# Evidence runs -- captured 2026-10-25 (sphinx, Archive lane)

## 1. live run: bash CHECK.sh  -> exit 0

== long-context multimodal dossier gate ==
root: /root/.hermes/wizard-outputs/sphinx/archive-lcmm-metadata-2026-10-25

verify_lcmm_dossier.py -- long-context multimodal dossier (sphinx) v2
root: /root/.hermes/wizard-outputs/sphinx/archive-lcmm-metadata-2026-10-25

A. manifest integrity (re-measured from disk + recomputed from dossier)
  ok   sha256 dossier.json == manifest :: d77990948e812083
  ok   bytes dossier.json == manifest
  ok   sha256 source/atom_batch.xml == manifest :: d3150650da24e2ec
  ok   dossier.source_atom_sha256 == manifest
  ok   manifest per-paper hashes/fields recomputable from dossier.json :: 0 problem(s)

B. live arXiv API re-fetch (this is the evidence, not a replay)
  mode: LIVE -- compared against the arXiv API fetched at check time
  ok   every id resolved :: 11/11

C. dossier vs LIVE truth
  ok   all papers match on title/version/published/authors/abstract/categories :: 0 mismatch(es)

D. scope + report-rendering checks
  ok   paper_count >= 10 :: 11
  ok   every abstract non-empty
  ok   every primary_category allowed :: ['cs.CL', 'cs.CV', 'cs.IR']
  ok   every paper has a lane note
  ok   REPORT paper table == dossier.json (id/version/cat/authors/date/title) :: 0 drift(s)

E. controls (negative ones MUST go red, the positive one MUST stay green)
  ok   P0 an unmodified copy stays green :: 0 failure(s)
  ok   F1 a flipped paper title is caught by field compare :: 1 failure(s)
  ok   F2 a dropped paper is caught by field compare :: 1 failure(s)
  ok   F3 an edited abstract is caught by field compare :: 2 failure(s)
  ok   F4 a bumped version string is caught by field compare :: 1 failure(s)
  ok   F5 a lying paper_count is caught by manifest recompute :: 1 failure(s)
  ok   F6 an altered manifest hash is caught by manifest recompute :: 1 failure(s)
  ok   F7 an altered report row is caught by report table check :: 1 failure(s)
  ok   all 8 controls ran :: 8 control(s)

mode: LIVE   checks run: 21   failures: 0
VERDICT: dossier holds
verify: PASS (mode LIVE)

report table: in sync with dossier.json (11 rows)
report table reproducible from dossier.json: ok

report cross-references build + verify + render: ok

CHECK: GREEN -- dossier holds

## 2. fallback branch: LCMM_ARXIV_API=http://127.0.0.1:9/api/query bash CHECK.sh -> exit 1 (CACHED is red by design)

== long-context multimodal dossier gate ==
root: /root/.hermes/wizard-outputs/sphinx/archive-lcmm-metadata-2026-10-25

verify_lcmm_dossier.py -- long-context multimodal dossier (sphinx) v2
root: /root/.hermes/wizard-outputs/sphinx/archive-lcmm-metadata-2026-10-25

A. manifest integrity (re-measured from disk + recomputed from dossier)
  ok   sha256 dossier.json == manifest :: d77990948e812083
  ok   bytes dossier.json == manifest
  ok   sha256 source/atom_batch.xml == manifest :: d3150650da24e2ec
  ok   dossier.source_atom_sha256 == manifest
  ok   manifest per-paper hashes/fields recomputable from dossier.json :: 0 problem(s)

B. live arXiv API re-fetch (this is the evidence, not a replay)
  (live fetch unavailable after 4 attempts/host: RuntimeError("URLError(ConnectionRefusedError(111, 'Connection refused'))"))
  mode: CACHED -- compared against archived source/atom_batch.xml
  ok   every id resolved :: 11/11

C. dossier vs CACHED truth
  ok   all papers match on title/version/published/authors/abstract/categories :: 0 mismatch(es)

D. scope + report-rendering checks
  ok   paper_count >= 10 :: 11
  ok   every abstract non-empty
  ok   every primary_category allowed :: ['cs.CL', 'cs.CV', 'cs.IR']
  ok   every paper has a lane note
  ok   REPORT paper table == dossier.json (id/version/cat/authors/date/title) :: 0 drift(s)

E. controls (negative ones MUST go red, the positive one MUST stay green)
  ok   P0 an unmodified copy stays green :: 0 failure(s)
  ok   F1 a flipped paper title is caught by field compare :: 1 failure(s)
  ok   F2 a dropped paper is caught by field compare :: 1 failure(s)
  ok   F3 an edited abstract is caught by field compare :: 2 failure(s)
  ok   F4 a bumped version string is caught by field compare :: 1 failure(s)
  ok   F5 a lying paper_count is caught by manifest recompute :: 1 failure(s)
  ok   F6 an altered manifest hash is caught by manifest recompute :: 1 failure(s)
  ok   F7 an altered report row is caught by report table check :: 1 failure(s)
  ok   all 8 controls ran :: 8 control(s)

mode: CACHED   checks run: 21   failures: 0
VERDICT: dossier holds
verify: FAIL -- no live evidence (mode was not LIVE)

report table: in sync with dossier.json (11 rows)
report table reproducible from dossier.json: ok

report cross-references build + verify + render: ok

CHECK: RED -- dossier does NOT hold

## 3. report-drift probe: copy of the tree with one table cell changed to 2605.22907v1 -> exit 1

== long-context multimodal dossier gate ==
root: /root/.hermes/profiles/sphinx/cache/scratch/drift-probe

verify_lcmm_dossier.py -- long-context multimodal dossier (sphinx) v2
root: /root/.hermes/profiles/sphinx/cache/scratch/drift-probe

A. manifest integrity (re-measured from disk + recomputed from dossier)
  ok   sha256 dossier.json == manifest :: d77990948e812083
  ok   bytes dossier.json == manifest
  ok   sha256 source/atom_batch.xml == manifest :: d3150650da24e2ec
  ok   dossier.source_atom_sha256 == manifest
  ok   manifest per-paper hashes/fields recomputable from dossier.json :: 0 problem(s)

B. live arXiv API re-fetch (this is the evidence, not a replay)
  mode: LIVE -- compared against the arXiv API fetched at check time
  ok   every id resolved :: 11/11

C. dossier vs LIVE truth
  ok   all papers match on title/version/published/authors/abstract/categories :: 0 mismatch(es)

D. scope + report-rendering checks
  ok   paper_count >= 10 :: 11
  ok   every abstract non-empty
  ok   every primary_category allowed :: ['cs.CL', 'cs.CV', 'cs.IR']
  ok   every paper has a lane note
  FAIL REPORT paper table == dossier.json (id/version/cat/authors/date/title) :: 1 drift(s)
       2605.22907 report version '2605.22907v1' != dossier '2605.22907v2'

E. controls (negative ones MUST go red, the positive one MUST stay green)
  ok   P0 an unmodified copy stays green :: 0 failure(s)
  ok   F1 a flipped paper title is caught by field compare :: 1 failure(s)
  ok   F2 a dropped paper is caught by field compare :: 1 failure(s)
  ok   F3 an edited abstract is caught by field compare :: 2 failure(s)
  ok   F4 a bumped version string is caught by field compare :: 1 failure(s)
  ok   F5 a lying paper_count is caught by manifest recompute :: 1 failure(s)
  ok   F6 an altered manifest hash is caught by manifest recompute :: 1 failure(s)
  ok   F7 an altered report row is caught by report table check :: 2 failure(s)
  ok   all 8 controls ran :: 8 control(s)

mode: LIVE   checks run: 21   failures: 1
VERDICT: dossier does NOT hold
verify: FAIL (rc=1)

report table: STALE at line 39
  on disk: | 2605.22907 | 2605.22907v1 | cs.CV | 6 | 2026-05-21 | VideoOdyssey: A Benchmark for Ultra-Long-Context and Omni-Modal Video Understanding |
  dossier: | 2605.22907 | 2605.22907v2 | cs.CV | 6 | 2026-05-21 | VideoOdyssey: A Benchmark for Ultra-Long-Context and Omni-Modal Video Understanding |
report table reproducible from dossier.json: FAIL

report cross-references build + verify + render: ok

CHECK: RED -- dossier does NOT hold
