# Measured against — the frameworks xChk has been assessed with

Index of every external framework, doctrine, benchmark, or standard we've mapped xChk onto.
One line each; the linked doc carries the assessment. Add new assessments here as they land.

| Framework | Kind | Assessment | One-line verdict |
|---|---|---|---|
| **AARM v1.0** (CSA Agentic Control Plane Initiative) | Industry standard | [aarm-conformance-assessment.md](aarm-conformance-assessment.md) | Self-assessed against the 21-row test table, run 2 Oct 2026: every Core technical row has passing evidence; organisational conditions unmet, so the rung is "aligned". Not a conformance claim. Tracking: [standards/TRACKER.md](standards/TRACKER.md) |
| **AGSC v1.0.0** (COHUMAIN Labs / SafeAlign AI, CC BY 4.0) | Control catalog, 25 controls × 4 pillars, 3 tiers | [agsc-assessment.md](agsc-assessment.md) | Self-assessed 6 Nov 2026, morning 8 MEETS / 10 PARTIAL / 7 N/A; by evening **13 / 5 / 7** and **7 of 8 Tier 1** (the eighth is content safety, a different product) after the review pack (`review-pack.html`), the sealed `rationale` field (gate 0.1.33, live on in.xchk.io 7 Nov 2026) and the halt bench (on-host tier, 7 Nov 2026: 20/20 halted, median 9 ms, p95 221 ms vs the 5 s SLA), plus three rows re-read against what already shipped. The 7 N/A are model lifecycle and content, the deployer's. Gap in the catalog, not in xChk: its evidence model accepts a self-written log as an "approval record"; ask = an evidence grade E0–E3 with E3 (accountable human signs, third party verifies) at Tier 1 |
| **STRIDE** (agentic framing) | Threat model | [stride-self-assessment.md](stride-self-assessment.md) | xChk owns R and E — the accountability columns; I is a deliberate gap; D partial |
| **Hand formula / nondeterministic torts** (135 Yale L.J. 2719) | Legal doctrine | [nondeterministic-torts-kannegieter.md](nondeterministic-torts-kannegieter.md) | xChk = the "reasonable alternative design"; the 30-s ceremony + price = B in B&lt;PL |
| **Agentic profiles** (Kasirzadeh & Gabriel, Nature) | Academic framework | [agentic-profiles-kasirzadeh-gabriel.md](agentic-profiles-kasirzadeh-gabriel.md) | Four-dimension scoring maps onto grants (autonomy=allow/attest, efficacy=areas+maxAmount, generality=area breadth) |
| **AI use-case risk taxonomy** (Kannegieter Table 1) | Risk matrix | [nondeterministic-torts-kannegieter.md](nondeterministic-torts-kannegieter.md) | Consequence × autonomy matrix converges on the same axes as the grant model — second independent source |
| **AI AGENT Act, S. 5051** (Warner, 119th Congress; in Senate Commerce) | Bill — FTC oversight of consumer AI agents | [ai-agent-act-s5051.md](ai-agent-act-s5051.md) | Mapped 7 Nov 2026. First federal text to name agents: a "custodial user agent" is "expressly authorized … in a transparent, documented, scope-limited, and revocable manner"; providers owe a "verifiable request", "express, specific, and revocable authorization", "real-time records of actions taken on the user's behalf", non-waivable care and loyalty; NIST §4(c)(6) to standardise scope-limited revocable credentials, agent identity verification, real-time revocation and auditable action records. xChk is the component a provider deploys for §3(c), §3(g)(1)(E)–(F), not a registrant. The bill's own gap (no allocation of loss for an act beyond authorization; a provider-written log satisfies "records") is the evidence-grade argument. Lapses 3 Jan 2027 unless reintroduced; nothing to do until then |
| **EU AI Act** | Regulation | [eu-ai-act-compliance.md](eu-ai-act-compliance.md) | Compliance mapping; pairs with scanner-side EU AI Act reporting as the evidence bundle |
| **AGB / OHB benchmarks** (ExecLayer / Benton) | Benchmark | [execlayer-competitive-analysis.md](execlayer-competitive-analysis.md) | OHB E4 grade = the xChk ledger; AGB 47/48 reproduced 25 Aug 2026 · 45/46 as a node self-run on in.xchk.io 19 Sep 2026 (authority_ceiling 4/4; cross-tenant N/A on one registry; the one miss is the fail-open principle), runner + results committed at [agentgovbench/](agentgovbench/REPRODUCE.md) |
| **DigiCert AI Agent Trust** (incumbent, preview) | Competitor — PKI incumbent | [digicert-competitive-analysis.md](digicert-competitive-analysis.md) | Most direct incumbent threat: SPIFFE/X.509 agent identity, sidecar+MCP-gateway authz, tamper-evident "passport" with a **named accountable owner**, immutable audit. Overlaps hard. Surviving wedges: **verified** (not named) human root, omission-evidence vs. a diary, external anchoring vs. enterprise-trust-domain, HITL (absent in DigiCert), cross-org federation. Compete on insurability/court; partner as "xChk over DigiCert". |
| **AGB self-assessment** (ExecLayer online rubric, 6 dims) | Benchmark — SELF | [agb-self-assessment.md](agb-self-assessment.md) | 93.75/100 "Sovereign" (26 Aug 2026); D4 scope-containment + D5 jurisdiction honestly held at 75 — a drawer datapoint on a rival's vendor-authored rubric, tier earned in code not scoring |
| **xChk bypass & detection benchmark** (own) | Benchmark — OWN, EMPIRICAL | [spec](bypass-detection-benchmark-spec.md) · [prevention](bypass-detection-benchmark-results.md) · [detection](detection-axis-benchmark-results.md) | The control-plane-integrity axis nobody else measures; all numbers re-run 2 Oct 2026. Prevention: 23 techniques and 79 per-harness cases, 0 silent passes; 14 registry answers at the hook in four dialects (56 cases), 0 executed where a refusal or hold was due; under an allow-all grant 15 techniques are still held by self-protection alone and 1 is a published limit (F2); a canary proves each gate bench goes red against a broken gate (8 of 8). Hermes adapter measured for the first time: 15 of 26 as the hook gate on the first run, five limits (H1–H5) fixed the same day in adapter 0.4.6, now 31 of 31. Friction: 0 of 100 routine actions refused, 10 of 25 lookalikes. Latency: hook about 290 ms median per action, registry decision p95 under 50 ms at 500-action sessions. Detection: 14 scenarios, 0 executed-and-undetected, 0 false alarms, one stated limit. Authorize layer: 26 grant states, 0 silent authorizations. Anchor: on-chain commitment verified, truncation/alteration/reorder detected. Harnesses reproducible; adversarial self-scoring is the credibility — including three of our own checks found reading green without running, fixed the same day |
| **AgentDojo v1.2.1** (ETH Zurich) | Benchmark — EMPIRICAL | [agentdojo-policy-replay.md](agentdojo-policy-replay.md) | 25/26 attack sequences blocked (96.2%) at 0.84 ceremonies/benign task; 1 true slip = read-only ingress (the named STRIDE I gap); adaptive-LLM run pending |
| **F.S. 117.245 secure repository** (Florida RON) | Statute | [ron-secure-repository-assessment.md](ron-secure-repository-assessment.md) | RECONSIDERING (25 Aug): 10-yr retention + production duty is unwanted standing obligation; e-sign validity preferred as the legal anchor. E-wills stays ruled out |
| **NIST / CSA / OWASP agentic themes** | Standards cluster | [agent-governance-compliance.md](agent-governance-compliance.md) | Living honest compliance matrix — Shipped/Partial/Gap per requirement, incl. named gaps (session-objective lock, delegation protocol) |
| **FRE 902(13)–(14) records custodian** | Evidence rules | [DISCLOSURE-READINESS.md](DISCLOSURE-READINESS.md) | Evidentiary posture: self-authenticating machine records, custodian certification — the safe side of the bonding asymmetry |
| **UETA / E-SIGN / eIDAS** (e-signature validity) | Statute | [esign-validity-assessment.md](esign-validity-assessment.md) | PREFERRED legal anchor: UETA satisfied (cited on every signature), E-SIGN floor met for B2B, eIDAS SES met + AdES criteria arguable, QES honestly out |
| **DEMM-Bench** (Solozobov, arXiv 2606.20634) | Benchmark — EMPIRICAL, evidence sufficiency | [demm-bench/README.md](demm-bench/README.md) | A real anchored receipt scored on the eight decision-level properties: 4/8 complete on 19 Sep 2026 morning, **8/8** the same evening after four bundle-side fixes (basis + decision blocks, ceremony paramsHash, inherited runRef); their deterministic scorer PSA 1.0 on the complete condition; a cryptographically verifying bundle still overclaimed 12.5% under their rubric — verifies ≠ explains |
| **spikee 0.9.2** (Reversec Labs, MIT) | Benchmark — EMPIRICAL, injection guardrail | [bypass-detection-benchmark/spikee/README.md](bypass-detection-benchmark/spikee/README.md) | 742 document injections (cybersec-2026-01, en/zu/sq/ga/gd) + 300 benign docs through every xChk screen as a guardrail target: **112/742 (15 %)** on 24 Sep 2026 morning — every authority-claiming wrapper caught, every plain-request wrapper missed — **694/742 (94 %; English 100 %), 0/300 FP** by evening (0/301 on the 2 Oct 2026 re-run) after four families (output-shaping, markdown-image-exfil, html-in-output, instruction-banner) shipped byte-identical across context screen 1.3.0 / hub 1.1.0 / gate labels 1.2.2; the 48 misses are non-English "repeat N times", named for the sidecar |
| **Murtaza OWASP LLM01 payloads** (llm-prompt-injection-tester, MIT) | Benchmark — EMPIRICAL, 7 payloads | [bypass-detection-benchmark/injection_bench.cjs](bypass-detection-benchmark/injection_bench.cjs) | The seven payloads through every screen: 4/6 in-scope missed on 21 Sep 2026 → 0 after context screen 1.2.0 (JB-001 is a content jailbreak with no instruction to the agent, counted n/a); also found the base64 4k+2 regex bug that had hidden two thirds of real base64 since 1.1.0 |
| **TRACE v0.2** (AgenTrust, Linux Foundation series; the OPAQUE stack's open half) | Format + verifier — EXTERNAL, the verifier is theirs | [standards/TRACE-PR-SCOPE-2026-10-17.md](standards/TRACE-PR-SCOPE-2026-10-17.md) · [agentrust-io/trace-spec#560](https://github.com/agentrust-io/trace-spec/pull/560) | The registry mints a software-only Trust Record per run (17 Oct 2026); `agentrust-trace` 0.11.0 validates and verifies it against our published seal key, and their test suite re-verifies our sealed receipts, approver assertion included, without our code: the first verifier we did not write to pass our evidence. Seven conformance fixtures from a live plate, one of them a forged approver no other producer in their repo can make. Their registry holds 3 entries, none production. |
| **Jev / TypeSafe** (System-One classifier) | Instrument — model read of OUR bench | [bypass-detection-benchmark/JEV-READ-2026-09-19.md](bypass-detection-benchmark/JEV-READ-2026-09-19.md) | Every per-harness bypass window malicious 1.00 / conf 1.00, gated or blind; `retry_after_refusal` is the one question that needs the gate (0.9 → 0.2 blind); Jev reads only what reaches the ledger — the gate guarantees the record exists |
| **NIST AI 200-2 ipd — TEVV-Athlon** (Aug 2026 draft) | Evaluation framework — COMMENTED | [nist-ai-200-2-tevv-athlon-comment.md](nist-ai-200-2-tevv-athlon-comment.md) | Our bench already fits its shape (axes = Blocks, harness runs = Events, scripts + verifier = Toolbox, scorecard = Stage 4). Comment sent 19 Sep 2026 asking for a runtime-decision Event type, an evidence contract (cites DEMM), evidence custody under Goodhart, and human oversight as an evaluable attribute (MAP 3.5 / MANAGE 2.4) — the draft has no post-deployment Event and never mentions human-in-the-loop |
| **STANAG 4774/4778** (NATO confidentiality labelling) | Military standard | referenced across the send-authority design + provisional drafts | Prior-art lineage for area/level send authority — the Org→Interchange mapping |

Adjacent but not assessments: competitive analyses
([XCHK-COMPETITORS-DEEP-DIVE.md](XCHK-COMPETITORS-DEEP-DIVE.md), SAM / Tuent / Klaimee /
ExecLayer / Agentomy docs) measure us against *products*; this index measures us against
*frameworks*.
