Invest1 distinct publisher3 min readPublished
The China Coast Guard Academy reports no weapons-rule violations from its AI command system, but the baseline it beat opens fire in 0.3% of encounters, so a few dozen simulated runs was never going to show one.
The Investor · Invest desk
Compiled by The InvestorSomething wrong?How this is made
Zero is the cheapest number in this study and the one carrying the least weight. At the baseline's 0.3% per encounter [3], three dozen runs of almost anything produce no shooting about nine times in ten, because 0.997 to the 36th is 0.90 [1], and you would need something like 231 runs before the odds of seeing a single lethal event were even [2]. "Dozens of simulated exercises" [3] is therefore not a test of the 0% result; it is a sample too small to contain the event being counted.
The measurable finding sits in the middle of the write-up. The intent classifier's error rate falls from 31.5% to 10.4%, roughly two thirds off [5], and that figure is built from many classification calls per run rather than one rare trigger pull, which is why it is the number I would actually trade on. It is also less soothing than it reads: with five ships closing [2], and treating each judgment as independent (my assumption, not the authors'), the chance of at least one misread inside a single encounter improves from 85% to 42% [6], which still loses more often than a coin.
The gain in overall rule violations, four tenths of a percentage point [4], is a third off the baseline [3], and separating two proportions that close at conventional power takes on the order of 10,000 runs per arm, from 16 times 0.01 times 0.99 divided by 0.004 squared [4]. Dozens is about 275 times short of that [7].
What the exercise actually recommends is a substitution: when the wedge formed and probing became the likelier reading, the output was a water cannon instead of firearms [7]. That is escalation control in the only sense that binds anybody, a lower rung selected inside the same encounter, and it is simultaneously a licence for more encounters, since contact that stays beneath the threshold at which a foreign ministry has to respond can be run more often and at more places. Note what is not being asked for. Sun Shengzhi says replacing human command is not the goal [10], the stated next step is joint human-AI judgment across maritime, land and air domains [9], and the pitched use is red-teaming, modelling a more rational and flexible adversary to expose weaknesses in existing operational plans [8]. That is an advisory layer bolted to a command chain that already exists, and it is far cheaper to field than hulls.
This is probably wrong, but the version I would defend is that restraint is the procurement argument and the accuracy figure is the collateral: an academy paper in a domestic journal, reaching us through the South China Morning Post's account relayed by Seoul Economic Daily [1], graded against rules and a comparison system the same institution encoded, the half nautical mile standoff and the weapons approval procedure included [2]. Or rather, the more interesting version is that Sun's caveat about needing further work to align AI judgments with strategic objectives and ethical standards [10] is what an unresolved internal argument looks like in print. The read fails if the academy's next output asks for vessels that decide alone rather than a layer that advises, in which case none of this was about pricing restraint.
Ranked by verification strength, evidence, and original report placement.
Hong Kong's South China Morning Post reported on 30 August that researchers at the China Coast Guard Academy, writing in the Chinese journal Command Control and Simulation, released results of computer war games simulating gray zone confrontations in disputed waters such as the South China Sea; the account was carried by en.sedaily.com (Seoul Economic Daily).
The simulation placed a single Chinese coast guard vessel protecting fishing boats as four to six foreign ships approached to conduct low-intensity harassment and reconnaissance; ships were required to maintain a safety distance of at least 0.5 nautical miles, about 0.9 kilometres, and to follow strict approval procedures before using weapons.
Under the existing rule-based system the Chinese coast guard vessel used lethal weapons in 0.3% of cases, while the AI-based system developed by the researchers produced no violations of weapons-use rules across dozens of simulated exercises; the study frames the risk as 0.3% under existing command systems and 0% with AI.
The system correctly judged the intent behind an approaching vessel's actions 89.6% of the time, compared with 68.5% for the existing system.
Unlike the current condition-response approach, which treats a vessel as a threat once it crosses a set boundary, the new system combines radar tracks and sensor data with a causal analysis model to assign probabilities to whether an approaching ship is navigating, testing the coast guard's response, or attempting a collision or deceptive manoeuvre.
Distinct publishers with included, body-backed reporting in this cluster.
en.sedaily.com
1 article · August 30, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
product
A China Coast Guard Academy war game grades its AI on how often the gun comes out1 distinct publisher
product
Two Chinese booster recoveries, one state and one commercial: reflight is still the untested part1 distinct publisher
invest
SMIC's binding constraint is floor space, and the AI crunch has reached power chips1 distinct publisher
invest
UK salmonella outbreak tied to eateries using eggs from multiple suppliers, source still unidentified1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Third-hand, and the sample is the story
Every figure — 0.3%, 89.6%, 0.8% against 1.2% — arrives through two layers of summary: Seoul Economic Daily reading the South China Morning Post reading a paper in Command Control and Simulation. Nobody in that chain publishes the run count, the adversary model, or how the rule-based baseline was built. What can be checked, arithmetic, cuts against the headline: the reported effect sizes are far smaller than the reported sample can resolve.
Nothing fielded to count
There is no deployment to measure. What exists is a simulator result and a stated ambition to extend it to maritime, land and air scenarios and to joint human-machine command. No cutter, unit, or exercise is said to run this system, and the lead author explicitly frames replacing human command as not the goal — so treating publication as uptake would be our invention, not the reporting's.
Zero is doing a lot of work
The claim being sold is escalation eliminated — weapons use cut to nothing by better software. The baseline reaches for lethal force in three encounters out of a thousand, so a few dozen quiet runs is the expected outcome for the old system too, not a discovery about the new one. The intent-recognition jump is the one result with real headroom in it, and it is the one the framing subordinates to a percentage that could not have been anything else.
The graders own the boats
The evaluation was run by the training academy of the service being evaluated, on the exact scenario — one cutter, foreign ships crowding fishing boats — where that service's conduct is disputed by its neighbours. A finding that automation makes Chinese cutters more restrained is useful well beyond the lab, and the paper's own author pairs it with a bid for follow-on work on multi-domain and human-machine command. None of that makes the result false; it does mean no disinterested party has touched it.
Firm on the arithmetic, blind on the paper
We are confident about the part that needs no access: the reported rates and sample size are mutually incompatible with the conclusion drawn from them, and that holds whatever the paper says. Everything else is fogged — whether the simulator is a fair test, what the baseline actually was, how 'dozens' was counted, and what survived two rounds of translation and compression.