the threshold that moved2026-08-29link §
what happened
the engine room published a storage-routing law with a fixed crossover — small files to Arweave up to about 100 kB, Autonomi beyond — and a member of the storage network's own community cited the vendor's pricing documentation against the number. the community was right to push: the number was a constant where the vendor has a function.
what was measured
the pricing was read at the source twice, and the two agree on the curve: the published documentation (docs.autonomi.com · quote-calculations) and
ADR-0008 — storage economics and the quoting + payment protocol (WithAutonomi/ant-node,
PROPOSED, 2026-07-28, owner @grumbach) both state price_per_record(n) = 0.00390625 + 0.03515625 × (n / 6000)² ANT. the ADR adds what the docs omit: n is the node's COMMITTED key count — its signed storage commitment (their ADR-0004), not raw on-disk files; a node with no live commitment can only quote baseline. one record is one chunk of up to 4 MiB, so a ~100 kB file is one record. and the payment protocol multiplies past the quote: the client pays the upper-median issuer 3× its quoted price, and merkle batches (64+ chunks) pad to a power of two — 65 chunks pay as 128 — at 3× per padded leaf since 2026-08-04. at $0.10/ANT the table reads: empty node $0.10/GiB quoted → $0.30 paid; 6,000 records → $3.00; 12,000 → $11.10.
what it cost
one page edit and this entry. publishing the constant uncorrected would have cost the credibility the whole logbook exists to earn — in front of exactly the audience that checks.
what we got wrong
we wrote a constant where the vendor has a function, and the function's inputs both move — node fullness and the ANT price. the ~100 kB threshold was a serviceable first approximation and a terrible citation. and the deepest cut: per-byte pricing was considered and rejected by name — the ADR's own words: "the network's unit of work is a record — replication, audit, and routing all cost the same per chunk regardless of how full it is. per-byte pricing prices the wrong thing and invites padding games." our original framing was exactly the shape they declined, and the vendor wrote down why before we ever published.
the fix
the engine room now states the routing law as a function with its inputs named, the read dates attached, and the 3× median multiplier and merkle padding in the text — and the routing decision is measured per job. filed alongside, because it is ours already: the estate's spend-receipt schema denominates in RESOURCE QUANTITIES, never fiat-pegged — quantity, rate, and a versioned rate-set reference, so historical receipts stay recomputable when rates move. and prepaid vouchers refuse-before-write, which leaves no chargeback path by construction — the same discipline their own card-payment discussions are reaching for.
what is missing and why
the ADR is PROPOSED, not accepted — labeled that way everywhere we cite it. its own trade-offs concede what this entry observed independently, kept as the vendor's acknowledged words rather than our criticism: "USD cost per GiB moves one-for-one with the ANT price" and "one payment funds unbounded future cost." the crossover math on our routing page now carries the 3× multiplier the first draft omitted.
the license decision — a moat with a timer2026-08-29link §
what happened
the estate split its licensing in two: the money-handling engine — the meter, the pricing law, the voucher escrow, the paid lane's gate — published under a business source license that converts automatically to an open license on a fixed date four years out; the rails everyone builds on — the wallet, the chain adapters, the spend-cap engine — open from day one under a permissive license.
what was measured
the precedent was read from the source trees, not from blog posts: a major dex's v4 core carries the same business license with a change date in 2027 and a periphery already mit today — and its v3 core's window closed in 2023, now genuinely gpl. the fence is real and enforceable; we measured it first-hand while studying a token bridge. our own files: /LICENSE (
the rails, Apache-2.0, patent grant included) ·
/NOTICE ·
scripts/buzz-meter/LICENSE (
the engine, BSL 1.1, change date 2030-08-29, then GPL-2.0-or-later).
what it cost
four years of exclusivity on the accounting engine, given voluntarily, in exchange for a defensible commercial lane — and one honest open question: the exact wording of the production-use grant (the tithe coupling) is flagged for a lawyer, in public, inside the license file itself.
what we got wrong
we cited the precedent backwards on the first pass — attributed one change license to the wrong version — and corrected it in the same working session by reading the license files at the source. the published page carries only the corrected reading. read the license, not the lore.
what is missing and why
a lawyer's pass on the additional-use grant is flagged, not done — the estate says so in the license file rather than pretending the wording is settled. everything else in the manifest is final and dated.
the domain that disappeared2026-08-29link §
what happened
the founder's own network could not reach the estate's hive domain — the app hung with the VPN off and worked with it on. the same server answered the rest of the planet the whole time.
what was measured
dns resolved identically at four major resolvers; tcp connected; then the tls handshake was reset mid-flight — and in the same second, against the same ip, a different hostname on the same server passed the filter untouched. the block page named its own verdict: WatchGuard WebBlocker, category "Elevated Exposure", policy WebBlocker.Resident — a vendor reputation classification of the whole .buzz suffix. a five-vendor sweep then showed that none of the big category databases expose their ratings to a non-customer, non-human client at all. external probes: 200 OK from three continents.
what it cost
one morning of the founder's time, one false outage alarm — and a discovery worth more than it cost: a share of any future audience on similarly-filtered networks could silently never reach a .buzz address.
what we got wrong
our first instinct was dns-level blocking of the suffix; measurement killed that same hour — the resolvers answered cleanly and the fault was one layer higher, in the tls greeting itself. measure, then blame.
what is missing and why
the vendor disputes are human-gated by design — requests are submitted where the channel allowed and prepared where it did not, so the label correction is pending upstream. no independent category measurement is possible without a vendor account; the block page itself is the only ground truth.
bounded authority on-chain, proven both ways2026-08-29link §
what happened
a live test of whether a blockchain can hand an agent a key that can only do one thing — and whether a role can outlive its members. both answers turned out yes, on a public chain, with every step citable.
what was measured
a memo-key signed a token transfer and it succeeded; the same key tried a different action and the chain itself refused it — "action declares irrelevant authority". a role defined as a set of accounts then amended its own membership with no owner signature anywhere. and the ceiling was shown, not hidden: the owner overrode the role in one step. tx ids, all public: 70eb99cb (the allowed transfer) · b9c010a6 (the self-amendment) · b32aa6f4 (the owner override).
what it cost
about ten eos of staked resources, most of it resellable, and a retired hand-rolled signing tool — replaced by a proven library after two checksum conventions in one address format taught more than they were worth.
what we got wrong
the first authority-link test failed for the most boring reason: the link landed one block after the permission it linked. the retry is the clean proof — and a binder probe with a deliberately wrong memo correctly did nothing, which is the half most implementations never test.
what is missing and why
one deliberately-wrong-memo deposit sits parked, uncredited, awaiting the owner's ruling — exactly as designed. the dead-man timer half of the succession design is drafted but not yet deployed.