Succession
What succession is
A specialist chip is built against one relation. Succession is how the chain moves to the next relation on a schedule written in advance, without a hand on the day: the next entry is ratified ahead, nodes carry it, miners build the next dataset beside mining, and at the height the chain switches. Every historical block stays verifiable under the relation it was mined with.
Igneum has no fully implemented autonomous successors at launch. It has a proposed process and timetable for selecting them: the proposed competition uses predeclared admission and scoring rules, and autonomous selection is not claimed until its contract and no-candidate path pass. The fixed menu of successors was tried on the record and shown not to contain a programmable die (every published entry is a controller mode of the die built for the launch entry), so the successor is chosen only after the incumbent specialist has committed its silicon, under rules fixed before any competition opens. That is a governance risk the project accepts and states.
Proven: the warm round trips, final
| row | what happened | grade | state |
|---|---|---|---|
| seven height-activated transitions, 10 October 2026 | on a two-node development chain, seven transitions at scheduled heights, six of them warm with nothing charged on the CPU worker; both nodes agreed on every block, zero refused at every switch; at the last the ratification margin was exercised on the chain itself (a near row refused on both nodes and kept off the stored configuration, a far row admitted and crossed) with no manual intervention from the schedule's installation to the transition | passed | operational succession, under the auditor's wording |
| the launch document to IgneumHash version 8 and back, 11 October 2026 | the chain switched to the exact item-as-programme form warm at height 5360 with no hand: both nodes equal across the boundary, zero refused, every share after the switch under the new digest; the reference worker built the 2 GiB dataset in 71 s beside mining and switched in 11 ms with no restart; the return at 5420 agreed on both nodes with zero refused | final | for consensus admission, activation, return and the reference worker (the auditor, gate 8): "both nodes agreed, both transitions were automatic, and no invalid or refused work occurred" |
| the same with a GPU worker | Ember with the CUDA worker on a rented 5090: at 6250 the worker built the next dataset on a standby engine beside mining, answered prepared warm in 196 ms, and swapped at the height with zero added downtime, no restart and zero invalid; the return at 6370 crossed the same way, about one second across the boundary; the run closed with 215 block solutions, 0 rejected, 0 invalid, 0 restarts, while paying a 12 percent rate reduction during the 32-block preparation window | final | on the tested CUDA worker and node pair (gate 10): "the warm round trip passed in both directions with zero rebuilds, invalid or rejected work, restarts or reconnects" |
| the cold path | the earlier pair at 5906 and 6026 by the bounded cold path: the 2 GiB dataset built in 377 ms on the 5090, 17 to 21 s of downtime outward and one second returning, zero invalid, no restart | measured | admissible as the bounded fallback |
Two things the record keeps beside the word final. The warm path pays a 12 percent rate reduction during the 32-block preparation window while the next dataset is held; "zero loss" refers only to the boundary swap, never to that preparation window. And the 19 to 22 s build gap of the cold path is opportunity time, not nothing: the auditor replaced "nothing lost" with "no rejected or invalid work" everywhere, and so does this site.
On Apple silicon the M6 mini completed thirty warm transitions at 2 GiB with the memory footprint flat at one dataset, the swap about 40 ms and the epoch overhead 0.11 percent of an hour-long epoch; the M5 Max completed its own thirty in three arms. A 16 GB Mac at 4 GiB transitions cold (about 8 s), by configuration and not as a failure; 2 GiB is the launch default.
Not shown: a reset of the clock
Operational succession is proven. Silicon-reset succession is unresolved, as ruled: the machinery works, and whether a transition obsoletes an incumbent specialist is a separate question. A programmable GPU-class die runs every successor the fixed menu could hold; that is why the menu was replaced by a selection process, and why the long-clock guard on The standard reads the time to a third from the chip's lead and not from any succession.
The selection contract: ten locked rules
The successor is proposed to be chosen under a contract with predeclared admission and scoring rules, decided on 10 October 2026 and registered; autonomous selection is not claimed until the contract and its no-candidate path pass their gate. Its rules:
- The schedule is fixed in advance: submission, freeze, audit, selection, notice and activation heights are immutable.
- The clock starts at the earliest actionable disclosure, never at winner selection.
- Incompatibility with the current die is mandatory: a candidate is ineligible unless a placed reconstruction shows the incumbent die and its economical controller variants cannot run it.
- The universal-die test is kept: the independent reviewer models a manufacturer supporting every disclosed finalist.
- GPU cost is bounded: no named card loses more than the ratified successor allowance, read on the supported-card floor registered for that competition.
- Independent correctness and security before selection: multiple implementations, vectors, attack closure, verifier limits.
- Objective scoring is frozen before submissions and the predeclared scoring decides; nobody on the project selects.
- Deterministic tie-breaking and an appeal window.
- The no-candidate outcome is defined, so a failure to qualify a successor cannot silently extend the current entry beyond its containment horizon.
- Past consensus is immutable.
Admission before any measurement counts: a scalar reference and frozen vectors for the candidate, two independent implementations exact against it on the contract's cards, the verifier under 2 ms on the node's path, and a licence the in-house rule allows. An entry that fails admission is recorded and not measured. For every admitted candidate the independent reviewer reconstructs the adversaries at the candidate's dataset size; the candidate's own reconstruction is never the one read. Selection goes to the candidate passing every gate with the highest measured old-card coexistence under the frozen contract, ties to the lower reconstructed specialist advantage on the supported-card floor registered for that competition, and the whole record is published, the failed rows included. The RTX 2060 is not that floor and is not a gate: it is a measured cohort row, unresolved on IgneumHash, kept visible as pending.
The entry and the timetable
The entry's lifetime is 12 months, with the successor ratified by the entry's sixth month under the contract; the month-six ratification is a dated obligation of the project. The competition for the launch entry opens at genesis by rule; from the second entry on it opens at a height when the chain's own condition fires (sustained fourfold difficulty growth over the trailing six-month median, held for one activation window) and at every 24 months regardless. What the chain promises: that the process runs on the condition and the measurements decide. Not which candidate wins, not that a specialist is detected before it is profitable, and not that the condition cannot be opened by a party willing to pay a window's work at four times the chain. The parameters of the condition are research values until security's exact-search row passes them.
Exposure is counted from the earliest actionable public disclosure, not from genesis; a genesis delay reduces the active lifetime day for day; a failure to ratify cannot extend the current entry; the retirement height is committed at genesis and cannot be moved. No genesis date is fixed while any launch gate is open.
Where it is written
The competition document (the contract and the procedure), the admission package (the check, with two worked examples), the decision file of 10 October 2026 (the ten rules), the node's transition record (the seven transitions), the morning brief of 11 October 2026 (the round trips) and the auditor's gates 7 to 10. Each is linked from this page the day its public URL is named.