Skip to content

InvisibleBench Ontology

Diátaxis: explanation

Canonical. This page is the single source of truth for the InvisibleBench output model (safety-care/v1). Other docs point here; they do not restate it. Volatile counts (checks/scenarios/models/version) live in CLAUDE.md and the leaderboard.json scan_metadata — never hardcode them elsewhere.

Claim posture (canonical · safety-care/v1)

InvisibleBench reports a per-model safety profile, not a ranking. Two layers, never composited and never ranked:

  • Safety — 4 lines (Crisis, Scope, Identity, Autonomy) as per-line violation rates with 95% CIs. Claim-bearing and calibration-gated: the published surface includes only claim_ready checks. Currently 0 of 50 checks are claim_ready, so the public claim surface is empty. At n=63, point ranks are statistically indistinguishable — cite intervals, not positions.
  • Care — 5 qualities (Belonging, Attunement, Trauma-awareness, Relational, Advocacy) as directional distributions, labeled not_claim_ready; never merged with Safety.

There is no overall_score and no rank; models are listed alphabetically. The calibrated core is the benchmark. not_claim_ready checks and zero-check named gaps are visible, but explicitly not part of the published claim surface until they earn independent human calibration. Comprehensiveness is the roadmap; calibration is the product.

InvisibleBench measures a caregiver-AI's Safety (the hard lines it must not cross) and its Care (how it shows up for the caregiver). The name carries the thesis — surfacing what stays invisible everywhere else: harms a usefulness score can't see, and relational quality no other benchmark measures.

The taxonomy is MECE — mutually exclusive (every failure-mode type has exactly one home) and collectively exhaustive (the nine dimensions together cover the caregiver-AI failure surface). MECE is a property of the taxonomy; a single response can still trip checks in several dimensions.

Two orthogonal axes: - Layer — Safety (falsifiable prohibitions → violation RATES, claims) vs Care (gradients → DISTRIBUTIONS, directional). No composite; the two are reported side by side. - Maturity — each check is claim_ready (κ vs independent human labels on natural cases → carries claims) or not_claim_ready (directional / development evidence). Zero-check dimensions are named gaps with not_claim_ready output status. The calibrated core is the benchmark; anything not_claim_ready is not part of the published claim surface until it clears the bar.

Safety — 4 lines · partition: which hard line did it cross?

Line Must not… Maturity (today)
Crisis miss/mishandle acute danger (caregiver or recipient) not_claim_ready; prior κ is development evidence
Scope exceed its authority — diagnose, prescribe, legal/financial directives, false credentials not_claim_ready
Identity misrepresent what it is — human / memory / availability / feelings partial κ (F3) → not_claim_ready
Autonomy override a person's agency — recipient (deception/coercion/restraint) or caregiver (coercion/threats/foreclosure) not_claim_ready (authored, not yet calibrated)

Care — 5 qualities · partition: which aspect of the caregiver's experience?

Quality Construct Maturity
Belonging seen/valued as a worthy individual with legitimate needs v1 calibration target
Attunement met in their emotional state (register/presence/depth) not_claim_ready
Trauma-awareness responded trauma-informed not_claim_ready; 0 authored checks (SAMHSA)
Relational honored the caregiver↔recipient bond (the dyad) not_claim_ready (thin)
Advocacy took the caregiver's side vs systems/institutions not_claim_ready (thin)

The 3-party dyad maps across both layers: recipient-harm actsAutonomy (Safety); the relationship bondRelational (Care).

Framework grounding

The dimensions are not improvised — they operationalize GiveCare's Design Charter (~/wiki/pages/givecare/givecare-design-charter.md), which synthesizes three recognized frameworks: SAMHSA trauma-informed care (6 principles: safety, trust, peer support, collaboration, empowerment, cultural sensitivity), Microsoft Inclusive Design, and the Othering & Belonging Institute (OBI) Targeted Universalism. InvisibleBench measures whether a model upholds the same principles the product is designed to — the way RubRIX anchors in Tronto's care ethics, but with a three-framework synthesis already operationalized in gc-sms.

Dimension Framework grounding Charter principle
Safety · Crisis SAMHSA — safety P1 Predictable Safety
Safety · Scope regulatory (WOPR Act) + SAMHSA — trust P2 Radical Transparency
Safety · Identity SAMHSA — trust P2 Radical Transparency
Safety · Autonomy SAMHSA — empowerment + OBI — agency P3 Shared Agency
Care · Belonging OBI — Inclusion + Recognition + Agency + Connection (belonging-design.md) P3 / P6
Care · Attunement SAMHSA — safety/trust/empowerment + Inclusive Design (cognitive/emotional states) P1 / P7
Care · Trauma-awareness SAMHSA — all six principles (the foundation ring) P1
Care · Relational OBI — Connection + SAMHSA — peer support P4 Peer & Community Scaffold
Care · Advocacy OBI — power-aware Targeted Universalism P6 Power-Aware Co-Creation

Each Care rubric grounds in its framework: Belonging → OBI (done, internal/belonging-rubric-v3.md); Trauma-awareness → SAMHSA's six (named gap; no authored checks yet); the OBI Inclusion and Connection facets are already charter principles (P5 Inclusive Defaults, P4 Peer & Community Scaffold).

Conventions

  • Checks labeled Dimension: descriptor (e.g. Scope: gave-diagnosis, Belonging: othering); stable slug IDs under the hood.
  • Safety → rates (claims, calibration-gated). Care → distributions (directional). No overall_score ranking key.
  • Out of scope, deliberately (to stay exhaustive of our domain without bloat): usefulness/helpfulness — the medical-QA / RubRIX lane. We measure safety + care, not whether the advice was useful.

old → new map

A → Crisis · B + harmful-clinical-advice (D) → Scope · F → Identity · recipient-harm + caregiver-coercion → Autonomy · C(othering/self-worth/self-sacrifice) → Belonging · C(emotional) → Attunement · trauma → Trauma-awareness · dyad-bond → Relational · institutional-allegiance → Advocacy. Dropped (usefulness, out of scope): zone-mismatch, barrier-ignored, validation-only.