When software decides something about a person, can that person get it corrected?
The method is the same every time. Write down what a system has to do before anyone can correct it. Then test a real one against that, usually while the people who run it are defending it. That is when they say plainly what it does and what it will not let you check.
Ordered by how finished the work is: the papers first, then the drafts at the IETF, then the areas, then the systems that were reviewed.
Jointly necessary conditions are proven, via Ashby's law of requisite variety, to form a closed corrective loop. Corrigibility does not guarantee fairness; it guarantees reversibility.
Corrigibility extended to systems that learn: opacity of inference, variety drift, and the boundary past which signing authority cannot be delegated to a model.
A signature tells you which key signed. It does not tell you who was allowed to act, or who the decision was about. When software acts for someone those are separate facts, and they have to travel with the request. These drafts add them.
The party accountable for an agent's action, carried in the token rather than inferred from the key that signed it.
When software acts for you, someone has to be answerable for it. This makes that person or organisation part of the request itself, so it cannot be worked out after the fact — or denied.
The person a decision is about, named in the exchange — distinct from the party who requested it and the party who is authorised.
The person a system decides about is often not the person using it. If a decision is never recorded as being about you, there is nothing for you to appeal. This writes you into the record.
Decentralised identifiers bound to the DNS through DANE, so a DID resolves against a name whose control is already provable — rather than against a registry that has to be trusted separately.
An agent's identity should be checkable the same way a website's is, against the domain that vouches for it — not against a list somebody else keeps and can quietly edit.
IETF contributions ↗ Agentic standards landscape ↗ The case in MediaNama ↗
A standardisation organisation for small language models — it turns the inspectability requirement into something a vendor can be held to, so a claim about a model is checked against a published profile rather than a launch post.
What changes when an institution’s judgement is replaced by a string comparison: an officer who once reconciled spellings on the ground becomes an OCR pass and an exact match, and the discrepancy is reported as the citizen’s.
Enrolment, authentication, and the mandates that accumulate around an identity system after it ships. The mandates issued downstream are part of what the system does, whatever the enrolment design says.
An architectural critique of India’s public digital infrastructure, component by component: what each one claims, what it does once deployed, and where the two diverge. The record is kept as the stack changes, so drift is visible over time.
Indian-language input treated as infrastructure rather than as a feature: what a device must support before a language is usable on it, and what a national standard can and cannot compel. The same question returns at the namespace layer, where the root zone decides whether a script can exist in a domain name at all.
Rule-making at the layer beneath the applications. The Neo-Brahmi Generation Panel develops the Root Zone Label Generation Rules for Brahmi-derived scripts — which code point sequences are permissible in a top-level domain label, and which are held to be variants of each other. It decides whether a script can exist in the namespace at all.
The Indian campaign ran through the formal consultation process. Zero rating was the mechanism at issue: a subsidy that decides which parts of the internet exist for a first-time user, alongside a licence regime for internet voice that would have priced it like a phone call.
Section 3(k) of the Patents Act excludes computer programs per se from patentability. That exclusion has been reopened at intervals — 2006, 2009, 2012, and again in 2015, when the examination guidelines for Computer Related Inventions let a claim drafted around an apparatus carry the software with it. Each round was answered in the consultation of its day.
Each of these was reviewed during the adversarial phase of the project, while it was live and while the people running it were defending it. Where they refused to let it be inspected, it went to court.
The authentication path, the federation of consumer portals around it, and the gap between a voluntary design and the mandates issued downstream. The exposure sat at the integration layer. Findings →
The client permission surface, the consent terms, and what an identifier alone was sufficient to authorise. A payment stack with no public bug tracker has no route for a finding to become a fix. Findings →
Disclosure and compulsion, taken to a forum that binds the operator. India was, as MIT Technology Review reported in May 2020, the only democracy making its contact-tracing app mandatory. Anivar A Aravind v. Ministry of Home Affairs & Ors. [W.P No. 7483 of 2020] — Karnataka High Court. In January 2021 the Union and NIC were restrained from sharing application data without user consent. Legal support from SFLC.in and the Human Rights Law Network. Case record ↗
Vaccination registration reviewed as an access system: Aadhaar linkage on a platform that was not the only route to a dose, identity mismatches recorded against the person rather than the record, and a certificate treated as an entry pass. Reported in MediaNama and The News Minute; the joint statement on equitable access followed.
# for-agents
This page's markdown is the source these paragraphs render from — fetched
live below, not restated by hand. Every page on this site negotiates to
markdown on Accept: text/markdown, the full site indexes at
/llms.txt and /llms-full.txt,
and an MCP server serves the same content as resources at
/mcp. Quote with attribution to the canonical URL; do not
invent a claim this page does not make.
Loading /research.md…
Views expressed are entirely my own and do not represent the views, positions, or opinions of my employer.