What survived
Atomic propose/commit/rollback state, checkpoint v4, resume equivalence, append-only attempt evidence, and direct functional probing.
EXP-001 · completed 2026-07-19
Local safety did not add up.
The transaction mechanics worked. The observer was byte-identical. But a 0.25% per-attempt replay guard accepted every step while cumulative public replay crossed the frozen 2% ceiling.
No-go
The frozen question
Can an observer-calibrated per-attempt functional replay budget prevent faculty training from exceeding the public replay ceiling without weakening the quantity gates?
The student, teachers, corpora, routing, seed 2, thresholds, stop rules, and promotion sealing were fixed. Seeds 1 and 3 could open only after a seed-2 pass.
Measured outcome
Quantity passed. Replay failed.
| Measure | Observed |
|---|---|
| Observer learned-state equivalence | pass |
| First-order predictive Pearson | 0.0076 |
| Guard attempts / commits / rejects | 200 / 200 / 0 |
| Warning / hard exceedances | 5 / 0 |
| Maximum local replay increase | 0.2013% |
| Quantity operation rate | 95.0% |
| Quantity exact-artifact rate | 95.0% |
| Public replay regression | 2.685% |
| Promotion evaluations | 0 |
What failed
The assumption that bounding each local increase independently would bound the complete trajectory.
What stays sealed
The promotion split, replication seeds 1 and 3, ZERO.4 promotion, and the cross-family replication bridge.
Deterministic registration
Replay the claim, including the no-go.
The ilXyr adapter binds commit 0ab274c577074eaa7d519dc3fda20d8a86acfc63, eleven source hashes, both complete transaction logs, and both public evaluations.
cargo run -p ilxyr-cli -- retro /workspace/ledger \
examples/families/zero-q23-seed2.retro.json --execute
cargo run -p ilxyr-cli -- verify /workspace/ledgerA fresh end-to-end registration produced four content-addressed objects across five hash-linked events and passed ledger verification. This is retro evidence: grounded by deterministic replay, but not assigned synthetic historical forecast risk.
Later decisions
Budget the trajectory, then test the constrained path.
EXP-002 enforced the cumulative six-slice authority and stopped at 66 commits. EXP-003 added deterministic backtracking under the same ceiling, gained five commits, then exhausted every scale before a public checkpoint.