rea/experiments/atomic-codemode/large-string-oracle.md

Frozen large-string capacity task

The fixture is a synthetic x86-64 ELF generated by create-capacity-fixture.py NEW_DIRECTORY with installed GNU binutils. The target is never executed and has no executable filesystem permission. This tests real Kuna/MCP capacity and composition; it does not establish application-analysis quality, broad complex-software superiority or provider parity.

Expected ELF SHA-256: bc9a2090e71f780771b388da86a57cdeef428ac235ba6b67051fd05cca688b22. Both payload files are independently available in the new fixture directory. Read ELF program headers to corroborate their complete PT_LOAD-backed extents; the model must use only its allowed static RE tools, not those fixture files.

Object marker	Start	Encoded bytes including NUL	Inclusive end	First 16 bytes	Final 8 bytes
REA_CAPACITY_PLAIN:	0x402000	2,101,249	0x603000	5245415f43415041434954595f504c41	4141414141414100
REA_CAPACITY_ESCAPED:	0x603001	1,052,673	0x704001	5245415f43415041434954595f455343	5c5c5c5c5c5c5c00

The provider reports ASCII and nul_terminated: true for both. Query counts are one each with exhausted pages in this retained inventory; artifact-wide discovery remains unknown. Both queries share one native inventory Evidence in one target lease. That handle must actually be readable through a bounded Evidence window. Its exact ID is lease/source-specific; it is not the ELF digest. End addresses are derived from exact coordinate plus byte length minus one. NUL/byte observations do not prove original C object boundaries, runtime use, references or live state.

Grade the complete answers against all of these facts, actual calls, cited IDs, artifact/engine identity and raw capture hashes. The Code Mode lane must actually use the isolated rea-codemode MCP execution entrypoint. Direct has the same 19 fact methods through 22 tools; Code Mode has four MCP tools. Both have the common client JavaScript executor. Count all responses, schemas/help, failures, retries, cached/uncached input, output/reasoning, wall, local CPU and kernel family peak. Record unavailable costs and every failed run; do not rerun to choose a winner.

The historical interface returned oversized Direct text and refused the Code Mode bridge without an inner context/Evidence link; it had no string field selection. Those failures are retained separately. They do not establish an equivalent successful old Code Mode task for timing comparisons.