The Calibrated Authority Index codes how knowledge institutions construct trust in generative AI. It is a small, opinionated instrument applied with one discipline above all: every score traces to the institution's own public policy and a verbatim quote.
Trust does not form one way. Humans extend it forward, on the expectation of reciprocated fair dealing — Anticipatory Reciprocity. Agentic systems earn it in reverse, in proportion to what they can prove — Calibrated Authority. The claim of this corpus is that the split is driven not by substrate but by verification economics: where proof is cheap, trust is granted evidentially and is fungible; where proof is scarce, trust runs forward on relational grounds.
We trust machines exactly as far as we can check them — past that, it's faith.
The Index is the observed-in-the-wild evidence. 55 institutions, each coded; mean Calibrated Authority 9.8/12.
Each institution's public AI policy is scored on six dimensions. Each dimension takes 0 (absent / silent), 1 (partial / implied), or 2 (explicit / enforced). The six sum to a composite Calibrated Authority (CA) score from 0 to 12.
| D1 | Traceability & inspectability |
| D2 | Human authorship & accountability |
| D3 | Disclosure & labeling |
| D4 | Synthetic-identity / fabrication prohibition |
| D5 | Human validation in loop |
| D6 | Evidential-trust emphasis |
For each institution we locate its public, generative-AI-specific policy, read it in full, and assign each dimension by what the policy actually says — not by reputation or intent. Alongside the six scores we record three reads: posture (Prohibitive · Balanced · Enabling), verification-boundary fit (does the policy land where the thesis predicts — fits · partially · contradicts), and trust-logic (Evidential · Relational · Both-split · Neither). We also flag twilight framing — precedent collapse, feedback delay, exponential fog — where the policy reasons that way, usually without naming it.
The Index practices the discipline it measures. No public claim ships without a traceable source. Every record carries the institution's own policy URL and a verbatim quote — the exact line the score rests on. The author has no relationship to the coded institutions; scores are read from public documents, and the underlying data is open (JSON, CSV, per-institution atoms) so any reader can re-derive a score from the same evidence.
The Index stands next to two strong traditions, and owes something to both. Archival projects — the Open Terms Archive and the University of Bremen's Platform Governance Archive — continuously version the terms and policies of commercial AI platforms, keeping a daily, diffable record of the raw text. Academic audits score the substance: cross-sectional studies code the AI policies of universities, journals, and newsrooms on substantive dimensions at a single point in time, the nearest being a Peking University study tracking the author-facing AI policies of 5,114 journals across two waves (arXiv:2512.06705, Dec 2025). Archivists version the text but don't score it; academics score the text but don't version it. This Index does both at once — it continuously versions and scores knowledge institutions' AI policies on one multi-dimensional instrument, and, to the author's knowledge, it is the only one measuring how the division of work between humans and AI moves over time.
verify_status in their machine record; treat their quote attribution as good-faith but secondary.The Index and all its data are released under CC-BY-4.0. To cite the dataset:
To cite a single institution, use its permalink page (https://calibrated-authority.chrishuberreitz.com/institutions/<id>) and the verbatim quote carried there. A machine-readable record for every institution lives at /institutions/<id>.json, and the full corpus is enumerable in one fetch at /institutions/index.json.