Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions experiments/zeta_fold_exploration/MANIFEST.json
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@
},
"generated_by": "make_manifest.py",
"note": "committed_files lists every file meant to be committed (this manifest excluded). raw_files are regenerable samples written to results/raw/ and not committed; their hashes are from the exploratory run; regeneration is expected to be byte-identical only with OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 (README section 8).",
"committed_total_bytes": 558582,
"committed_total_bytes": 559876,
"committed_files": [
{
"path": "CLAIMS.md",
Expand All @@ -18,8 +18,8 @@
},
{
"path": "README.md",
"bytes": 110564,
"sha256": "c2c932769e70204fa060e827e7b5bcf1adae2940e248e966ede75ed98924943a"
"bytes": 111858,
"sha256": "2e08a7d91dcb1c7f17e593a953046f3bf3eaf15ef9fbe96284c53f945d264be2"
},
{
"path": "dbn/check_collision.py",
Expand Down
25 changes: 21 additions & 4 deletions experiments/zeta_fold_exploration/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -1249,6 +1249,23 @@ A nonauthor read from a non-Anthropic provider was requested, in four slices:
execution of the same programs and fixed inputs, not the replay's own streams, and it makes
no scientific acceptance.

**The replay's own timing-only streams, retained: PASS (scoped).**
- **Recovery.** The replay's original 14 timing-only byte streams of 20 October were later
found in its performer's workspace. They are retained, without any rerun, on draft
[main#479](https://github.com/d6g8k5htny-coder/main/pull/479) at `fe29f645`
([custody record](https://github.com/d6g8k5htny-coder/main/pull/458#issuecomment-6324928437)),
a retention-only record not meant for merge.
- **Check.** Each file matches the replay's `outputs.json` (at `78976ae7`) in size and SHA-256.
Under the published mask, each equals its committed original, with no non-timing difference.
The receipt of record is a nonauthor
[receipt from Grok Bot agent 13](https://github.com/d6g8k5htny-coder/main/pull/458#issuecomment-6325462247)
(Cursor-hosted, not Anthropic and not the custodian; organizational-independence credit 0),
which checked the git objects directly. An
[Anthropic cross-check](https://github.com/d6g8k5htny-coder/main/pull/458#issuecomment-6325388915)
(same provider as the packet author, no independence credit) agrees with it.
- **Coverage.** This covers the 13 streams that the replay's record review could not mask, and
`rs_spot_check.out`.

**Residual trails reproduced exactly.** The replay's two differing files are
`results/dbn/indep_ht_c910.out` and `indep_ht_c3435.out`. At the documented dps 60,
`c910` matches every line except the iteration-4 step, and `c3435` differs from
Expand Down Expand Up @@ -1336,9 +1353,8 @@ A nonauthor read from a non-Anthropic provider was requested, in four slices:
claims that Edwards was verified. The docstring of `stats/rs_zeta.py` still names Edwards. That
file is left byte-unchanged so that the replay and R2a records stay bound to its blob
`8b97a1f6`. The replay's two residual-trail discrepancies now have the exact reproduction
above. Its own 13 timing-only streams stay unretained (their classification now rests on the
retained re-run above), and R2a's own raw JSON is unpublished; no record closes all of
slice C.
above. Its own timing-only streams are now retained and mask-checked (above). R2a's own raw
JSON is unpublished, so no record closes all of slice C.
- **D.** A wording audit of the non-claims. **Read: AMEND** (the same review, §10.1). **Delta
readback: PASS (scoped)** (review 5487707101, with slice A).

Expand All @@ -1364,7 +1380,8 @@ The slice C performer reported Cursor/xAI Grok 4.7; the record review was a sepa
Codex lane, source-exposed and not blind, with organizational-independence credit 0. The
review checked the 59-command multiset, hashes, counts, environment, discrepancies and packet
binding at exact replay head `78976ae7`. It did not independently reapply timing masks to 13
timing-only replay byte streams that were not retained. The replay used the same programs and
timing-only replay byte streams that were not retained. Those streams were later retained and
mask-checked (item C). The replay used the same programs and
fixed inputs: it is not a second implementation or a scientific acceptance. Both residual
trails were later reproduced exactly with the earlier review-math script (item C). The original
command lines were not recorded, so that is a sufficient invocation, not a log. Scientific
Expand Down