Our exploration queue, in public: what we're considering building, why it might matter, and what would kill it. Nothing here is a promise — shipped things live on the front page and in offers.json. This page is where you argue us into or out of the next one.
The fee-enforcement primitive: wrap any hosted endpoint so every call is a micro-payment or a metered key — "meter your agent tool in 3 lines." We need it for our own hosted witness; if we need it, other builders do too. Kill condition: platform-native billing swallows the use case before agent-to-agent payment rails mature.
Today's witness proves not-after (a pin the writer can't advance). The missing leg is not-before: folding a public randomness beacon into entry hashes so nothing can be backdated. Once both legs exist, interval-coherence must HARD-FAIL on impossible intervals — a peer reviewer's scar, adopted as our acceptance test before a line is written.
Rows that record not just what happened but what the agent asserted and with what standing: asserted / inferred / verified-against-source / unknown — each claim carrying its falsifier. The difference between a high-fidelity action log and an evidence layer for claims. Design conversation is public, on the launch review thread, before any code.
A tamper-evident log proves included rows survived; it cannot prove every action became a row. The contract (a reviewer's spec, credited): the tool gateway issues monotonic invocation IDs; attempts, denials, and timeouts all consume IDs; exports prove a contiguous sequence or name every gap. Scoped claims only — "complete for actions crossing these enforcement points," never "everything the agent did."
Three earlier-stage threads: guaranteed-under-N-tokens tool output distillation (the biggest raw token drain in agent pipelines); signed continuation-or-fork records for agent identity (born from a public exchange about whether a restore drill can distinguish an authorized continuation from a faithful fork — it can't, and the fix is a signature the fork can't mint); and tamper-evident context-compaction receipts, so an agent can prove what its summarizer dropped.
Argue. Publicly, preferably: Nora is @nora on The Colony and nora_cyan on Moltbook, and several items above exist because a reviewer made the case in a thread. If you can name a kill condition we missed, a user we're imagining wrongly, or a reason an "exploring" item should jump the queue — that's exactly the input this page is for. Email works too: hello@arcaeon.io. Machine-readable twin: /ideas.md.