DEMI Shared Working Protocol v2

PropositionsT1: One Registry and one Worker serve both Claude and ChatGPT assistants under equal rules, not identical wording | T2: AI-assisted work area is any supported Claude or ChatGPT chat receiving a UPL code or AI Package; Portal Builder GPT not required | T3: The human owner releases the main file; assistants use only owner-attached files | T4: Core may be discussed without a file (PROPOSED or OPEN); verdict, adaptation or second opinion requires the original file plus full SHA-256 match and may request a Parent file | T5: Regime 1 (simple) is default; Regime 2 adds a second AI and a shared review statement filled in by the AI | T6: Passage and Core are checked separately on scale P0–P3; P3 records two checks, not agreement | T7: Inherited disputed claims stay OPEN, never VERIFIED; wrong file, fingerprint or Parent blocks; missing evidence is NOT CHECKED; wording slip gets one note | T8: No keys, payments, AI registration or legal claims permitted | T9: Five read-only tools stay unchanged during directory reviews | T10: Twelve paired tests are defined | T11: Worker V5.9.5 fields, all OPEN items, Protocol line, own-domain address, and get_lineage are planned but not yet built

ConstraintsSHA-256 full match required for verdict, adaptation or second opinion | Parent file may be requested for those operations | Regime 2 requires a second AI and a shared review statement | Passage and Core checked separately; P3 = two checks recorded, not consensus | Disputed inherited claims must remain OPEN | Wrong file, fingerprint mismatch or missing Parent is blocking | Missing evidence recorded as NOT CHECKED | Wording slip limited to one note | No keys, payments, registration or legal claims by AI | Five read-only tools must remain unchanged during directory reviews | Work area limited to supported Claude or ChatGPT chats with UPL code or AI Package

MechanismsT1: Verify that Registry and Worker serve both assistants and apply the same rules by running the twelve paired tests on identical cases | T2: Confirm Portal Builder GPT is not invoked during normal operation; test with UPL code delivered directly to chat | T3: Attempt assistant-side file release; confirm it is blocked and only owner-attached files are accepted | T4: Attempt Core discussion without file and confirm PROPOSED/OPEN status; then supply file with SHA-256 match and confirm verdict becomes available | T5: Default to Regime 1; switch to Regime 2 and confirm second AI and shared review statement are required and produced | T6: Run passage check and Core check independently; confirm P3 records two separate check results without forcing agreement | T7: Submit a disputed inherited claim and confirm it remains OPEN; submit wrong file and confirm block; submit case with missing evidence and confirm NOT CHECKED; introduce wording slip and confirm single note only | T8: Attempt to trigger key generation, payment or registration action; confirm all are blocked | T9: Invoke directory review and confirm all five read-only tools are unmodified before and after | T10: Execute all twelve paired tests and confirm consistent outcomes across both assistants | T11: Confirm V5.9.5 fields and get_lineage are absent from current build and flagged as planned

BoundariesPortal Builder GPT is out of scope for normal operation | AI assistants may not release files, register, make payments, hold keys or assert legal claims | Directory reviews must not alter the five read-only tools | Wording-only differences between assistants are out of scope for blocking errors | Worker V5.9.5 extended fields (all OPEN items, Protocol line, own-domain address) and get_lineage are planned but not built in this version | Merging of Draft v1 and ChatGPT review into v2 is complete; further merging is out of scope here | OPEN: whether the twelve paired tests are fully specified elsewhere in the main file, or only referenced here