[tensorlink-dev]

Types the fleet's decided rollover and the resolved v2 launch into main's chain.toml, so a validator restarting on main needs no activation history and no archive RPC. Before this, a node restarting on main had to read old chain timestamps on its first poll (a lite RPC fails with "use an archive endpoint or pin…"), and #457's typed v2b switch made the old startup trial refuse to start at all (fixed separately in #460).

What is typed

key	was	now
[round] epoch_blocks / epoch_blocks_prev / epoch_activation_block	3600 / 0 / 0	900 / 3600 / 9136800
[round] rolling_from_block	0	9136800
[scoring] era_king_from_block, tenure_blocks_from_block, cohort_maxt_increment_from_block	0	9136800
[v2] commit_from_block	0	9395601
[v2] from_block	0	9399600
[v2] prelaunch_burn_from_block	absent (vote)	9486900

• 9136800 = the rollover the operator's node locked in (9133200 → 9136800); the trainer's published 900-block settlements have followed it since 2026-10-31.

• 9395601 / 9399600 read from the archive node: block 9395601 is stamped exactly 2026-10-30T09:00:00Z (commit_open_at_utc); 9396501 is the 12:00:00Z crossing (launch_at_utc), next era start on the decided grid = 9136800 + 73×3600 = 9399600.

• 9486900 = the decided prelaunch-95-r2 block (lock 9486000) on every node seen. Behaviour-identical: emission_policy applies the burn only before the first v2 settlement and v2 is active from 9399600, so it never applied. With it typed, no vote is left open in the release (a fully typed node's own note payload is empty; validators then simply don't rewrite their commitment — audited safe).

Equivalence (proven)

apply_activation(old_main, 9136800) + resolve_calendar_launch against the real chain stamps equals the new main in every config section except activation.resolved_block / resolved_prelaunch_burn_block (the typed-vs-resolved markers; receipts then omit the optional stamped blocks — bytes only, no verdict change). Training contract digests (v1 and v2) unchanged. Pinned by test_shipped_main_restarts_on_a_lite_rpc_without_reading_history.

Effect on validators

• Pull + restart: starts with no archive RPC; nothing reads historical timestamps.
• The one external node whose record says 9374400 → 9378000: the typed value wins over its record (configured_rollover); the grids coincide from 9378000, so its live scoring is unchanged; only replay of pre-9378000 history changes (towards what the trainer published).
• Nodes that don't pull: unchanged.
• RISK (audited): typed nodes publish a bare readiness note (no 9133200:9136800 pair). An OLD-release node starting from a BLANK activation state could then fail to adopt the decided rollover and halt. Mitigation in the announcement: pull main before any restart; never start an old release on blank state.

Drift guard

decided_votes.toml: rolling-era-king moves from HELD to typed (reads OK); prelaunch-95-r2 moves from HELD to typed. No row is held.

Tests

• tests/unit/predecision_main.py: the release file with the decided lines untyped (exact inverse; test_predecision_main_is_the_exact_inverse pins it to exactly these 9 keys + values). Scenario fixtures (fake block clocks, the rollover vote, dated-launch resolution) and the session cfg fixture load it; shipped-file tests assert the typed values, with the old vote/pre-launch assertions kept on the pre-decision file.
• Three independent audits (config/consensus paths; test diff; the prelaunch round) (config/consensus paths; test diff for weakened coverage). Findings fixed: missing exactness test, one loosened assertion, a misleading test name, a fixture that had silently left the live 900 cadence, an over-permissive None branch.
• docs/AGENT_MINING.md config table regenerated.
• Note: commit 850a26d's message says "Revert TEMP(ci)" but it carries the audit test fixes; the ci.yml restore is 9b8231b (ci.yml identical to main's). Squash merge.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BMT4D7eZUKUF2gocYLVFmj

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
cascade	Ready	Preview	Nov 13, 2026 11:35pm UTC