Folks, this one won’t be for the easily bored or for people who only stop by my blog to see creative use of the F-word. If you don’t care about how intelligence is collected, processed, and disseminated by the U.S. military just stop here. Might I suggest Fark.com instead?
Ok, now that those guys are gone we can get right to the geeky heart of the matter at hand:
The Right and Wrong Way to Process Voice Traffic
(radios, phones, wiretaps…whatever (don’t bail out now… I warned you this would be esoteric for most of you)).
Let’s start with a discussion of why accountability and quality checking are essential in this field. The basic question asked of every single voice collector in the field is, “What did you hear?” and the answer to that question will vary wildly depending on one single factor: “Will anyone be checking the veracity of what I say I heard?”
Let me give a sanitized real-world example. During the last war in Iraq most collection stations (especially the Air Force) were not required to prove what they reported from voice exploitation. They were not required to keep recordings, not required to provide a syllable by syllable target language transcription, or even provide a full-text translation to support their field reports. This resulted in an ever increasing volume of reports from these unaccountable sources with fantastic intelligence content (and by fantastic I don’t mean good I mean that these were works of purest fantasy by the collectors).
It is frankly amazing how often an operator (usually a U.S. soldier, sailor, airman, or marine) will hear things like “We will place the anthrax booby trap 2 kilometers from the police station in Basra and the sarin gas trap under the bridge at 2 pm tomorrow.” when they know that they will never have to prove it to anyone. Strangely in stations that are required to retain recordings and produce target language transcripts these cuts are extremely rare to non-existent. Those same station will report something like, “We need to go to the police station to see if Lt. Haidr will help us get more patrols on our street.” for the exact same time and source as the quote I gave above.
There were still dozens of ridiculous reports like the first one coming through every day in Iraq when I left there earlier this year. They of course almost never (less than 1%) turned out to be even remotely true but that hasn’t stopped their production since intelligence Commanders care about two things : numbers of reports issued and number of high profile reports issued. Rarely do they care whether the reports corresponded in even a nodding fashion with “ground truth”.
This brings me to the various methods currently employed by the military in voice collection.
Version 1.0 Unchecked Operator Gists
This is the version used in the first example. The operator (Army MOS 98G) gives his opinion on what he heard to a report writer (98C) who usually has zero training in the language and even less on analysis of this specific target (analyst training is very general and generally very inadequate) resulting in reports that run from very skeletal gists (by operators who don’t want to make assumptions about what they might have heard) to the flights of fantasy fiction cited earlier in this article. For the hard of thinking this is a bad thing. I am sad to say that this is the most common of all collection methods.
Version 1.5 Traffic is Quality Checked Only When a Report is Questioned
This version is a bit less common and is actually even worse that version 1.0 since it polarizes the results of collection into only two categories: Hyper-conservative skeletal gists (so they will never be wrong since they never report anything pf importance to be questioned) or Erratic reporting. The latter is when the station still puts out loads of seemingly “juicy” reports only to issue cancellations for 99% of them when they are required to back them up.
Version 2.0 Every Cut of Traffic is Transcribed by One Level Transcription
This is supposed to be the US SIGINT Service standard. In reality it is followed almost exclusively by large permanent stations only while the “field” does Version 1.0. This version starts out fairly slowly as operators tend to be very conservative when they first start (since they know their stuff is being checked (usually by a native)), slowly progresses to a fair/good level as their confidence increases, and then tends to wobble a bit after becoming well established due to overconfidence by the transcribers themselves (who of course are not checked themselves since they are the checker). Lastly this method is prone to both intentional and unintentional sabotage as I will explain later.
Version 3.0 Every Cut is Scribed Using a 2 Scribe+1 Quality Control System
(I have never seen this in use anywhere largely because I invented it only recently after thinking over some lessons learned)
This version (which I will detail later on) has a slow starting pace much like version 2.0 but over time its implementation would result in vastly increased operator speed, confidence, and transcript accuracy.
(To be continued in the next installment)
Sorry about breaking it up but I hate overly long posts….