GB1 verdict — alignment and range access

This page records why blob extension §28 marks GB1 closed. It is a record of evidence, not part of the specification: it adds no requirement, and the gate’s text on the security page is what it answers to.

1. The gate

GB1 alignment and range access: an SOP-BLOB-1 object imported into the transfer layer; a plaintext range served via one verified range request from a non-decrypting holder; decrypt of only the fetched chunks; tamper anywhere in a served range and any chunk tag detected.

2. Where the evidence was read

• Commit. af7e168e6e2b574e42549cca1fd6a8195d0254de, on develop. Every test named below exists at that commit, and on main.
• Run. CI run 60665, job Rust Test (sovm workspace), 2026-10-03T01:24Z, step cargo test (sovm workspace on loopback), which checked out exactly that commit. Every test named below passed in it. None of them is ignored by default, so the same tests run in that job on every merge.

3. The clauses, and the tests that carry each

Every test below except the last row’s lives in the GB1 prototype’s gb1 tests. The prototype is a member of the sovm workspace. It runs against the transfer layer’s pinned version, so a dependency bump that changes any measured behavior fails the build.

Clause	Tests
an SOP-BLOB-1 object imported into the transfer layer	an_sop_blob_1_object_imports_under_the_storage_id_the_spec_derives: the identifier the transfer layer assigns on import equals the storage identity the specification derives over the encrypted object
a plaintext range served via one verified range request from a non-decrypting holder	one_range_request_moves_exactly_the_chunk_and_nothing_around_it: a holder with no key serves one chunk’s ciphertext span through one verified request, and the bytes moved are exactly that span
decrypt of only the fetched chunks	the_reader_decrypts_the_fetched_chunk_and_holds_nothing_else: the reader recovers the requested plaintext from the fetched chunk and holds no other chunk
tamper anywhere in a served range detected	a_single_flipped_bit_anywhere_in_a_served_range_fails_verified_streaming: every single-bit flip across a served range is refused by verified streaming
any chunk tag detected	a_flipped_byte_in_the_fetched_ciphertext_fails_the_chunk_tag: ciphertext that passes the transfer layer but has been altered fails at the chunk’s authentication tag
the group size the gate’s prose says to prototype against	the_verification_group_is_the_sixteen_kib_format_section_7_assumes and verified_ranges_are_served_in_one_kib_units_not_in_verification_groups: the verification group matches the reference assumption, and the serving unit is one 1 KiB chunk
negative control on alignment	a_span_shifted_off_the_grid_over_fetches_one_serving_unit_per_edge: a span shifted off the chunk grid pays one extra serving unit per edge, so the exact-span result in the second row cannot pass vacuously
the serving unit holds at the binding’s transfer path	gi2_a_ranged_get_moves_only_the_requested_range_not_the_whole_object in sovm-iroh’s gates tests: a one-byte request moves exactly one serving unit, asserted by equality rather than by a ceiling

4. The rulings that interpreted the gate

The chunk constant is confirmed, not re-derived. Section 28 asks for a re-derivation if the group size or the hashing behavior differs from the reference assumptions. The group size matched. The hashing behavior differed, but in the finer direction: verified ranges are served in 1 KiB units, not in 16 KiB groups, and 1 KiB divides the 1 MiB chunk exactly. So the reason the constant was chosen still holds. The ruling recorded that the trigger fired and the answer did not change, rather than recording that the trigger did not fire. The same measurement corrected two passages. One is format §7’s statement of the serving granularity. The other is this gate’s fourth clause, which previously said “tamper in any 16 KiB group”. Section 7’s number was accepted only on the condition that something fails when it stops being true. The exact-equality assertion in the table’s last row is that condition.

The transported length is unauthenticated, and GB1 holds anyway. The prototype found that the object length the transfer layer carries is outside range authentication: the_bao_length_prefix_is_outside_the_range_verification accepts a share of single-bit flips in it. That does not fail any clause of GB1, because an SOP-BLOB-1 reader takes the object’s length from the signed manifest and derives every span from the chunk grid. Format §7 now states this as a requirement on the reader instead of leaving it as an unstated dependency.

The gate closes on its own text. GB1 closed by itself, while GB2 and GB3 stay open. This follows the rule SOVM/1’s other gated pages already apply: a gate closes when its own text is satisfied, not when all of its siblings are.

5. What the evidence does not show

The evidence covers the reference transfer layer only, at its pinned version. It says nothing about seek-and-play performance or decode-path overhead (GB2). It does not cover padding or storage overhead (GB3). It does not show a range read through the binding’s own fetch interface end to end: the prototype drives the transfer layer directly, and the binding’s ranged fetch is measured separately by its own gate tests.

6. Reproducibility

The workspace these tests live in is not released, so an outsider cannot re-run them today. Reproducibility is a target, owed under decision D-0002 by project P-0058, which releases the node code the tests exercise; until then this page is a record an outsider can read and not yet one they can check.