A summary keeps the gist and drops the particulars. Names, numbers and exact sequences are the first casualties, while conclusions usually survive.
Survives a summary | Often lost in a summary |
|---|---|
Decisions and their reasons | Exact addresses and bit fields. |
Open questions and owners | Ordering of similar events. |
Overall verdicts | Rare corner cases and caveats. |
A 6000-token debug history compresses to a 400-token summary. The ratio is large and the kept fraction is small, which is exactly why follow-up questions suffer.
original 6000 tokens, summary 400 tokens
ratio = 6000 / 400 = 15, so 15:1
kept = 400 / 6000 = 0.0666... = about 6.7%
dropped = 5600 / 6000 = 0.9333... = about 93.3%
check: 6.7 + 93.3 = 100.0The program below stands in for a model with a fake scripted function. It keeps sentences holding a digit and drops the rest, then checks what survived.
import sys
SENTENCES = [
"The regulator output is 1.8 V with a 50 mV ripple limit.",
"The enable pin needs a 10 ms delay after power-up.",
"The lab bench is near the window and gets warm afternoons.",
"The retry limit is 3 attempts with a 5 ms backoff.",
"Coffee tastes better after a passing regression.",
]
def fake_summarise(sents):
# FAKE scripted stand-in for a model: keeps sentences holding a digit.
return [s for s in sents if any(ch.isdigit() for ch in s)]
def words(sents):
return sum(len(s.split()) for s in sents)
def main():
kept = fake_summarise(SENTENCES)
dropped = [s for s in SENTENCES if s not in kept]
bad = 0
if not (len(kept) < len(SENTENCES)):
bad += 1
if not (words(kept) < words(SENTENCES)):
bad += 1
if not any("1.8" in s for s in kept):
bad += 1
if any("Coffee" in s for s in kept):
bad += 1
if not any("Coffee" in s for s in dropped):
bad += 1
if bad == 0:
print(f"PASS: kept {len(kept)} of {len(SENTENCES)} sentences and {words(kept)} of {words(SENTENCES)} words")
else:
print(f"FAIL: {bad} checks wrong")
sys.exit(1)
main()Three of five sentences survive because they hold digits. The lab note and the coffee line are gone, so questions about them are unanswerable from the summary.
About 6.7% is kept. Detailed follow-ups likely need the original, since 93.3% was dropped.
The fake rule keeps sentences holding a digit. The coffee line holds none, so it is dropped with its detail.