Skip to content

Commit bf91c5d

Browse files
committed
Replay full strength domains from pinned raw journals
1 parent 33a8000 commit bf91c5d

12 files changed

Lines changed: 5456 additions & 5 deletions

File tree

‎.github/workflows/ci.yml‎

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -90,6 +90,7 @@ jobs:
9090
python -m arena.run config/arena-ci.json runs/ci-arena --workers 1
9191
python -m arena.statistics runs/ci-arena
9292
python -m tests.match_journals --campaign runs/ci-arena --output runs/ci-match-journals
93+
python -m tests.strength_bundle --campaign runs/ci-arena --output runs/ci-strength-bundle
9394
python -m tests.strength_controls --campaign runs/ci-arena --output runs/ci-strength-controls
9495
python -m tools.evidence --checkout . --engine build/caissa --network models/diagnostic-v2.cai --arena runs/ci-arena --output runs/ci-evidence
9596
python -m verification.verify runs/ci-evidence --checkout . --expected-commit "$GITHUB_SHA" --profile diagnostic
@@ -110,6 +111,8 @@ jobs:
110111
runs/ci-arena/SUMMARY.json
111112
runs/ci-match-journals/*.json
112113
runs/ci-match-journals/*.log
114+
runs/ci-strength-bundle/*.json
115+
runs/ci-strength-bundle/*.log
113116
runs/service.log
114117
runs/ci-source-binding/CONTROLS.json
115118
runs/ci-strength-v4-profile/CONTROLS.json

‎docs/MATCH_JOURNALS.md‎

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -24,7 +24,10 @@ PGN results/colors/moves, boolean identities, an invented75-move draw, an
2424
illegal PV, a missing PGN and an extra pair file. The intended outcome is one
2525
positive reconstruction and nine semantic rejections. They are run remotely
2626
while the local frozen v4 strength sequence owns the timed host. Their actual
27-
results must be inspected before claiming this increment verified.
27+
results passed on both Linux and macOS at source33a8000 in correctness run
28+
40212451591. [The complete platform logs](../results/match-journals-2026-10-25/VERIFICATION.json)
29+
also retain all26 original actor lineage controls. This verifies the small
30+
journal increment, with no full-sample or Elo claim.
2831

2932
This strengthens the raw-artifact portion of G14/G15. It supplements the
3033
fixed-sample strength verifier and does not replace any of its game counts,

‎docs/STRENGTH_BUNDLE.md‎

Lines changed: 32 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,32 @@
1+
# Full strength domain replay
2+
3+
G08 and G09 accept the `strength-journals-v1` adapter. Their `strength.campaigns`
4+
array contains one neural comparison or the three external settings, in frozen
5+
order. Each descriptor contains one exact `config` artifact reference. Its
6+
parent directory must hold CONFIG.json, PROTOCOL.json, SUMMARY.json, every
7+
paired JSON record, every separate game JSON and every PGN. All these files
8+
must be present in both the immutable manifest and that domain's artifacts.
9+
Missing, extra, linked, changed or undeclared journals reject the domain.
10+
11+
The adapter stages pinned bytes outside the evidence directory, reconciles
12+
pair/game/PGN files, and independently replays moves, claims, request budgets,
13+
timings and PVs. It then recomputes counts, clustered bootstrap intervals and
14+
relative Elo using the existing full strength verifier. G08 retains the
15+
4000-game sample, identical classical actor, at least50 Elo neural gain and
16+
strictly positive95% lower bound. G09 retains all2000 games across the three
17+
frozen settings and never produces an absolute rating from their option names.
18+
The caller's original wall and active deadline also bounds bootstrap work.
19+
20+
Saved pass receipts cannot discharge either gate. A negative neural endpoint
21+
preserves its independently observed statistics in STRENGTH.json and remains
22+
failed. The full report retains that failed endpoint and can continue to G09;
23+
it never lists the negative result as a verified gate. Even successful G08/G09
24+
replay leaves whole-project acceptance false
25+
while other domains are pending.
26+
27+
Hosted controls stage the four actual native CI games, reject seven artifact
28+
and sample mutations, invoke the actual full CLI against a rehashed false pass
29+
receipt, and interrupt bootstrap with the caller's original deadline. These
30+
small controls do not certify a full strength sample. Full-size positive replay
31+
still requires the completed frozen campaign. Local timed matches continue
32+
from their original archived owner and verifier code.
Lines changed: 30 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,30 @@
1+
{
2+
"schema": "caissa-match-journals-controls-v1",
3+
"source_commit": "33a80006763bc4ec8efd7f8f5a7e56b8db958bf4",
4+
"workflow_run": 40212451591,
5+
"jobs": {
6+
"linux": 121226407663,
7+
"macos": 121226407953
8+
},
9+
"status": "passed",
10+
"per_platform": {
11+
"real_original_lineage_controls": 26,
12+
"actual_journal_games": 4,
13+
"journal_semantic_rejections": 9
14+
},
15+
"scope": "Small real native CI arena, pair/game/PGN agreement and actual original actor lineage controls. No full strength sample or Elo result.",
16+
"logs": [
17+
{
18+
"file": "linux.log",
19+
"bytes": 503227,
20+
"sha256": "cb5ffd9f1efbc5444deac82bf9459cd7d39d70bc3d9262afeb6adaaae0a83dca"
21+
},
22+
{
23+
"file": "macos.log",
24+
"bytes": 102632,
25+
"sha256": "3063b24d87e9a8e314f451e0174021a24c57ca64d3c5e4d87aaa8dc135ccc641"
26+
}
27+
],
28+
"full_strength_sample_certified": false,
29+
"full_acceptance": false
30+
}

0 commit comments

Comments
 (0)