Accountability and escalation: what the example found
How to read the tamper-detection table: each row is a log design tested against the same 200-entry log and the same five attacks; "yes" means the design's own verify() caught the tampering.
Log
Modify
Delete middle
Reorder
Truncate tail
Rewrite all
Plain
no
no
no
no
no
Hash chain
yes
yes
yes
no
no
Hash chain + anchor
yes
yes
yes
yes
yes
HMAC chain
yes
yes
yes
no
yes
HMAC chain + anchor
yes
yes
yes
yes
yes
What to notice: truncating the tail is the one column where only the two anchored designs say yes. An attacker who controls the storage can always recompute hashes forward, so nothing stored inside the log itself can catch a chain that simply stops early.
How to read the severity table: each row is a different combination of impact, reach and reversibility run through the same formula.
Impact, reach, reversibility
Grade
Minor, 1 record, reversible
SEV4
Minor, 50 records, reversible
SEV3
Serious, 5 records, reversible
SEV3
Serious, 500 records, irreversible
SEV2
Critical, 5 records, irreversible
SEV1
Critical, 5,000 records, irreversible
SEV1
What to notice: serious impact on 5 records and minor impact on 50 records both land on SEV3. The reach band (under 10, 10 to 999, 1,000 and over) can outweigh a one-point difference in raw impact once both cases sit in the same band.
How to read the time-to-acknowledgement table: p is the probability each on-call person is reachable for the whole incident; p95 is the minutes below which 95 percent of incidents were acknowledged; "never" is the share that hit the 480-minute give-up point.
p available
Escalate: p95 / never
Re-page same person: p95 / never
0.9
23 min / 0.1%
480 min / 9.8%
0.7
39 min / 2.5%
480 min / 30.0%
0.5
480 min / 12.3%
480 min / 50.0%
What to notice: escalating keeps the p95 under 40 minutes at 90 and 70 percent availability. But once availability drops to 50 percent, even escalation's p95 hits the 480-minute cap. Escalation helps most at realistic availability. It helps far less once availability is already bad.
How to read the circuit-breaker table: limit is how many suspicious events in 60 seconds open the breaker; harm is the mean units of damage before it opens during a simulated attack; false trips counts how often ordinary traffic (5 percent flagged) opened it over 8 hours.
Limit
Harm before halt (mean of 20 runs)
False trips per 8 h
2
1.0
27.9
3
2.3
3.1
5
5.5
0.0
10
56.6
0.0
No breaker
68.5
0.0
What to notice: limit 3 cuts harm by 97 percent (68.5 to 2.3) for about 3 false trips a day, while limit 10 barely helps (56.6). The useful threshold sits where the harm curve has already dropped sharply but false trips have not yet climbed, which here is 3 to 5, not 10.
Verifying a 100,000-entry chain end to end took about 0.5 seconds in this example, which is cheap enough to run on every shift change rather than only when something looks wrong.
Tamper detection matrix, escalation versus re-paging, and the breaker threshold trade-off.