Blog
Signal from the disconnected edge.
Notes on offline AI, multilingual intelligence, and defense acquisition, written by Azcuna Dynamics for the people who buy, build, and operate language systems where the cloud is not an option.
Three rules. Every item is summarized from, and linked to, its public source. We cover model releases only when the weights are open and the provenance is disclosed. We cover acquisition only when there is a solicitation number. Opinion is labeled as opinion.
Start here
Evaluating language models
LLM abstention: how to make a model say "I don't know" and measure it
A system that answers from doctrine or manuals must refuse when they are silent, and a made-up answer reads like a sourced one.
5 min readRead →
Markings and data rights
DoD distribution statements A through F: levels, distro F, and DoDI 5230.24
The letter on a marked technical document sets who may receive it without further approval, so a misread letter can put it in front of the wrong audience.
13 min readRead →
Disconnected operations
How to verify that software makes no network calls
Software for a disconnected network can still call out for updates or telemetry, and each way of testing that claim covers only part of the question.
4 min readRead →
Deep dives
The engineering behind the claims, written for evaluators. Each is a long read with the questions to ask any vendor, including us.
Explainers
Short, sourced answers to the questions that come up in every evaluation. Each one lists the public documents it was written from.
Technical manuals
Army MWO: modification work orders under AR 750-10
An MWO is the only publication that authorizes a change to Army equipment, so no unit may modify equipment until one is published.
4 min readRead →
Technical manuals
ATA chapter numbers: the full list of ATA numbers in aviation
The numbering reaches regulation: FAA airworthiness directives name their subject by ATA code, such as ATA Code 05, Time Limits/Maintenance Checks.
5 min readRead →
Technical manuals
DA Form 2028: how to fill it out and where to send it
A DA Form 2028 is how an error found on the flight line reaches the office that owns a manual, and ATP 3-04.7 makes aviation quality control inspectors track every one their unit submits.
5 min readRead →
Technical manuals
How Army doctrine is numbered: the three types of publications
The prefix and the number tell a reader how much weight a publication carries: ADP 1-01 gives ADPs precedence over FMs, and FMs over ATPs, when publications conflict.
13 min readRead →
Technical manuals
How to read an Army technical manual number
The last digits of a TM number tell a repairer which level a manual serves, and the CH-47F field maintenance manual once numbered -23&P is now the -20&P, dated December 30, 2025.
5 min readRead →
Technical manuals
Illustrated parts breakdowns and RPSTLs, explained
The parts list is an ordering authority, not a catalog: MIL-STD-40051E says the RPSTL authorizes the requisitioning, issue, and disposition of parts according to their SMR codes.
6 min readRead →
Technical manuals
Logistics product data (LPD): GEIA-STD-0007 and DI-SESS-81758A
DI-SESS-81758A ties every LPD delivery to GEIA-STD-0007, so the contract has to name the revision and the data elements being bought.
4 min readRead →
Technical manuals
Maintenance allocation chart (MAC): what it is and how to read it
The MAC decides which level may do a task: AR 750-1 sends any corrective action the MAC does not authorize at field level to sustainment maintenance, with turn-in documents.
5 min readRead →
Technical manuals
MIL-STD-3001, explained
NAVAIR work package manuals are prepared to this 8-part standard, and Part 1 says it is incomplete without the other 7.
5 min readRead →
Technical manuals
MIL-STD-3031 Army S1000D business rules
An Army S1000D delivery needs the correct specification issue, Army rules, and project decisions to meet its publication requirements.
6 min readRead →
Technical manuals
MIL-STD-38784, explained
MIL-STD-38784 reaches every technical document assigned a TM identification number, so a contract that cites it fixes the page layout, numbering, and warning format of the manual.
5 min readRead →
Technical manuals
MIL-STD-40051 work packages explained
Reading the correct work package and initial setup connects an authorized maintenance task to its required people, tools, and conditions.
6 min readRead →
Technical manuals
MIL-STD-40051-1, explained
MIL-STD-40051-1 was canceled on March 25, 2025, so a contract that still cites it points to a standard ASSIST no longer lists as active.
7 min readRead →
Technical manuals
MIL-STD-40051-2, explained
MIL-STD-40051-2 was canceled on March 25, 2025, so a contract that still cites 40051-2 or 40051/2C points to a standard ASSIST no longer lists as active.
5 min readRead →
Technical manuals
MIL-STD-40051, MIL-STD-3001, and MIL-STD-3031: the U.S. technical manual standards
Technical manuals are prepared under these standards, so a program office or contractor must know which standard and revision applies before handling a manual.
5 min readRead →
Technical manuals
MIL-STD-963 and DIDs, explained
A data deliverable on a defense contract is defined by a DID on a CDRL line, so the DID decides what the government receives.
5 min readRead →
Technical manuals
NAVAIR 00-25-100 and how a NAVAIR technical manual changes
Naval aviation maintainers may work only from approved, current technical data, and the NAMP gives a squadron 2 working days to incorporate a rapid action change.
4 min readRead →
Technical manuals
NAVAIR technical directives (TDs), explained
No configuration change may be made to naval aviation systems unless a NAVAIR technical directive directs it, and an Immediate Action TD is due before the next flight, the next use, or within 10 flight hours.
7 min readRead →
Technical manuals
S1000D standard numbering system (SNS) and the data module code
Every S1000D data module is filed and retrieved by its code, and the Issue 7 schema makes eleven parts of that code mandatory.
5 min readRead →
Technical manuals
Technical manual validation and verification, explained
Army procedures require hands-on verification of new and changed operating, maintenance, and troubleshooting procedures, and the PMCS table is verified hands-on in full.
4 min readRead →
Technical manuals
TMDE meaning (Army): test, measurement, and diagnostic equipment and calibration
A maintenance manual's torque value is only as good as the wrench that applies it, and a 2021 Army story relays that a Fort Campbell helicopter crash was traced to an uncalibrated torque wrench.
14 min readRead →
Technical manuals
Warning, caution, note in aviation manuals: what each means
These notices are how a manual warns a maintainer of injury and equipment risk, and the three standards that define them do not word them the same way.
4 min readRead →
Technical manuals
What is a BREX data module?
The Army and Air Force S1000D business rules both require a project BREX that references the next higher layer, so a manual delivered without one fails those rules.
5 min readRead →
Technical manuals
What is an IETM? Interactive electronic technical manuals explained
A technical manual shown on a screen is not necessarily interactive, and its class says what a reader can expect from it.
5 min readRead →
Technical manuals
What is ASD-STE100 Simplified Technical English?
STE became a requirement for ATA 100, now iSpec 2200, and is a recommendation in S1000D, and the Army and Air Force S1000D rules carry a language code for it.
6 min readRead →
Technical manuals
What is ATA iSpec 2200?
Commercial derivative aircraft in Navy and Marine Corps service come with manufacturer manuals built to iSpec 2200, and the FAA's own system codes derive from the same ATA numbering.
5 min readRead →
Technical manuals
What is S1000D? The specification, current issue, CSDB, and SNS
S1000D builds a technical manual from XML data modules in a common source database, with a Defense Interest Group speaking for defense customers.
5 min readRead →
Markings and data rights
CG-RN-1, CG-RN-2, and the NNPI CUI category
Unclassified NNPI is CUI with a NOFORN control by DoD rule and is exempt from mandatory public disclosure, so a holder cannot treat it like other unclassified technical data.
5 min readRead →
Markings and data rights
Classification markings: portion marks, banner lines, and the classification block
Markings tell a reader what is classified, at what level, and for how long, and 32 CFR 2001.21 sets where each one goes.
6 min readRead →
Markings and data rights
Controlled technical information (CTI), explained
Technical information a contractor creates for a DoD contract can be covered defense information under DFARS 252.204-7012, so the clause can reach a contractor's own drawings, data sets, and code.
5 min readRead →
Markings and data rights
CUI category: Export Controlled (EXPT)
An export-controlled document in DoD hands carries a CUI banner, a distribution statement, and a separate export-control warning, and each comes from a different rule.
5 min readRead →
Markings and data rights
CUI category: General Procurement and Acquisition
The Procurement Integrity Act makes disclosing or obtaining bid, proposal, or source selection information before award a crime, and the CUI marking is how a holder knows which documents those are.
5 min readRead →
Markings and data rights
CUI category: Privileged Safety Information (PSI)
A mishap safety report is protected so that witnesses and investigators speak freely, and the same facts in a legal investigation are not protected at all.
5 min readRead →
Markings and data rights
Data rights legends on technical data: unlimited, government purpose, and limited rights
A legend, or its absence, shows what limits a contractor asserts on delivered technical data, and unmarked data is presumed to carry unlimited rights.
6 min readRead →
Markings and data rights
DD Form 2345 and the Joint Certification Program, explained
DoDI 5230.24 bars release of Statement C or D export-controlled technical information outside the Government or to contractors without first verifying a current, valid DD Form 2345.
5 min readRead →
Markings and data rights
Derivative classification study guide (IF103.16)
If you make a new document from classified information, you are responsible for marking it correctly, and improper classification can bring sanctions.
10 min readRead →
Markings and data rights
Derivative classification, explained: EO 13526, SCGs, and the two-year training rule
Anyone who restates classified information in a new document must mark it, and under 32 CFR 2001.71 a derivative classifier without training in the past two years has that authority suspended.
13 min readRead →
Markings and data rights
DFARS 227.7202: commercial software licenses in DoD contracts
For commercial software the Government's rights are the ones in the license, so the license you attach to the contract decides what DoD may do with the software.
7 min readRead →
Markings and data rights
DoD distribution statements A through F: levels, distro F, and DoDI 5230.24
The letter on a marked technical document sets who may receive it without further approval, so a misread letter can put it in front of the wrong audience.
13 min readRead →
Markings and data rights
DoD Mandatory CUI Training answers: every knowledge check, explained
CUI marking and handling mistakes are what the training is built to prevent, and the same rules decide how you mark, store, send, and destroy CUI at work.
7 min readRead →
Markings and data rights
DoDI 5230.09 clearance for public release
Public release requires the right approval route, even when the proposed material is unclassified.
7 min readRead →
Markings and data rights
Export control warning statement: the DoDI 5230.24 wording and when to use it
A DoD technical document that holds export-controlled data must carry the warning and a B, C, D, or E distribution statement, and a wrong release can be an unlicensed export.
7 min readRead →
Markings and data rights
FOUO meaning, and whether legacy markings automatically become CUI
Documents marked FOUO are still in circulation, and 32 CFR 2002.36 requires any CUI reused from them to go into a new document marked as CUI.
13 min readRead →
Markings and data rights
ITAR vs EAR: USML, CCL, the 600 series, and paragraph (x)
An item described on the U.S. Munitions List is a defense article under the ITAR, and the Commerce order of review says such an item needs no CCL review at all.
6 min readRead →
Markings and data rights
MIL-STD-130 and IUID marking requirements
The physical mark and delivery record must identify the same individual item throughout its service life.
6 min readRead →
Markings and data rights
MIL-STD-31000B and technical data packages: levels, types, and DIDs
A product level TDP is meant to let a competent manufacturer build the item without going back to the original designer, so what a contract orders shapes what the Government can buy and repair later.
5 min readRead →
Markings and data rights
Patent rights under federal contracts, explained
A contractor that misses a disclosure or election deadline under its patent rights clause can be asked to assign the invention's title to the Government.
4 min readRead →
Markings and data rights
SBIR data rights: what the legend protects and for how long
The SBIR/STTR legend shows whether data is still in its 20-year protection period and what rights the Government holds when the period ends.
6 min readRead →
Markings and data rights
What else must a derivative classifier carry forward from an authorized source?
A derivative classifier who drops the downgrading or declassification instructions can leave information protected too long or not long enough.
7 min readRead →
Markings and data rights
What is a data spill? Classified data spills explained
DoD policy handles every classified spill as a possible compromise until an inquiry decides whether an unauthorized disclosure occurred, so the receiving system's status alone sets off the response.
4 min readRead →
Markings and data rights
What is CUI? CUI Specified vs CUI Basic, categories, and markings
The CUI banner and designation indicator show how a document is controlled and which agency designated it, so a holder needs to read both.
5 min readRead →
Markings and data rights
What is NOFORN?
NOFORN and REL TO decide whether a page can be shown to a coalition partner, so anyone handling marked documents needs to read the dissemination control, not just the classification.
5 min readRead →
Naval aviation
Airborne early warning at sea: from the E-1 Tracer to the E-2D
Designations, squadron names, and dates for the E-1 and E-2 differ across official Navy pages, so a record about either aircraft needs its source.
4 min readRead →
Naval aviation
ALSSA multi-Service brevity codes (April 2025), explained
A brevity code means only what the users agreed it means, so a tool that invents a meaning gives an operator a fabricated fact to act on.
13 min readRead →
Naval aviation
AN/APY-9 meaning and JETDS equipment names
Correctly reading a full equipment designation helps match technical documents and parts to the intended configuration.
6 min readRead →
Naval aviation
Carrier catapults and arresting gear, explained
DOT&E's fiscal year 2025 report says EMALS and AAG reliability and maintainability still hurt sortie generation, the greatest risk to proving the Ford class in operational testing.
4 min readRead →
Naval aviation
Cooperative Engagement Capability (CEC), explained
DOT&E reported in FY2020 that the Navy had not shown the E-2D's CEC supporting its air and missile defense mission, and in FY2024 that it had taken no action on a retest.
7 min readRead →
Naval aviation
E-2D Advanced Hawkeye: what it is, crew, and specs
The E-2D supplies early warning and command and control to a carrier strike group, so anyone supporting naval aviation needs its crew, office, and dates right.
4 min readRead →
Naval aviation
E-2D crew positions and training: what a Naval Flight Officer does aboard the E-2
Training, manuals, and software for the E-2D crew depend on who is aboard and what the Navy calls them, and the public pages use two titles.
4 min readRead →
Naval aviation
Flight deck jersey colors: what each color means on a Navy carrier
The Naval History and Heritage Command says color-coded jerseys keep order and safety on the flight deck, so a wrong color in a manual, caption or training aid assigns a job to the wrong team.
4 min readRead →
Naval aviation
MIL-HDBK-516 airworthiness certification criteria
An approved certification basis connects the aircraft configuration, compliance evidence, accepted risks, and permission to fly.
6 min readRead →
Naval aviation
MIL-STD-1553 data bus explained
Reading the bus roles and exact standard revision helps an integration team specify compatible aircraft interfaces.
6 min readRead →
Naval aviation
NATOPS interim changes, explained
An interim change is how a safety-of-flight fix reaches every cockpit in days, so anyone who maintains or quotes a NATOPS manual needs to know how ICs are numbered, released, and recorded.
5 min readRead →
Naval aviation
NAVAIRINST 13034.1 and flight clearances, explained
No Navy or Marine Corps aircraft may fly in a nonstandard configuration without a NAVAIR flight clearance, so every new store, box, or software load runs through this process.
6 min readRead →
Naval aviation
Navy Fleet Readiness Centers (FRC): locations and what they do
GAO found that Navy aviation depots finished maintenance on time or early only 45 to 63 percent of the time in fiscal years 2014 through 2019, and GAO says such delays cut the time aircraft are available for operations and training.
6 min readRead →
Naval aviation
OPNAVINST 3710.7 and CNAF M-3710.7, explained
CNAF M-3710.7 is the rule set every naval flight is planned and flown under, so anyone who quotes a Navy crew rest or weather limit needs the right document and edition.
7 min readRead →
Naval aviation
OPNAVINST 3750.6 and the Navy aviation safety program
A Class A aviation mishap triggers a telephone report to the Naval Safety Command within 60 minutes, so every aviation command needs the current thresholds and clocks.
5 min readRead →
Naval aviation
The Aircraft Signals NATOPS manual, explained
CNAF M-3710.7 directs every naval activity to comply with the signals in NAVAIR 00-80T-113, so a wrong signal in a manual, poster, or training aid is a flight deck hazard.
4 min readRead →
Naval aviation
The CV NATOPS manual (NAVAIR 00-80T-105), explained
Every carrier launch and recovery is run under a General Series NATOPS manual, so anyone who writes or checks carrier procedures needs the right manual, number, and edition.
4 min readRead →
Naval aviation
The NAMP and COMNAVAIRFORINST 4790.2E, explained
The NAMP sets maintenance policy at every level of naval aviation, and the November 2024 release message says Revision E incorporated 84 changes and 36 maintenance advisories.
13 min readRead →
Naval aviation
What is a carrier air wing (CVW)?
The Navy calls the air wing's striking power the cornerstone of carrier power projection, and the Chief of Naval Operations told Congress in May 2026 that a full air wing has more than 70 aircraft.
4 min readRead →
Naval aviation
What is a Fleet Replacement Squadron?
Every pilot, flight officer, aircrewman and maintainer reaches a fleet aircraft through an FRS, so the FRS syllabus and its instructors set the baseline the fleet inherits.
5 min readRead →
Naval aviation
What is a NATOPS pocket checklist?
Every naval aircrew member flies with a PCL, and every page in it must trace back to a promulgated NATOPS flight manual, so the derivation and change rules decide whether a checklist is current.
5 min readRead →
Naval aviation
What is a Naval Flight Officer?
NFOs run the sensors, weapons and battle management in the E-2D, P-8A, F/A-18F and EA-18G, so training, manuals and software for those aircraft have to name the NFO's role correctly.
5 min readRead →
Naval aviation
What is Link 16? Frequency band, standard, and J-series messages
A January 2025 Navy program office sheet calls Link 16 a standardized system used by U.S., NATO and coalition forces, and says the office delivers its terminals to 58 nations and NATO.
5 min readRead →
Naval aviation
What is NATIP?
NATIP and NATOPS together form a naval aircraft's permanent flight clearance, so anyone quoting a weapon or mission system limit needs to know which product holds it and who may change it.
4 min readRead →
Naval aviation
What is NATOPS? Meaning, manuals, and the NATOPS check
NATOPS standardizes how naval aircrew and aviation facilities operate and are evaluated, so anyone who quotes or maintains its manuals needs the change process.
13 min readRead →
Army aviation
15U CH-47 helicopter repairer: job, ASVAB score, and AIT
Army doctrine ties the repairer's work to the manual: TC 3-04.71 says that in all cases maintainers must follow published requirements in Army technical manuals.
5 min readRead →
Army aviation
AOAP Army: the Army Oil Analysis Program in 2026
Oil analysis can flag a failing engine or transmission before it fails, and a 2025 Army story cites a message reporting that use of AOAP labs fell 78 percent in a decade.
4 min readRead →
Army aviation
AR 95-1, Flight Regulations: what it covers and what is current
AR 95-1 sets who may fly Army aircraft and how crews train, and it makes the aircraft operator's manual and checklist the primary references for operating each aircraft.
5 min readRead →
Army aviation
Army aviation two-level maintenance: field vs sustainment
The authorized level decides where an aircraft goes for repair; Corpus Christi Army Depot reports that depot teams sent forward cut induction timelines from 180-270 days to about 15-30 days.
5 min readRead →
Army aviation
Army maintenance test flights: the Army flight test after maintenance
Under AR 95-1, an aircraft whose test procedures are in an MTF manual may be test flown only by qualified maintenance test pilots or evaluators, so the pilot's designation matters as much as the repair.
5 min readRead →
Army aviation
ASAM in the Army: aviation safety action messages vs SOF and SOUM
An emergency safety-of-flight message can ground an entire fleet at once, and only the Chief or Vice Chief of Staff of the Army can approve one.
5 min readRead →
Army aviation
CH-47 Chinook variants: from the CH-47A to the CH-47F Block II
Parts, training and manuals must match the model in service, and a 1978 GAO report counted 104 A, 74 B and 183 C models in one Army modernization plan.
4 min readRead →
Army aviation
CH-47 flight engineer: the job, the path, and the manuals
The Army lists the CH-47's minimum crew as a pilot, a copilot and a flight engineer, so training, checklists and support tools have to fit that third seat.
5 min readRead →
Army aviation
Corpus Christi Army Depot (CCAD): what it does, size, and history
CCAD is the Army's rotary-wing depot, and an April 2026 AMCOM story puts it at nearly 2,100 personnel and nearly 3 million square feet of production space.
4 min readRead →
Army aviation
DA Form 2408 series: Army aviation maintenance records, explained
A status symbol on an aircraft's maintenance record decides whether it may fly, and the Army planned to train about 51,000 users on the GCSS-Army system that now integrates aviation maintenance data.
16 min readRead →
Army aviation
DoD and Army mishap classes A through E, explained
A mishap's class sets how it is investigated and reported; the Army requires immediate telephone notice to its Combat Readiness Center for Class A, Class B and aviation Class C mishaps.
5 min readRead →
Army aviation
FMC meaning in the Army: fully mission capable and mission capable rates
Mission capable rates are the main public measure of fleet health, and GAO reported in November 2022 that the CH-47F met its mission capable goal in none of fiscal years 2011 through 2021.
5 min readRead →
Army aviation
Forward area refueling equipment: FARE and AAFARS
Confusing FARE-2, AAFARS, and HTARS can put the wrong flow rates, fuel requirements, or manuals into a refueling plan.
6 min readRead →
Army aviation
Forward arming and refueling points (FARPs), explained
A FARP lets aircraft refuel and rearm without flying back to base, and ATP 3-04.17 estimates an eight-point FARP needs at least ten 92F fuelers and sixteen arming personnel.
7 min readRead →
Army aviation
Helicopter brownout and degraded visual environments (DVE), explained
The Army's safety center reported in 2014 that 88 of 367 Class A and B accidents since the start of the Afghanistan and Iraq wars involved degraded visual environments.
5 min readRead →
Army aviation
Helicopter sling load manuals: TM 4-48.09, 4-48.10, and 4-48.11
Every Army sling load must pass a qualified inspector before the aircraft arrives, and these manuals define the inspector, the rigging and the load.
4 min readRead →
Army aviation
How Army aviators are trained: flight school at Fort Rucker
A 2022 Army story quotes the Cairns airfield manager saying every Army aviator goes through the Lakota training program, and in August 2026 the Army chose a contractor to run initial entry training for 800 to 1,500 pilots a year.
5 min readRead →
Army aviation
Phase maintenance: how aircraft phase inspections work in Army aviation
Phase inspections take aircraft off the flight line for weeks: ATP 3-04.7 sets goals of 28 working days for a CH-47F 200-hour phase and 39 days for a 400-hour phase.
5 min readRead →
Army aviation
PMCS meaning (Army): preventive maintenance checks and services, explained
The Army calls operator PMCS the cornerstone of its maintenance system, and a service more than 10 percent late makes equipment count as not mission capable.
6 min readRead →
Army aviation
What is a combat aviation brigade?
Units, manuals and sustainment plans are built around the brigade's battalions, and the Army's FY2026 budget, as quoted by CRS, supports 11 combat aviation brigades.
5 min readRead →
Disconnected operations
Commercial Solutions for Classified (CSfC), explained
NSA's CSfC FAQ says NSA/CSS policy makes CSfC the first option to be considered for a cybersecurity requirement, and only products on the CSfC Components List may be used to protect classified data this way.
7 min readRead →
Disconnected operations
Continuous ATO (cATO), explained
The February 2022 memo says DoD's Risk Management Framework practice focused on obtaining ATOs but fell short on continuous monitoring once a system was authorized.
6 min readRead →
Disconnected operations
DoD cloud impact levels: IL2, IL4, IL5, and IL6 explained
Under DISA's June 2026 guide, every cloud offering at Impact Level 4 and above needs a provisional authorization at that level before it hosts mission systems, and FedRAMP baselines do not address national security systems.
5 min readRead →
Disconnected operations
DoD zero trust strategy: 4 goals, 7 pillars, and the FY2027 deadline
A July 2025 directive-type memorandum, which reached its stated expiration on July 17, 2026, told DoD Components to reach at least target level zero trust on all unclassified and classified systems by the end of fiscal year 2027.
5 min readRead →
Disconnected operations
FedRAMP explained: what it is, the classes, and FedRAMP 20x
By law, the assessment in a FedRAMP package is presumed adequate for an agency's authorization to operate, yet FedRAMP says its certification does not guarantee a service meets every requirement for agency use.
6 min readRead →
Disconnected operations
How much memory does an LLM need? Weights plus KV cache
On a machine with no network the model must fit the memory on hand, and a Hugging Face guide computes that one model's 16,000-token cache needs about 15 GB, half as much as its weights.
6 min readRead →
Disconnected operations
How to verify that software makes no network calls
Software for a disconnected network can still call out for updates or telemetry, and each way of testing that claim covers only part of the question.
4 min readRead →
Disconnected operations
ICD 703, explained
Every program that touches SCI works under ICD 703, which decides who may hold it, where it may be used, how it is marked, and how it reaches people outside the Intelligence Community.
5 min readRead →
Disconnected operations
ICD 705, explained
Any system meant for a SCIF will run inside a facility accredited under ICD 705, so its builders need to know what the directive, its standards, and its specification require.
6 min readRead →
Disconnected operations
ICS 705-1, explained
Any program that will run equipment inside a SCIF inherits the perimeter, alarm, access, acoustic, and TEMPEST rules of ICS 705-1, so its builders need to know what the standard demands and who decides.
5 min readRead →
Disconnected operations
ICS 705-2 (ICD 705-2): SCIF accreditation and reciprocity
A program that needs SCIF space will either accredit a facility or borrow one under ICS 705-2, so its schedule and its reciprocity depend on the rules in this standard.
5 min readRead →
Disconnected operations
MIL-STD-810H environmental test methods explained
A durability claim is useful only when the test evidence matches the equipment and its intended operating conditions.
6 min readRead →
Disconnected operations
NIST SP 800-171 in one read
Contracts involving controlled unclassified information require NIST SP 800-171, and DoD contracts under class deviation 2026-O0025 name Revision 2.
5 min readRead →
Disconnected operations
NIST SP 800-172 enhanced security requirements
Using the wrong edition can produce an assessment checklist that does not match the defense contract's requirements.
5 min readRead →
Disconnected operations
NIST SSDF (SP 800-218): 2026 updates, practices, and attestation status
From September 2022, OMB required federal agencies to obtain a secure development attestation built on SP 800-218 before using covered software, until a January 23, 2026 memo made it each agency's choice.
5 min readRead →
Disconnected operations
Removable media and one-way transfer into isolated networks
NIST's September 2025 guidance says portable storage media remain useful for moving data to and from operational technology, for example to update firmware on a device in an isolated network.
4 min readRead →
Disconnected operations
SCIF types and SCIF vs closed area: closed storage, open storage, T-SCIF
Software proposed for classified work runs in one of these spaces, and each one has its own approver, its own storage rule, and its own view of what may connect to the outside.
5 min readRead →
Disconnected operations
Security clearance levels, explained
A clearance at a level is only eligibility: under Executive Order 13526, access also takes a signed nondisclosure agreement and a need-to-know.
5 min readRead →
Disconnected operations
Signed software updates and releases: how code signing works
A signature on software carried into a disconnected network can be checked only against keys and roots already on the verifying machine, and the cosign README marks its air-gapped section out of date.
5 min readRead →
Disconnected operations
The Update Framework (TUF) and offline software updates
An attacker who can answer a client's update requests can replay old metadata, hold back new updates, or serve files that never belonged together, unless signed metadata says otherwise.
7 min readRead →
Disconnected operations
What does air-gapped mean?
NSA and CISA assessment teams have often reached operational networks described as fully air gapped, through special purpose, forgotten, or accidental connections, according to an October 2023 advisory.
4 min readRead →
Disconnected operations
What does DDIL mean?
A system that needs a steady network stops working where military networks are denied or cut, and DoD and Army plans expect systems to work in DDIL conditions.
5 min readRead →
Disconnected operations
What FIPS 140-3 validated means
The Cryptographic Module Validation Program says non-validated cryptography is viewed as providing no protection, so the data would be treated as unprotected plaintext.
5 min readRead →
Disconnected operations
What is a cross domain solution? File transfer, NSA, and DISA roles
Under DoDI 8540.01, a DoD cross domain requirement must be met from the baseline list, and any solution not on it needs a letter of exception.
5 min readRead →
Disconnected operations
What is a DoD authority to operate (ATO)?
Under DoDI 8510.01 of July 19, 2022, every DoD system has an authorizing official, who must be government personnel and may not delegate the authorization decision.
4 min readRead →
Disconnected operations
What is a SAPF? Meaning, requirements, and SAPF vs SCIF
Special access program information may be handled only inside an accredited SAPF or an accepted SCIF, so software proposed for one must work with no connection to the outside and without the personal devices the manual generally bars.
13 min readRead →
Disconnected operations
What is a SCIF?
Sensitive compartmented information can be handled only inside an accredited SCIF, so software proposed for one cannot assume a connection to the outside.
6 min readRead →
Disconnected operations
What is a STIG? What STIG stands for and how it works
DoD Instruction 8510.01 has the DISA Director keep STIGs consistent with NIST SP 800-53, and STIGs give assessors a basis for checking those controls under the Risk Management Framework.
5 min readRead →
Disconnected operations
What is an SBOM?
Executive Order 14028 listed giving a purchaser an SBOM for each product among the practices federal software guidance must cover, and a January 2026 OMB memo still lets agencies require one on request.
4 min readRead →
Disconnected operations
What is retrieval-augmented generation (RAG)?
A system that answers from a closed set of manuals can show the passage behind each answer, yet a 2024 preregistered study found retrieval-based legal research tools still hallucinated 17% to 33% of the time.
5 min readRead →
Disconnected operations
What is TEMPEST? Standards, testing, and the TEMPEST assessment
Any system that processes classified information electronically in a SCIF or SAPF gets a TEMPEST countermeasures review, so what the equipment radiates and what it connects to are design questions, not afterthoughts.
5 min readRead →
Disconnected operations
What is the NISPOM?
Every classified contract runs on the NISPOM, so a vendor that wants its software inside a cleared facility works under the facility security officer's rules, not its own.
5 min readRead →
Evaluating language models
AI agent benchmarks: SWE-bench, tau-bench, GAIA, WebArena, and what they measure
An agent that acts on files, databases or the web is judged by whether the job gets done, and at publication the best model resolved 1.96 percent of SWE-bench's real software issues.
5 min readRead →
Evaluating language models
CDAO Responsible AI Toolkit: what it requires, GenAI guardrails, and fielded models
A contractor whose AI project will be assessed by a Department program office needs to know the questions the toolkit asks before the program office asks them.
6 min readRead →
Evaluating language models
Direct vs indirect prompt injection, and the OWASP LLM01 defenses
Any document a language system reads is an input channel, and the 2023 paper that named indirect injection notes that even offline models retrieving documentation files could be injected through poisoned input data.
5 min readRead →
Evaluating language models
DoD AI ethical principles and DoD Directive 3000.09
The principles apply to both combat and non-combat functions, and DoDD 3000.09 requires AI in autonomous and semi-autonomous weapon systems to be consistent with them.
4 min readRead →
Evaluating language models
How to measure whether a cited source supports the answer
A citation shows that a source was attached, not that it supports the claim. An operator who acts on a cited page relies on the second.
5 min readRead →
Evaluating language models
ISO/IEC 42001 explained: the AI management system standard
Vendors are starting to offer ISO/IEC 42001 certificates as evidence of responsible AI, and a program office has to know that the certificate covers a management system, not the model or its outputs.
5 min readRead →
Evaluating language models
LLM abstention: how to make a model say "I don't know" and measure it
A system that answers from doctrine or manuals must refuse when they are silent, and a made-up answer reads like a sourced one.
5 min readRead →
Evaluating language models
LLM calibration: what "Language Models (Mostly) Know What They Know" found
A system that refuses when unsure is only as good as its confidence score, and the GPT-4 Technical Report shows post-training raising calibration error on an MMLU subset from 0.007 to 0.074.
5 min readRead →
Evaluating language models
LLM hallucination: what it is, why it happens, and how to measure it
NIST warns that confabulated logic or citations can mislead people into inappropriately trusting a system's output, so a fluent answer is not evidence of a correct one.
5 min readRead →
Evaluating language models
LLM-as-a-judge: how it works, how accurate it is, and its biases
Evaluations of language systems are often scored by another model, and in the 2023 MT-Bench study GPT-4 gave the same verdict after two answers swapped places in only 65 percent of cases.
5 min readRead →
Evaluating language models
Long-context evaluation: how to test whether an LLM uses its whole context window
A system that loads whole manuals into a model's input relies on the model using all of it, and in RULER's results table only 10 of 17 models claiming 32K tokens or more held up at 32K.
4 min readRead →
Evaluating language models
Model cards and datasheets, explained
A model card is often the only description of a model that reaches a program office, and a 2024 study of 32,111 cards found a filled-in limitations section in only 17.4 percent of them.
6 min readRead →
Evaluating language models
Model quantization explained: bits, GGUF types, and accuracy
A model must fit the memory of the machine where it runs, and each step down in bits costs accuracy that the file name does not show.
5 min readRead →
Evaluating language models
NIST AI 600-1, explained: the Generative AI Profile
NIST AI 600-1 names 12 generative AI risks and more than 200 actions to manage them, so a requirement or a compliance claim that cites it has to match what the document says.
5 min readRead →
Evaluating language models
NIST AI Risk Management Framework (AI RMF 1.0): summary, 4 functions, Playbook
The AI RMF gives buyers and vendors a shared vocabulary for AI risk, and as of November 2026 NIST says version 1.0 is being revised under the July 2025 White House AI Action Plan.
6 min readRead →
Evaluating language models
OWASP top 10 for LLM applications
A security review must cover retrieval, tools, data, outputs, and resource limits as well as the model's responses.
5 min readRead →
Evaluating language models
RAG evaluation: metrics, RAGAS vs ARES, and how to run one
A system that must answer only from its documents can be judged by an automated faithfulness score, and a 2024 study scored two such faithfulness metrics at 7.83 and 8.80 for correlation with human preferences, on a scale where two human annotators correlated at 70.09.
5 min readRead →
Evaluating language models
Red teaming language models: what it is, how it is done, and what it finds
NIST's generative AI profile names red teaming as a pre-deployment test, and one 2022 study collected 38,961 human red team attacks yet called its own data incomplete.
5 min readRead →
Evaluating language models
Semantic entropy, explained
A system that refuses when unsure needs a signal for when it is guessing, and semantic entropy reached an AUROC of 0.790 against 0.691 for plain entropy across 30 task and model pairs.
5 min readRead →
Evaluating language models
Towards Understanding Sycophancy in Language Models (ICLR 2024), and LLM sycophancy explained
A model that defers to the person asking will confirm a mistaken premise, and in one 2023 study an assistant wrongly admitted a mistake on 98 percent of questions when a user pushed back.
4 min readRead →
Evaluating language models
What is a held-out test set? Definition, splits, and leakage
A score from a test set that leaked into training or tuning overstates how a system will perform on documents it has never seen.
5 min readRead →
Evaluating language models
What is a system card? AI system cards explained
A program office that buys or integrates a frontier model gets the developer's own account of the model's testing in a system card, and has to know what that document can and cannot settle.
5 min readRead →
Evaluating language models
What is FAISS? Full form, index types, and is it a vector database
Retrieval decides which passages a system answers from, and a recall figure shows how well an index matches exact search, not whether it found the right page.
6 min readRead →
Evaluating language models
What is perplexity in language models? Formula, examples, limits
A vendor's perplexity figure is only comparable with the same tokenizer and test text, and a low number does not show that a model answers questions well.
5 min readRead →
AI agents and markets
Algorithmic collusion: what the pricing experiments found
Pricing software already runs in real markets, and a 2024 study of German gasoline stations found that margins in duopoly and triopoly markets rose only when every station had adopted it.
5 min readRead →
AI agents and markets
Anchoring bias: what it is and how language models show it
The 1974 paper that named anchoring was supported by the Defense Department's Advanced Research Projects Agency, and the same bias shows up when language models estimate a number after seeing one.
6 min readRead →
AI agents and markets
Dictator game vs ultimatum game: the difference and the results
The ultimatum game models any take-it-or-leave-it offer, and a 2024 test judged ChatGPT-4 more human-like than a random person in 5 of 8 comparisons drawn from classic behavioral games.
4 min readRead →
AI agents and markets
Double oral auction: Vernon Smith's market experiments
Smith's design gives a market whose right answer is known in advance, and a September 2026 replication found that markets of language-model traders did not fully reach it.
4 min readRead →
AI agents and markets
Experimental asset bubbles: what lab markets show and why
In the 1988 experiments, 14 of 22 markets produced bubbles and crashes even though every trader could compute the asset's value from announced dividends.
4 min readRead →
AI agents and markets
Generative Agent Simulations of 1,000 People, and generative agents explained
Simulated people are proposed as stand-ins for survey respondents and test subjects, and a 2024 study found that synthetic survey answers varied less than real ones and shifted with small wording changes.
4 min readRead →
AI agents and markets
Herding and information cascades, explained
Once a cascade starts, each new decision adds no information, and in the 1998 worked example a cascade is wrong with probability 0.247 even when each private signal is right 70 percent of the time.
5 min readRead →
AI agents and markets
Information aggregation in experimental markets, explained
A market is only as informative as its prices, and in Plott and Sunder's 1988 single-security markets with diverse traders, efficiency averaged only 47 percent even though pooled clues identified the state.
5 min readRead →
AI agents and markets
LLMs as economic agents: what the experiments show
Agents that look rational in one format can fail in another: in one 2023 study, average rationality scores in a social-preference task fell from 0.997 to 0.698 when only the price wording changed.
5 min readRead →
AI agents and markets
On the Conversational Persuasiveness of GPT-4, and what LLM persuasion studies found
Persuasion runs in both directions for AI agents, and a 2024 study found that persuasive misinformation flipped GPT-4's correct factual answers 20.7 percent of the time.
4 min readRead →
AI agents and markets
Prediction markets: what they are and how well they forecast
Prediction markets are a growing source of probability estimates, with more than $25 billion traded on CFTC-registered venues in 2025, and published tests find language models still trail the best human forecasters.
5 min readRead →
AI agents and markets
Public goods game: rules, payoffs, and what people do
Any shared account rewards free riding, and in a 2026 test none of 11 language models matched human giving by the final round of a public goods game.
4 min readRead →
AI agents and markets
The p-beauty contest game: rules, equilibrium, and results
Keynes likened the stock market to guessing what others will guess, and in Nagel's 1995 sessions with p below 1, no first-round subject chose the equilibrium number, zero.
5 min readRead →
AI agents and markets
The repeated prisoner's dilemma and tit for tat, explained
Agents that deal with each other again and again face the same choice between trust and exploitation, and in a 2025 study GPT-4 never resumed cooperating with a partner that defected once.
4 min readRead →
AI agents and markets
The winner's curse in common-value auctions
A bidder that ignores what winning implies overpays, and in the 1983 jar-of-coins auctions the average winning bid was $10.01 for jars worth $8.00.
5 min readRead →
Document processing
BM25 and hybrid search: the formula, k1 and b, and RRF
Parts and maintenance queries turn on exact strings, and in a 2021 test of questions about named entities a dense retriever found the answer passage 49.7% of the time against 72.0% for BM25.
5 min readRead →
Document processing
Dense vs sparse retrieval: the difference, and when each wins
A retriever that misses a part number or a clause number cannot be fixed downstream, and in BEIR's zero-shot tests the dense model DPR scored 47.7% below BM25 on average.
6 min readRead →
Document processing
Document cracking vs chunking: how documents are cut up for retrieval
On a procedure, a cut between a warning and its step returns an action without its hazard. A chunk with no source record cannot be checked.
6 min readRead →
Document processing
Document layout analysis: what it is, datasets, and tools
A Mask R-CNN layout model trained on scientific articles scored 93 mAP on its own test set and 30 on DocLayNet's varied pages, which include manuals.
5 min readRead →
Document processing
How to read a model license: open weights are not all open
A model license sets the terms of use. The Llama 4 acceptable use policy lists prohibited uses, and one item names military and warfare.
5 min readRead →
Document processing
LayoutLM, LayoutLMv2, and LayoutLMv3, explained
A model that reads a form by its layout can pull a field from a scanned maintenance record, but it reads whatever the OCR gave it and two of the three versions are licensed for non-commercial use.
5 min readRead →
Document processing
OCR on scanned manuals: how to do it and what goes wrong
A misread digit in a part number or a torque value looks valid and is wrong. On a maintenance manual that is a safety problem, not a typo.
6 min readRead →
Document processing
Product quantization for vector compression, explained
A corpus of millions of document embeddings only fits on one machine if the vectors are compressed, and the compression chosen decides how many right passages the search misses.
5 min readRead →
Document processing
Recall@k, MRR, and nDCG, explained
A retrieval score decides which search method feeds an answering system, and in one 2002 rejudging, 25.0 percent of documents TREC had counted as relevant were rated not relevant on a graded scale.
14 min readRead →
Document processing
Reciprocal rank fusion, explained
Hybrid search merges a keyword list and a vector list, and the merge rule decides which passage a reader or a language model sees first and cites.
6 min readRead →
Document processing
Table extraction from PDFs: pdfplumber, Camelot, Tabula, and scans
A torque table or parts list read with its columns shifted by one cell gives a technician a wrong value that looks right, and most extraction tools only work on PDFs that still contain text.
6 min readRead →
Document processing
Tesseract OCR: install it, run it, and get accurate text
Tesseract is the OCR engine inside many document pipelines, and a scanned manual's part numbers and torque values are only as good as what the engine read.
6 min readRead →
Document processing
What are Matryoshka embeddings? Truncating embedding dimensions
An embedding index for a document collection is sized by dimensions times vectors, and a model trained the Matryoshka way lets that size be cut after the fact without re-embedding.
5 min readRead →
Document processing
What is a reranker?
A system can cite only the passages it keeps, and in a 2022 study a trained reranker raised the rate of correct source passages on Natural Questions from 64.65 to 70.78 R-Precision.
6 min readRead →
Document processing
What is an embedding model?
The embedding model decides which passages a document system can find at all, so its language coverage, input limit, and benchmark scores are part of what is being bought.
6 min readRead →
Document processing
What is ColBERT? Late interaction and MaxSim, explained
In the ColBERT paper late interaction was competitive with BERT re-rankers at two orders of magnitude lower latency, but the index it needs is many times the size of a single-vector index, and that cost has to be planned for.
6 min readRead →
Document processing
What is cross-lingual retrieval? English queries, Russian documents
A search in English can rank English documents ahead of a better Russian match, and it returns the Russian passage, not a translation.
5 min readRead →
Document processing
What is HNSW? The HNSW index, M, efSearch, and memory
Faiss, Lucene and pgvector all ship an HNSW index, and the parameters chosen at build time decide how often the right passage is missed and how much RAM the index needs.
6 min readRead →
Document processing
What is the BEIR benchmark? Datasets, results, and how to run it
A retriever picked from a leaderboard was scored on English web, science, and news text, and the benchmark's own authors say its judgments favor keyword systems and its documents run a few hundred words.
6 min readRead →
Document processing
Why a PDF manual is hard to read by machine
When software guesses a manual's structure, a warning can end up detached from the step it governs, and the extracted text will not show it.
5 min readRead →
Languages
Chinese word segmentation for search and translation
The U.S. Department of Defense entered the 2005 international Chinese word segmentation bakeoff, and the organizer's report still called unknown words the field's main problem.
5 min readRead →
Languages
chrF metric for machine translation: formula, chrF vs BLEU, and COMET
A chrF, BLEU, or COMET number can only be compared with another one scored on the same test set, with the same variant, settings, and model version.
7 min readRead →
Languages
Egyptian Arabic (arz) for machines
Egyptian Arabic is the most widely understood Arabic dialect, and its chat and social media text mixes Arabizi with spellings that follow no convention, so a system scored only on Modern Standard Arabic has not been tested on the text it will actually meet.
6 min readRead →
Languages
Farsi, Dari, and Pashto: what a language system has to tell apart
A document labeled with the wrong language, or at the wrong code level, goes to the wrong translation or search path and will not match later records.
5 min readRead →
Languages
ILR scale vs DLPT score: levels, tests, and what counts
Under DoDI 5160.71, DLPT system tests are the only battery authorized to certify DoD personnel for foreign language proficiency bonus or pay, except as law or the instruction allows.
7 min readRead →
Languages
Iraqi (Mesopotamian) Arabic (acm) for machines
Iraqi Arabic reaches analysts with letters the standard Arabic alphabet lacks and words no MSA dictionary holds, and the largest NLLB-200 model's metrics file shows English into Iraqi as its weakest Arabic direction.
5 min readRead →
Languages
Korean for machines: Hangul blocks, spacing, and particles
Korean spelling rules allow a military term such as 중거리 탄도 유도탄 to be written with or without spaces, so a search or glossary that holds one spelling can miss the other.
5 min readRead →
Languages
Kurdish for machines: Kurmanji, Sorani, and two scripts
In a 2023 test, Sorani text typed with Arabic letter conventions lost about 26 BLEU points in translation to English, and character normalization recovered almost 20 of them.
5 min readRead →
Languages
Language identification: the machine learning survey, fastText lid.176, and NLLB
A 2020 study found that human-judged language identification accuracy in web-crawled corpora was only around 5 percent for many lower-resource languages, although the models scored well on held-out test sets.
5 min readRead →
Languages
Levantine Arabic (apc, ajp) for machines
Levantine Arabic is the everyday language of Syria, Lebanon, Jordan, and the Palestinian territories, and a system that reads only Modern Standard Arabic will miss the function words that carry most of a Levantine sentence.
5 min readRead →
Languages
Machine translation quality estimation, explained
A document translated for use has no reference translation to score against, and in WMT 2023 even the top quality estimation systems missed some hallucinated translations.
4 min readRead →
Languages
Machine translation terminology: how to make MT use your glossary
DoD Instruction 5025.12 makes the DoD Dictionary the primary terminology source for DoD correspondence, including policy, strategy, doctrine, and planning documents.
5 min readRead →
Languages
Najdi and Gulf Arabic (ars) for machines
A system that reports a Najdi Arabic score has been tested on a FLORES-200 set that the NLLB paper says matches Modern Standard Arabic almost exactly, so the number says little about Emirati, Kuwaiti, or Qatari text.
5 min readRead →
Languages
Named entity recognition across languages
On human-annotated news and forum text from the DARPA LORELEI program, name taggers trained on gold data scored between 42.2 and 70.7 F1 across nine languages.
5 min readRead →
Languages
NLLB-200 and what low-resource language means
The weights are licensed for non-commercial use only, and the model cards rule out production, documents, and medical or legal text, so check both before you build on NLLB-200.
6 min readRead →
Languages
One person, ten spellings: transliterating names
The same person can appear in three documents under three correct Latin spellings, so a name search that compares spellings directly can miss the match.
6 min readRead →
Languages
Pashto machine translation: script, resources, and pitfalls
Pashto documents reach analysts in a script with twelve letters Arabic lacks, in two main dialects, and often as scans, so a translation score alone does not say whether a system will read them.
6 min readRead →
Languages
Post-editing machine translation: ISO 18587, light vs full, and TER
HTER, an edit-rate measure of post-editing effort, was slated for the 2006 machine translation evaluations of GALE, a DARPA program, according to the paper that defined it.
4 min readRead →
Languages
TER translation edit rate: formula and examples
A defense translation evaluation needs consistent references and scoring settings before TER results can support a system comparison.
6 min readRead →
Languages
The Unicode Bidirectional Algorithm (UAX #9), explained
The W3C notes that a browser search, which captures text in logical order, can fail to match right-to-left text stored in visual order.
6 min readRead →
Languages
TMX translation memory: the file format, and which TMS vendors restrict export
A fuzzy match score decides which stored translation a translator sees, and the retrieval method used by commercial translation memory tools is typically not disclosed.
4 min readRead →
Languages
Tokenizers and language cost: why some languages need more tokens
A model reads a fixed number of tokens, and one 2023 study found the tokenizer of a widely used commercial model needed about 3 times as many tokens for Arabic as for English.
5 min readRead →
Languages
UDHR in every language: download all 533 translations as text and XML
OHCHR lists 578 translations of the Universal Declaration of Human Rights, and the same preamble and 30 articles in each one make it the widest parallel text in common use.
5 min readRead →
Languages
Ukrainian vs Russian alphabet: how to tell the Cyrillic languages apart
The Library of Congress tables romanize the same Cyrillic letter г as g for Russian and h for Ukrainian, so a name's Latin spelling depends on which language a system decided it was reading.
4 min readRead →
Languages
Urdu and the Nastaliq script: why OCR and search struggle
In a 2023 study, an Urdu OCR model scored 99.17 percent on synthetic test lines and 64.73 percent on real printed lines after training on the synthetic set.
4 min readRead →
Languages
What is DOMEX? Document and media exploitation explained
Translation is one of the four steps in the DoD definition of DOMEX, and DoD policy requires DOMEX capabilities to be maintained at all times.
4 min readRead →
Languages
What is FLORES-200?
Most vendor claims about low-resource translation quality rest on FLORES-200 scores, so a program office needs to know what the benchmark measures and what it leaves out.
5 min readRead →
Languages
What is sacreBLEU?
A translation score in a proposal is only evidence if the tokenizer, case handling, smoothing, and reference set are known, which is exactly what a sacreBLEU signature records.
5 min readRead →
Languages
What is SentencePiece?
The tokenizer decides how many pieces a Pashto or Burmese sentence becomes before a model ever sees it, which sets context limits, speed, and quality for the languages a program cares about.
6 min readRead →
Languages
Whisper speech recognition: accuracy, model sizes, and offline use
Whisper's word error rate was 3.0 percent in Spanish and 93.7 percent in Pashto on the same benchmark, so a single accuracy claim says little about a given language.
5 min readRead →
Languages
Why Arabic is hard for machines
One Arabic word appears in several written forms, so a query for one form can miss documents that hold another.
6 min readRead →
Acquisition
Army xTech prize competitions, explained
The Army reported in September 2026 that xTech|Search had awarded more than $14 million in cash prizes to more than 820 companies across its first 9 rounds.
4 min readRead →
Acquisition
CMMC levels 1, 2, and 3: what each level requires
A defense contract can require a current CMMC status at a stated level, and contracting officers are told not to award without one.
6 min readRead →
Acquisition
Commercial solutions openings, explained
A CSO lets DoD buy commercial products, commercial services, and nondevelopmental items through a general solicitation and expert review, with fixed-price awards treated as commercial.
5 min readRead →
Acquisition
CPARS: what it stands for, the 5 ratings, and how to respond
CPARS is the official source for past performance information, and agencies use its evaluations for up to 3 years after the evaluated work ends, or 6 years for construction and architect-engineer work.
6 min readRead →
Acquisition
DFARS 252.204-7012, explained
Nearly every Department of Defense contract that is not solely for commercial off-the-shelf items carries this clause, and a contractor that cannot meet it cannot hold the information the work requires.
6 min readRead →
Acquisition
DFARS 252.204-7021 CMMC contract clause
A missing CMMC status or affirmation can prevent award even when a contractor has completed its technical assessment.
6 min readRead →
Acquisition
Direct to Phase II and other SBIR paths, explained
Direct to Phase II skips Phase I entirely, so the feasibility proof has to come from work the firm did outside SBIR and STTR, and DoD solicitations reject proposals that cannot show it.
5 min readRead →
Acquisition
DoD colors of money: the appropriation types and how long each lasts
Charging the wrong appropriation can violate federal law. The PPBE Reform Commission reported 13 Antideficiency Act violations from FY2011 to FY2022 that used O&M money for software or system work.
5 min readRead →
Acquisition
DoDI 5000.73 and DoD cost estimating guides, explained
Under DoDI 5000.73, a milestone decision authority may not approve a major defense acquisition program entering a milestone phase until it has considered an independent cost estimate conducted or approved by CAPE.
7 min readRead →
Acquisition
Earned value management (EVM), explained
Contractor EVM data drive estimates of final cost, and GAO calculated in 2024 that the Columbia class lead submarine's likely overrun was more than six times the shipbuilder's own estimate.
5 min readRead →
Acquisition
Facility clearance (FCL): requirements and how to get one
Under 32 CFR 117.9, neither a contractor nor its employees may access classified information until the cognizant security agency has made an entity eligibility determination, such as issuing an FCL.
5 min readRead →
Acquisition
FAR 52.204-21 basic safeguarding requirements
Contractors must identify covered systems and meet the safeguarding requirements incorporated into their contracts.
5 min readRead →
Acquisition
FAR 52.204-25 and Section 889 explained
A contractor must examine both its offered equipment and its own covered use before making telecommunications representations.
5 min readRead →
Acquisition
FAR 52.219-14 limitations on subcontracting
A small-business prime needs payment and subcontractor records that support the applicable performance requirement.
5 min readRead →
Acquisition
Federal contract types: fixed-price, cost-reimbursement, T&M, and more
Since April 30, 2026, Executive Order 14402 has made fixed-price contracts the default, and a DoD non-fixed-price contract at or above $100 million needs the agency head's written approval.
6 min readRead →
Acquisition
How DoD SBIR topics are released, and when a proposer may contact the topic author
Under the DoD SBIR BAA, direct contact with a topic author is allowed only during pre-release; after that, questions go through a public Q&A that closes two weeks before the topic closes.
5 min readRead →
Acquisition
How to write a statement of work
A clear work statement gives the government and contractor the same definition of required performance.
5 min readRead →
Acquisition
MIL-STD-882E system safety and risk matrix
A traceable hazard record lets the responsible authority judge the remaining risk before people or equipment face it.
7 min readRead →
Acquisition
Organizational conflict of interest and FAR 9.5
Contractors and agencies need controls that address the actual conflict before relying on a mitigation plan.
5 min readRead →
Acquisition
Other transaction authority (OTA): what an OT is and how a small business wins one
Other transactions are outside the FAR, and DoD obligations through prototype and production OTs reached over $18 billion in fiscal year 2024, according to GAO.
5 min readRead →
Acquisition
PPBE and the POM: the POM cycle, FYDP vs POM, and DoDD 7045.14
PPBE decides which programs DoD asks Congress to fund, and a CRS primer says a cycle typically begins more than two years before the year the money is spent.
13 min readRead →
Acquisition
RDT&E budget activities 6.1 to 6.8, explained
The budget activity tells a reader what stage of work a program element funds. In the FY2027 request, operational system development alone is $132,750,742 thousand.
5 min readRead →
Acquisition
SAM.gov registration, UEI, and CAGE code: how to get all three
Offerors must have an active SAM registration when they submit an offer and at award, and the CAGE code is required before award.
5 min readRead →
Acquisition
SBIR commercialization benchmarks and the transition rate, explained
A firm that fails a benchmark loses Phase I eligibility for a year, and the numbers come from the commercialization report it files itself.
6 min readRead →
Acquisition
SBIR Phase I, II, and III, explained
Each SBIR phase has its own purpose, funding source, and rules, and Public Law 119-83 changed several of them in 2026.
5 min readRead →
Acquisition
SBIR Phase III: how the award works and what 15 U.S.C. 638(r)(4) says
Phase III is where SBIR-developed technology moves to follow-on work. The statute covers its funding, the sole-source preference, and data rights.
5 min readRead →
Acquisition
SBIR TABA (technical and business assistance), explained
TABA is the one pot of SBIR money that pays for commercialization help rather than research, and each agency sets its own request rules.
5 min readRead →
Acquisition
Sequential Phase II and Phase II enhancements, explained
A second Phase II, a Phase IIB, and an enhancement are all SBIR money with SBIR limits, while Phase III is not, and the labels decide what a firm may receive.
6 min readRead →
Acquisition
Should-cost and will-cost, explained
In its March 2016 assessment, GAO found 26 programs reporting $21.2 billion in realized should-cost savings, yet the programs could not account for where roughly a third of it went.
5 min readRead →
Acquisition
Small business set-aside programs, explained
The Small Business Act sets a governmentwide goal of at least 23 percent of prime contract dollars for small businesses, and SBA reports that agencies awarded nearly 28 percent in fiscal year 2025.
5 min readRead →
Acquisition
STTR vs SBIR: the differences in law and policy
SBA's directive allows no deviation from the STTR work split, because the statute sets it, while the SBIR split can be varied with written approval.
5 min readRead →
Acquisition
Technology readiness levels, explained
Before Milestone B, the milestone decision authority must confirm it considered whether a program's technology was demonstrated in a relevant environment, a test DoD's guidebook ties to TRL 6.
5 min readRead →
Acquisition
The Adaptive Acquisition Framework, explained
The pathway a program chooses sets its reviews, documents, and clocks; a middle tier program, for example, will not exceed 5 years after its start without a waiver.
5 min readRead →
Acquisition
The SBIR/STTR Policy Directive, explained
Every SBIR and STTR solicitation, award, and data rights clause a program office or small business signs is written to conform to this one SBA document.
5 min readRead →
Acquisition
The source selection decision: what the record must show
Since March 16, 2026, DoD contracting officers have used the overhauled FAR Part 15 under a class deviation, so the same documentation rules now sit under new section numbers.
5 min readRead →
Acquisition
What is a broad agency announcement?
Under 10 U.S.C. 3012, a research award chosen through a general solicitation and peer or scientific review counts as competitive, even though no two proposals answer the same statement of work.
5 min readRead →
Acquisition
What is a CDRL? DD Form 1423 explained
A contractor needs the full data requirement to price, prepare, and deliver an acceptable contract submission.
5 min readRead →
Acquisition
What is a trade study?
Before Milestone B, 10 U.S.C. 4252 has the milestone decision authority ensure that trade-offs among cost, schedule, technical feasibility, and performance have been made.
7 min readRead →
Acquisition
What is an Analysis of Alternatives?
An analysis of alternatives shapes which solution a major defense program pursues, and GAO found in 2009 that 10 of 32 major programs it reviewed had not conducted one.
5 min readRead →
Acquisition
What is an IDIQ contract? How IDIQ contracts and task orders work
Winning a multiple-award IDIQ contract guarantees only its stated minimum; work above that comes through orders you compete for against the other awardees.
7 min readRead →
Acquisition
What is an IGCE?
An IGCE is the Government's benchmark for judging offered prices, yet GAO found in 2017 that 49 of the 62 service-contract IGCEs it reviewed were not well documented.
5 min readRead →
Acquisition
What is SPRS?
A contracting officer checks SPRS before award, and a missing or expired NIST SP 800-171 score or CMMC status can stop a Department of Defense award to an otherwise qualified offeror.
6 min readRead →
Acquisition
What replaced JCIDS?
The January 2026 Joint Staff requirements manual put actual JCIDS timelines at about 300 days on average and set a nominal 55 business days for the process that replaced it.
5 min readRead →
