Intent: the maintainer, 2026-11-03: "let's get the docs locked in", for the clients waiting on a clear specification. The reading battery's first run (ba9105b3cc), gemma4:31b on three checks twice, got 44 of 48 right, and each miss was the text's: in section 2.1 it read "with no info string or json" as refusing a fence opened with json, which src/ocx/messages.cr accepts and OCX itself writes; and in 2.12, an exit of 137 as surely signal 9, though the code writes 128 plus N for a signal as a shell does, so the two are written alike. A reading of the rest found more a client would trip on: 2.3 said a run with no sandbox runs unconfined while refusal 5 refuses it on a host without ROOT/settings/unconfined; 2.3 bounded runs from R by what the clone's owner allows, a file decision 132 removed; 2.8 wrote $release$ and $releases$ with one dollar each; and 2.11's filter 1 was refusal 1's row copied, which no code applies. Filter 5 also held back posts from any clone's user, which src/ocx/acting.cr's FILTERS never did, and the maintainer, 2026-11-03, of that constraint: "the constraint never made sense".

Expression: docs/protocol-v2.md: 2.1, an opening of three backticks alone or then json, spaces around json, and any other opening holding no message; 2.3, a run from R with the isolation refusal 1 asks, and an absent sandbox refused unless the host allows it (refusal 5); 2.8, $$release$$ and $$releases$$; 2.11, the filters as the code holds them: known, after, no message, not its own (the process's user alone), where, heard; 2.12, exited for a process ended by itself or a signal with no stop or limit, and an exit above 128 saying neither. 399 lines. battery/readings: section names may be the sections' anchors, as models cite them; clients.json adds a fourth check, on docs/protocol-v2-tools.md: messages written by other means, receipts, post's refusal, tolerant answers, a misplaced stop, ok and stopped, dry_run, and create from a checkout starting no thread.

Demonstrates: crystal run battery/main.cr -- --read all, gemma4:31b and qwen2.5-coder:7b, two trials each, on this text: gemma4:31b right on every question of every check, 64 of 64 (envelope, acting, runs, clients); qwen2.5-coder:7b 27 of 64 right, 23 partial, 14 wrong, its reasons often quoting the rule its numbered choice then contradicts, as the proposal's section 5 found of it. spec/battery/reading_spec.cr, spec/guards/documents_spec.cr, spec/guards/delineation_spec.cr, spec/guards/size_spec.cr, spec/guards/protocol_names_spec.cr and spec/guards/legibility_spec.cr: 18 examples, 0 failures.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KLNdgQ5Qp4fkwUDDWT

Intent: the maintainer, 2026-11-03: "let's get the docs locked in", for the clients waiting on a clear specification. The reading battery's first run (ba9105b3cc), gemma4:31b on three checks twice, got 44 of 48 right, and each miss was the text's: in section 2.1 it read "with no info string or json" as refusing a fence opened with json, which src/ocx/messages.cr accepts and OCX itself writes; and in 2.12, an exit of 137 as surely signal 9, though the code writes 128 plus N for a signal as a shell does, so the two are written alike. A reading of the rest found more a client would trip on: 2.3 said a run with no sandbox runs unconfined while refusal 5 refuses it on a host without ROOT/settings/unconfined; 2.3 bounded runs from R by what the clone's owner allows, a file decision 132 removed; 2.8 wrote $release$ and $releases$ with one dollar each; and 2.11's filter 1 was refusal 1's row copied, which no code applies. Filter 5 also held back posts from any clone's user, which src/ocx/acting.cr's FILTERS never did, and the maintainer, 2026-11-03, of that constraint: "the constraint never made sense".

Expression: docs/protocol-v2.md: 2.1, an opening of three backticks alone or then json, spaces around json, and any other opening holding no message; 2.3, a run from R with the isolation refusal 1 asks, and an absent sandbox refused unless the host allows it (refusal 5); 2.8, $$release$$ and $$releases$$; 2.11, the filters as the code holds them: known, after, no message, not its own (the process's user alone), where, heard; 2.12, exited for a process ended by itself or a signal with no stop or limit, and an exit above 128 saying neither. 399 lines. battery/readings: section names may be the sections' anchors, as models cite them; clients.json adds a fourth check, on docs/protocol-v2-tools.md: messages written by other means, receipts, post's refusal, tolerant answers, a misplaced stop, ok and stopped, dry_run, and create from a checkout starting no thread.

Demonstrates: crystal run battery/main.cr -- --read all, gemma4:31b and qwen2.5-coder:7b, two trials each, on this text: gemma4:31b right on every question of every check, 64 of 64 (envelope, acting, runs, clients); qwen2.5-coder:7b 37 of 64 right, 23 partial or wrong, its reasons often quoting the rule its numbered choice then contradicts, as the proposal's section 5 found of it. spec/battery/reading_spec.cr, spec/guards/documents_spec.cr, spec/guards/delineation_spec.cr, spec/guards/size_spec.cr, spec/guards/protocol_names_spec.cr and spec/guards/legibility_spec.cr: 18 examples, 0 failures.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KLNdgQ5Qp4fkwUDDWT