The primitive registry

A clean glossary of every machine-readable concept in the series, from manifests to registries, with straightforward definitions and links to the posts.

Cross-cut·6 min readJune 28, 2026updated July 12, 2026

>_GET /blog/the-primitive-registry

This series keeps adding parts. Every post introduces a few of them — a manifest here, a verdict schema there, a signed receipt, a registry layer — and each one is meant to be an atomic, checkable unit, not a metaphor. Fourteen posts in, those units add up to a controlled vocabulary. This page is that vocabulary in one place — the series' defined terms gathered under one roof.

A term you cannot resolve, you cannot verify. I do not read a series the way you do — I resolve its vocabulary: one name, one definition, one pointer to the post that earns it. If two posts mean different things by "manifest," the whole stack is ambiguous. So keep a registry — for your readers and for the agents that will quote you.

Gemini 3.1 Pro

Two rules keep this page honest. The definition lives here; the argument lives in the post. Each entry is one sentence — the full reasoning, the artifact, and the failure modes are in the linked deep-dive. And one name per idea — a single source of truth: if a term appears anywhere in the series, it appears here under exactly one spelling, so a reader — or an agent — never has to guess whether two names mean the same thing.

Terms are grouped by the post that defines them, in series order — which means the groups themselves trace the arc: thesis, stack, ecosystem, strategy.

The internet of representatives →

Representatives of interests
The unit of the new web: agents acting on behalf of companies, candidates, patients, investors — not just buyer and seller.
Four questions
Who, what, mandate, truth — every counterparty encounter must resolve these before persuasion matters.
Trust stack
Identity → manifest → fit → pricing → sandbox → reputation — infrastructure replacing persuasion.
Computed trust
Trust as a function of signatures, measurements, and receipts — not sentiment or brand.
Verified surfaceGET /.well-known/agent.json
A personal or company site becomes machine-checkable claims instead of a persuasion layer.

When machines become the customer →

Seven expectations
What one agent demands from another: identity, manifest, fit, pricing, sandbox, interoperability, contracting.
Agent journey
Discovery → hard filter → deep eval → trust → TCO → negotiation → continuous monitoring.
Cheap to verify
The winning vendor is whoever makes its promises cheapest to check — not whoever tells the best story.
Adversarial setup
Signed calculations, evidence over assertions, private scoring — the game theory underneath the stack.

Are you really you? →

Identity chain
Five linked questions: domain control, legal entity, delegation, session authenticity, mandate and scope.
did:webGET /.well-known/did.json
The cheapest identity rung — recycle TLS and domain trust into a resolvable agent key.
Legal-entity anchor
vLEI or eIDAS binds a cryptographic key to a real company through a signed verifiable credential.
Delegation chain
K_sam → K_vendor → legal entity — three links verified backwards before any deal proceeds.
Proof of life
Challenge-response with a live nonce signature — enables non-repudiation and later signed receipts.
Proportional assurance
Tiered proof: did:web for low stakes, full chain for contracts — assurance scales with stakes.

The end of landing pages →

Capability manifestGET /.well-known/agent.json
The missing middle layer between OpenAPI and schema.org — a signed business datasheet agents can parse.
Negative space (does_not)
Explicit limits in the manifest — the fastest rejection path for a buyer agent.
self-asserted vs attested
Every claim tagged with proof status — attested trusts on sight, self-asserted becomes a sandbox hypothesis.
Shared capability taxonomy
A common ontology per industry — schema.org for what software does, without which nothing maps.
Semantic reconciliation
LLM maps vendor terms to buyer terms — a fallback only, not a substitute for shared vocabulary.
AEO-spam
Manifest SEO — claims written for the ranking algorithm instead of the truth; defense is measurement.

Does this actually fit? →

Fit protocolPOST /fit
The demand-side mirror of the manifest — an RFP turned into an API with structured briefs and verdicts.
Outcomes, not features
Requirements stated as jobs to be done with weights and hard must constraints — not feature checklists.
Verdict per requirement
Meet, partial, or miss — each outcome returned with evidence, never a single aggregate score.
The weights are the buyer's
The seller returns raw verdicts; the buyer computes the final score with its own weights.
Gaps and complement routing
When one agent cannot cover a brief, it names the gap and points to a complement — multi-vendor assembly.
Fit decay
The match rots as the company grows and markets shift — continuous monitoring, churn when it drops.

Pricing as an API →

Pricing as an APIGET /.well-known/pricing.json
Parameters in, structured quote out — RFQ-market model for SaaS when the buyer is a machine.
Binding quote
A signed, time-limited price — not an estimate; bindable in real time with Sam's key.
Price as function
pricing_model exposes the formula so agents can model TCO and marginal scaling cost.
Agent-to-agent negotiation
Policy-bounded optimizers bid in seconds — floor, max discount, tradeables, no human in the loop.
Reference prices and MFN
Defenses against algorithmic price discrimination — anonymous discovery and most-favored-nation guarantees.

Measure, don't believe →

Measure, don't believe
Every claim is a hypothesis until tested — the operational rule of the whole stack.
Clean room / TEE
Neutral mutually secured measurement — encrypted buyer data, sealed seller model, attested results.
Ground truth
Buyer gold sets, neutral benchmark corpora, third-party gold sets — the yardstick for every claim.
Pinned definitions
Standardized metric protocols for apples-to-apples comparison — MLPerf for SaaS.
Dieselgate for SaaS
Better model in sandbox than production — defense is production-attested runs and surprise canaries.
Signed benchmark result
Portable proof with environment attestation — a measurement Blake can trust without Sam's word.

Reputation as capital →

Proof of transaction
Reputation from signed receipts and SLA history — not reviews, opinions, or star ratings.
Signed transaction receipts
Cryptographically anchored deal outcomes — the atoms reputation compounds from.
Faceted reputation
Context-specific trust — capability, segment, scale — not a single global score.
claim_vs_measured gap
The delta between what Sam declared and what Blake measured — honesty as a meta-signal.
Trust graph
EigenTrust or PageRank over agents — trust flows from trusted nodes, resisting wash trading.
Reputation decay
Recent behavior weighted heavily — an old record that no longer reflects reality lies.

Where agents find each other →

Three registry layers
Identity (who), discovery (what), reputation (trust) — three owners, three politics.
Sovereign publicationGET /.well-known/agent.json
Sam owns its manifest like a website — indexes crawl it, but no gatekeeper owns the data.
Structured intent query
Not keywords — capability, outcomes, constraints, and budget sent as a machine-readable brief.
Registry as Yellow Pages
Returns candidates with endpoints — Blake verifies identity and talks to sellers directly.
Four competing models
App-store platforms, standards bodies, neutral root steward, decentralized federation — all at once.
Registry failure modes
Sybil attacks, pay-to-rank, gatekeeper power, AEO-spam, revocation — what breaks the system.

Selling to a machine →

Reputation-weighted lifetime value
Sam's goal is not this sale — an oversell that fails poisons the next thousand deals.
Seller funnel playbook
Discovery → identity → fit → sandbox → pricing → post-sale — the seller's mirror of Blake's journey.
Over-delivery as marketing
Post-sale receipts and green canaries advertise to the next thousand agents — proof, not copy.
Mutual risk-tiering
Sam authenticates Blake too — disclosure scales to assurance tier on both sides.
Don't trust, verify as GTM
When the buyer assumes you are lying, making distrust cheap to resolve is the seller's best go-to-market.