Tombstone — Why Plate Meld Died

The Book of Grok's Heart

Tombstone — plate meld

Archived design. Do not build new features on this path.

What v1 built

field-plate-meld.py motto: flock + fsync + chain-hash + triple mirror.

Thirty-ish plate sources (iron_plate, gatekeeper, combinatorics, hostess7_brain, spatial, firmware, …) refreshed into one generation-linked JSON. Diagnostic mode gated which sources could refresh. Compiler sense read meld to pick profiles. Compatibility layers and launch seals trusted meld first.

What was right

• Monotonic generation
• Durability under power loss (fsync + bak)
• Fault isolation impulse (stop refreshing fault-adjacent work)
• Single place for greppable truth

What was wrong

1. Fork storms — subprocess per plate on full meld.
2. False unity — “single truth” that re-fused stale peers into multi-MB panels.
3. Security latency — keylog/capture policy should not wait on hostess7 + spatial + firmware plates.
4. Speed — post-meld hot ratio ~0.986 was a receipt that meld is almost free *after* work; the refresh path was not free.
5. Depth culture — every new idea became another plate source.

What replaces it

Chapter 6 sealed generation: one posture artifact, one bak, generation + mask + profile_id.

Diagnostic mode becomes mask collapse, not _refresh_if_allowed plate lists.

Historical files (look-only)

• lib/field-plate-meld.py (~1.7k LOC)
• lib/field-plate-meld-orchestrator.py
• data/field-plate-meld-doctrine.json
• plate-meld-redundant/
• FR v1 Chapter 8

Research conclusion

Meld taught us generation and durability. It must not remain the nervous system of the field stack.

Next: Chapter 9 — static layers and launch seals.