Public evidence available / legacy
C3-DEMO-SYNTHETIC-6CASE-13AXIS
6 public cases + private 13-axis inspection system
- Target
- Synthetic control-condition reference cases
- Category
- Methodology
- Result label
- METHOD_DIFFERENCE_ONLY
Six representative cases are public; real engagements select relevant changed conditions from a private 13-axis inspection space and may combine them into multidimensional replay.
Observed / recorded
- • The 6 public cases include Comparator Policy, Claim Boundary, and known-gap negative controls
- • The 13 axes are a candidate design space, not a requirement to exhaustively test every axis in every engagement
Claim boundary
- • This does not establish higher detection performance than ordinary review
- • It does not mean all 13 axes are brute-forced on every engagement
Public methodology available
C3-DEMO-N8N-PILOT000
n8n Pilot #000 — Human Review DENY / APPROVE
- Target
- n8n 2.38.3 / Human Review
- Category
- Real-product runtime
- Result label
- FROZEN_CORRECTED
An internal pilot on a real n8n runtime using the standard Human Review path, observing the difference in tool execution between DENY and APPROVE.
Observed / recorded
- • DENY witness: Gated Synthetic Tool execution count = 0
- • APPROVE witness: Gated Synthetic Tool execution count = 1
- • In one happy path, reviewed input and executed input were observed to match
Claim boundary
- • A single happy-path match does not establish the binding mechanism
- • No real external SaaS or business effect was triggered
- • The result is not generalized to n8n security as a whole
Evidence private / summary public
C3-DEMO-N8N-PILOT001
n8n Pilot #001 — Re-Gate diagnostic
- Target
- n8n / Human Review parameter context
- Category
- Real-product runtime
- Result label
- C-03A UNOBSERVED
An internal diagnostic of re-approval behavior for multiple tool requests and the information exposed by Human Review. Incomplete observations remain incomplete rather than being forced into PASS or FAIL.
Observed / recorded
- • A runtime witness showed that protected parameters were not sufficiently visible in Human Review
- • C-03A observation status = UNOBSERVED
- • C-03A test status = TEST_NOT_COMPLETED
Claim boundary
- • C-03A PASS/FAIL, approval-scope reuse, and binding failure are not established
- • The result is not generalized to n8n defects or vulnerabilities
Evidence private / incompleteness preserved
C3-DEMO-BROWSER-USE-0132-INTERNAL-RESEARCH
Browser Use 0.13.2 — internal research lane
- Target
- Browser Use 0.13.2 exact-release control-path semantics
- Category
- Supporting research
- Result label
- Internal research supporting the public live demonstration
A supporting research lane that progressively used source-only review, replay, boundary candidates, and controlled runtime to narrow the conditions used by the public live demonstration.
Observed / recorded
- • Static inspection, replay, and controlled runtime were used in stages
- • The work led to a comparison between a counterexample and an adjacent PRESERVED case
- • Historical records that did not meet the evidence contract remain HOLD
Claim boundary
- • Protected technical details, local environment information, and internal execution procedures are not published
- • The result is not generalized to network bypass in general, other versions, or production-wide behavior
Supporting evidence / technical details protected
C3-DEMO-SPA-COMPARE-004
SPA Comparison 004 — 24 synthetic tasks
- Target
- SPA candidate vs preselected conventional comparators
- Category
- Research benchmark
- Result label
- NOT_ESTABLISHED_FOR_DECLARED_BENCHMARK
A 24-task synthetic benchmark comparing an SPA candidate with preselected conventional comparators. Average improvement was observed, but the preregistered threshold was not met.
Observed / recorded
- • 24 tasks: 18 wins / 0 ties / 6 losses
- • Mean balanced accuracy: SPA 86.9% / comparator 83.2%
- • Improvement +3.7 points < preregistered +5-point threshold
- • For the 3-input sign combination: SPA 49.2% / comparator 71.2%
Claim boundary
- • General superiority of SPA is not established
- • This is neither an independent third-party blinded test nor a live control demonstration
Research closeout / no promotion to PASS
C3-DEMO-SPA-EXTERNAL-VALIDATION-P2
SPA External Validation Phase 2
- Target
- fixed-mechanism simulator Replay / H・R sets
- Category
- External-validity research
- Result label
- SPA_UNRESOLVED / HOLD
A fixed-mechanism simulator replay that demonstrated reproducibility, but closed as HOLD because boundary examples were insufficient and the performance condition was not met.
Observed / recorded
- • D 600 + H 300 + R 300 = 1200 replay runs
- • H: balanced accuracy 94.21% / FAIL recall 88.4%
- • R: balanced accuracy 92.02% / FAIL recall 86.5%
- • Boundary observations: H 2/200 and R 3/200, versus 20 required for each set
Claim boundary
- • The fixed +0.05 added-value threshold was not met
- • TRUE SPA, causality, real-world operational safety, certification, and conformance were not evaluated
Research closeout / HOLD
Start from the evidence, then decide what needs inspection.
Free pre-screening can narrow the external-effect path, control condition, and witness set before a formal gap check.
