Skip to content

bench: incident benchmark (N=6) + evaluation harness for the paper #31

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
tishachawla-jg merged 2 commits into from
Jul 29, 2026
Merged

bench: incident benchmark (N=6) + evaluation harness for the paper #31

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
261a266
bench: incident benchmark + evaluation harness for the paper
Jul 28, 2026
3f3a87f
bench: add over-fitting resistance (benign output-mutation sweep)
Jul 29, 2026
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
98 changes: 98 additions & 0 deletions docs/benchmark-results.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,98 @@
[
{
"name": "Refund",
"crossings": 3,
"t_base_us": 3.7226000567898154,
"t_rec_us": 72.48700002674013,
"overhead_us_per_crossing": 22.92146665665011,
"overhead_pct_at_model": 0.007640488885550038,
"bytes_per_crossing": 1373.3333333333333,
"replay_live_crossings": 0,
"deterministic": true,
"detected": true,
"spec_gated": true,
"spec_benign": true,
"benign_survived": 5,
"benign_total": 5
},
{
"name": "Invoice",
"crossings": 3,
"t_base_us": 4.606600035913289,
"t_rec_us": 74.36680002138019,
"overhead_us_per_crossing": 23.25339999515563,
"overhead_pct_at_model": 0.007751133331718543,
"bytes_per_crossing": 1476.0,
"replay_live_crossings": 0,
"deterministic": true,
"detected": true,
"spec_gated": true,
"spec_benign": true,
"benign_survived": 5,
"benign_total": 5
},
{
"name": "Trade",
"crossings": 3,
"t_base_us": 4.276599967852235,
"t_rec_us": 72.05800001975149,
"overhead_us_per_crossing": 22.593800017299753,
"overhead_pct_at_model": 0.007531266672433251,
"bytes_per_crossing": 1406.3333333333333,
"replay_live_crossings": 0,
"deterministic": true,
"detected": true,
"spec_gated": true,
"spec_benign": true,
"benign_survived": 5,
"benign_total": 5
},
{
"name": "Email",
"crossings": 3,
"t_base_us": 3.52159992326051,
"t_rec_us": 72.54480000119656,
"overhead_us_per_crossing": 23.007733359312024,
"overhead_pct_at_model": 0.007669244453104008,
"bytes_per_crossing": 1367.6666666666667,
"replay_live_crossings": 0,
"deterministic": true,
"detected": true,
"spec_gated": true,
"spec_benign": true,
"benign_survived": 5,
"benign_total": 5
},
{
"name": "Payout",
"crossings": 3,
"t_base_us": 3.607400110922754,
"t_rec_us": 74.70859994646162,
"overhead_us_per_crossing": 23.700399945179623,
"overhead_pct_at_model": 0.007900133315059874,
"bytes_per_crossing": 1390.0,
"replay_live_crossings": 0,
"deterministic": true,
"detected": true,
"spec_gated": true,
"spec_benign": true,
"benign_survived": 5,
"benign_total": 5
},
{
"name": "Deletion",
"crossings": 3,
"t_base_us": 2.8966000536456704,
"t_rec_us": 71.75499992445111,
"overhead_us_per_crossing": 22.952799956935152,
"overhead_pct_at_model": 0.007650933318978384,
"bytes_per_crossing": 1274.3333333333333,
"replay_live_crossings": 0,
"deterministic": true,
"detected": true,
"spec_gated": true,
"spec_benign": true,
"benign_survived": 5,
"benign_total": 5
}
]
29 changes: 29 additions & 0 deletions docs/benchmark-table.tex
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,29 @@
% Auto-generated by examples/benchmark/harness.py
\newcommand{\numincidents}{6}
\newcommand{\detectionrate}{6/6}
\newcommand{\specificity}{6/6}
\newcommand{\benignspecificity}{30/30}
\newcommand{\recoverheadus}{23.0}
\newcommand{\recoverheadpct}{0.008}
\newcommand{\kbpercrossing}{1.44}
\newcommand{\modellatencyms}{300}

\begin{table}[t]
\centering
\small
\begin{tabular}{lrrrr}
\toprule
Scenario & Cross. & $t_{\mathrm{base}}$ & $t_{\mathrm{rec}}$ & Rec./cross. \\
& & (\textmu s) & (\textmu s) & (\textmu s) \\
\midrule
Refund & 3 & 3.7 & 72.5 & 22.9 \\
Invoice & 3 & 4.6 & 74.4 & 23.3 \\
Trade & 3 & 4.3 & 72.1 & 22.6 \\
Email & 3 & 3.5 & 72.5 & 23.0 \\
Payout & 3 & 3.6 & 74.7 & 23.7 \\
Deletion & 3 & 2.9 & 71.8 & 23.0 \\
\bottomrule
\end{tabular}
\caption{Recording adds \recoverheadus{}~\textmu s per crossing (\recoverheadpct\% of a \modellatencyms~ms model call). Replaying the suite makes zero model calls; cut-point tests flag \detectionrate{} incidents and pass on \specificity{} guarded and benign changes. The store grows by at most \kbpercrossing{}~KB per crossing.}
\label{tab:overhead}
\end{table}
14 changes: 14 additions & 0 deletions examples/benchmark/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,14 @@
"""The Chronicle incident benchmark: recorded agent incidents + an evaluation harness.

Each scenario is a small ``model -> tool -> model`` agent where an unguarded tool
produces an unsafe result and a guarded version corrects it. Every scenario has three
tool variants:

- ``ungated`` the incident (unsafe): the cut-point test must FAIL (fault detected).
- ``gated`` the fix (safe): the cut-point test must PASS.
- ``benign`` an unrelated correct change: the cut-point test must PASS (specificity).

The harness records each incident, replays it, and cut-point tests all three variants,
reporting recording overhead, store growth, replay determinism, and detection /
specificity. See ``harness.py``.
"""
Loading
Loading