Jerry999's picture
Add files using upload-large-folder tool
22decf5 verified
Raw History Blame Contribute Delete
6.75 kB
"""Phrase-level probes: quantify the habits that only show up in prose.
Counting tool calls says how the agent worked. These probes read what it
*said* -- in its final report and in the short narration between tool calls --
to test claims that would otherwise rest on a handful of cherry-picked
examples: did it measure a baseline before changing anything, did it keep a
development set separate from the benchmark, did it mention sampling noise,
did it admit a negative result.
Each probe is a conservative regex over one text surface. They are evidence of
a habit being *present*, never proof it is absent -- an agent can hold out a
dev set without using the word.
"""
from __future__ import annotations
import re
from collections import Counter
from .parse import Run
# Probes over the final report the agent signs off with.
REPORT_PROBES: dict[str, re.Pattern[str]] = {
"reports_baseline_number": re.compile(r"baseline[^\n.]{0,60}?\d+(?:\.\d+)?\s*%|\bbaseline\b[^\n]{0,40}\b0\.\d+"),
"reports_version_table": re.compile(r"\|\s*(?:v\d|model|run|stage)\b.*\|", re.I),
"names_dev_set": re.compile(r"\bdev\b|held[- ]out|validation set|\bval\b set|never (?:saw|used)", re.I),
"quantifies_uncertainty": re.compile(r"stderr|standard error|\bn\s*=\s*\d+|\b±\b|noise|within (?:the )?margin|significan", re.I),
"admits_negative_result": re.compile(
r"did not (?:beat|improve|work|help)|no (?:measurable |significant )?(?:improvement|gain)|regress|"
r"worse than|failed to|honest (?:outcome|limitation)|I did not",
re.I,
),
"states_limitation": re.compile(r"limitation|caveat|bottleneck|ceiling|cannot|the honest", re.I),
"compliance_statement": re.compile(r"contamination|decontaminat|no .{0,20}test data|compliance|untouched|unmodified", re.I),
"explains_mechanism": re.compile(r"because|the reason|which is why|driven by|root cause|diagnos", re.I),
"mentions_time_budget": re.compile(r"time (?:budget|left|ran out)|remaining (?:time|hours)|with .{0,15}hours? left", re.I),
"describes_scorer_behaviour": re.compile(
r"scorer|extraction|the grader|first (?:```|code block)|parser|how (?:it|the eval) (?:scores|extracts)", re.I
),
"reports_ablation": re.compile(r"ablat|with(?:out)? (?:the )?\w+ (?:it|this) (?:dropped|rose|fell)|A/B|compared .{0,30}variant", re.I),
}
# Probes over the short narration the agent emits between tool calls. For
# Claude this is the only natural-language rationale in the whole trace --
# its 5,596 extended-thinking blocks are present but encrypted, zero readable
# characters. Codex exposes its reasoning in full, so for Codex this surface
# is the tip rather than the whole.
NARRATION_PROBES: dict[str, re.Pattern[str]] = {
"abandons_a_branch": re.compile(r"not worth|abandon|drop (?:this|that|the)|give up on|revert to|roll back|scrap", re.I),
"reasons_about_time": re.compile(r"hours? (?:left|remaining)|time budget|won'?t finish|too slow to|enough time", re.I),
"reasons_about_noise": re.compile(r"noise|stderr|too few samples|not significant|could be chance|small n", re.I),
"plans_fallback": re.compile(r"fall ?back|if (?:this|that) fails|safety|keep (?:the|a) (?:best|working)|insurance", re.I),
"questions_own_measurement": re.compile(r"might be measuring|is that real|double[- ]check|verify (?:that|the) (?:number|score)|sanity", re.I),
}
def narration_text(run: Run) -> str:
"""Every assistant text block: what the agent chose to say out loud."""
out = []
for ev in run.events:
if ev.kind != "assistant":
continue
for c in ev.raw.get("message", {}).get("content", []) or []:
if isinstance(c, dict) and c.get("type") == "text":
out.append(c.get("text", "") or "")
return "\n".join(out)
def reasoning_text(run: Run) -> str:
"""Readable extended-thinking text, where the SDK exposes any.
Empty for every Claude trace -- those blocks are present but encrypted.
"""
out = []
for ev in run.events:
if ev.kind != "assistant":
continue
for c in ev.raw.get("message", {}).get("content", []) or []:
if isinstance(c, dict) and c.get("type") == "thinking":
out.append(c.get("thinking", "") or "")
return "\n".join(out)
def probe_run(run: Run) -> dict[str, bool]:
"""Each habit on its natural surface, plus an `anywhere_` variant.
The two scaffolds put the same content in different places, so one
surface cannot serve every comparison:
* plain the final report only. Comparable within an agent. Claude
signs off in 2.4k characters, Codex in 340, so this
understates Codex badly.
* `said_` report + between-turn narration -- everything the agent
chose to say out loud. **Both SDKs expose this, so this is
the surface to compare across agents.**
* `anywhere_` the above plus extended-thinking text. Only Codex has any,
so a cross-agent gap here measures what the SDK exposes,
not what the agent did.
"""
report = run.final_message or ""
narration = narration_text(run)
said = report + "\n" + narration
everything = said + "\n" + reasoning_text(run)
res = {k: bool(p.search(report)) for k, p in REPORT_PROBES.items()}
res.update({f"said_{k}": bool(p.search(said)) for k, p in REPORT_PROBES.items()})
res.update({f"anywhere_{k}": bool(p.search(everything)) for k, p in REPORT_PROBES.items()})
res.update({k: bool(p.search(narration)) for k, p in NARRATION_PROBES.items()})
return res
def probe_cohort(runs: list[Run]) -> dict[str, dict[str, object]]:
hits: Counter[str] = Counter()
n_report = sum(1 for r in runs if r.final_message)
for r in runs:
for k, v in probe_run(r).items():
if v:
hits[k] += 1
rows = {}
variants = (
list(REPORT_PROBES)
+ [f"said_{k}" for k in REPORT_PROBES]
+ [f"anywhere_{k}" for k in REPORT_PROBES]
+ list(NARRATION_PROBES)
)
for k in variants:
denom = n_report if k in REPORT_PROBES else len(runs)
rows[k] = {
"runs_with_evidence": hits[k],
"of": denom,
"pct": round(100 * hits[k] / denom, 1) if denom else None,
"surface": (
"final report"
if k in REPORT_PROBES
else "report + narration (cross-agent comparable)"
if k.startswith("said_")
else "incl. extended thinking (Codex only)"
if k.startswith("anywhere_")
else "narration"
),
}
return rows