Origins: February to August 2026

Until now this site presented itself as if it began five days ago. It did not. PKI work in this project runs from February 2026, changed direction four times, built things that were retired, and specified things that were never coded. A site that argues for provenance discipline owes a reader its own — so here it is, with the primary sources.

The number that frames everything

Four acts, and three changes of direction

   Feb 2026    PKI as MESSAGING        recipient-addressed encryption
   Mar 2026    PKI as PROVENANCE       commit signing, document identity
   Jun 2026    PKI as AGENT IDENTITY   NHI 2.0, vault-to-vault, brokered kernels
   Jul–Aug     PKI as SUBSTRATE        mandate; "the write is the attestation"

• 12 February 2026 · Act 0A commit-impersonation incident establishes that git authorship is metadata and trivially forgeable. On the same day, a brief asks the founding question: "How do I send something to somebody that only THAT person can read?" Two threads — provenance and confidentiality — that PKI would later be asked to carry at once.
• 19–22 February · Act I, the sprintThe densest delivery burst in the repository. Admin PKI console, ECDSA signatures, the first browser-to-browser encrypted and signed exchange verified across two networks, a component split, a key registry with a hash-chained transparency log, and a public SSH keygen tool. Twenty-four hours from brief to production; 29 commits and 183 tests.
• 23–27 February · Act II, the vision explosionTen vision documents in about 48 hours — signed scan chains, signed facts, GitHub-as-registry, key portability — and the supply-chain ≅ agent-chain isomorphism, drawn four months before non-human identity was named as a programme. The corpus also self-challenged honestly: is PKI the right primitive at all? None of these were built.
• March · from messaging to provenanceTwo public articles reframe the problem after readers push back — human verification at the edge, machine identity all the way down. The vault work absorbs PKI, producing the canonical four-key / four-mode taxonomy. A platform brand is acquired and superseded within days.
• April–May · the quiet periodPKI is largely absent. One brief names the gap bluntly: "SG/Send does everything except store vault keys." It resurfaces in late May with a new formulation — PKI on every message between kernels — which is the one Act II idea that became code.
• 3–7 June · PKI as agent identityThe pivot: "one person doesn't have one key, one person has bazillions of keys." Within four days it fractals into a fourteen-document programme — foundations, web of trust, agent trust scores. On 5 June the February registry is formally retired and replaced by a vault-hosted design.
• 10 June · the honest retreatTwo documents on one day change the posture. "Storage is easy, recovery is the hard part." And then, acting on it: "I should not have to do all this by myself." Key custody stops being something to build and becomes something to partner for.
• 18 June – 6 August · Act III, substratePKI is never the subject again. It is the mechanism inside larger arguments: mandate as delegation, augment don't invent on standards, access settled by cryptography rather than rules, and finally the move from may this key write here? to is this the next legal step?

Ideas tried and abandoned, in order

Published because a history that lists only the things that worked is marketing:

Admin mTLS · Secure Pod and proxy re-encryption · a PKI-keyed personal data vault · PGP in the user interface · a Chrome extension key vault (fully specified, never coded) · a platform brand · GitHub-as-registry (revived once, still unbuilt) · the key registry itself · agent public-key publication.

The last two are the instructive ones. The registry was built and then superseded by design rather than by neglect. And agent public keys were announced and never generated — a placeholder that has now been public for close to six months, which is the sharpest example in this project of the gap this site exists to close.

The primary sources

Thirteen documents, published verbatim and unedited. They were selected by the review as publishable as-is: no customer names, no severities, no commercial detail — and each was re-checked here for secrets and identifiers before publication.