Methodology
Structured analytic techniques, applied to the documents you release.
FORCE runs the same eight stages over every engagement, from a released document set to a capability map. The pipeline was developed with a European defence prime, against a rubric an analyst had already applied by hand.
The pipeline
The eight stages of a FORCE analysis
One released document set goes in and a capability map comes out. Every stage writes its own output, and all eight ship with the assessment.
One analysis, in order
Eight stagesRetrieval
Documents are read from their text layer, and a page too thin to carry one is read by optical character recognition instead.
Screened for injected instructions
Output
One row per document
Passage extraction
Each document is swept once per combat function, and once more for strategic and cross-cutting material.
Verified verbatim against the source
Output
Passages, each with its character range
Combat function
Each passage is assigned one of six labels, from written definitions and thirty-eight written rules.
Output
One combat function per passage
Topic grouping
Passages are clustered inside each function, and any cluster carrying more than one concept is split before it is named.
Output
Named single-concept topics
Summary
Description, concrete examples and the reason it matters, so a finding can be read without opening the source.
Output
Four fields per passage
Capability or enabler
The ability a document argues for, or what has to exist for it — from nineteen written rules. The two never merge.
Output
One class per passage
Reporting
The merged result is written out, including a workbook in the review format a human grader marks up.
Output
Findings table, review workbook
Consolidation and ranking
Topics are ranked against each other and findings are consolidated into tiers from the agreed standpoint.
Output
Capability map, brief, dashboard
Provenance
Where the FORCE methodology comes from
The method was a manual one first, and it was automated rather than redesigned.
An analyst read a defence publication, marked the passages that mattered, assigned each to a combat function, separated capabilities from enablers, then compared every topic against every other one. The output was a spreadsheet.
FORCE built the pipeline against that method, together with a European defence prime, and against a set of documents the analyst had already labelled by hand. That hand-labelled set is still the reference a run is measured against.
How FORCE startedThe rule set
Each rule records one case where a run disagreed with the analyst, names which reading wins, and says what separates the two. The rules travel with the run.
Thirty-eight rules for the six combat functions
Nineteen for the capability-and-enabler distinction
A classification made today was made under a rule a person can read
Guards
Three checks that can stop a run
Three limits are enforced by the pipeline rather than by review, and each of them is about provenance rather than quality.
5%
Unverifiable passages
above that share, the run stops instead of delivering2,000
Characters per screened segment
four hundred of them overlapping the segment before10
Attempts per model call
then the run raises — never a partial answer
Stage 01
Screened before it is read
Released documents are treated as data, never as instructions. Contract and policy language passes, because it binds people and organisations. A clause addressed to an automated reader or to whoever summarises the paragraph stops the run — and garbled text from a poor scan is not treated as an attack.
Stage 02
Verified against the source
Every extracted passage is matched back against the document. A passage that does not match is broken into sentences, the sentences that do match are kept, and the remainder goes back for one more extraction pass over the region it sits closest to. What still fails is dropped and logged.
Every stage
No answer composed to fill a gap
A model call that will not return a usable answer is retried, and then the run fails. Nothing falls back to a partial result and nothing is substituted, because a finding invented to complete a table would be indistinguishable from one that was read.
Ordering
How the findings are ordered
Frequency is not the ranking. Topics are compared against each other in pairs, and the grounds recorded for each comparison ship with the result.
| Topics in the set | How the order is established | What ships with it |
|---|---|---|
| Under 80 | Every pair is judged | The complete duel matrix, and the reason recorded for each duel |
| 80 or more | Five rounds of a Swiss tournament, four pairings a round, then a Bradley–Terry strength estimate | The duels played, the strengths derived from them, and the basis recorded per topic |
In both paths the head of the ranking is not left to one judgement: the top thirty topics, plus a margin of ten, are completed into a full sub-matrix, and each missing pair there is judged three times with the majority taken.
Three weighted components decide the tier
The weights do not change between engagements, and only the first of them can be moved by the customer lens.
0.60
Importance
judged from the agreed standpoint0.25
Specificity
judged without the customer lens0.15
Novelty
judged without the customer lens
A score of 0.80 or above is Priority, 0.65 or above is Notable, and the rest are Mentioned — the three rows of the capability map. The score, its three components and the tier ship per finding, so a row that sits one tier lower than expected can be read rather than guessed at.
The capability mapStandpoint
What the standpoint changes, and what it cannot
A standpoint is one side and, optionally, one company: whose forces count as own, allied or adversary, and what the customer works on.
Detection and consolidation of capabilities and enablers run with no company in the prompt at all. Enrichment is split in two: a company-blind call returns the canonical name, the side, the specificity, the novelty and why a finding matters, and a second call returns the importance score and nothing else.
So two analyses of one set that differ only in the company differ in what is ranked highest, not in what was found. A disputed label is then an argument about the standpoint, which is productive, rather than about the method.
Why the standpoint is agreed firstFour sides
The standpoint is agreed before the run and shipped with the result, and it reaches only the calls that need it.
EU and NATO
Germany
Switzerland
A Russia-led adversary
Evaluation
What is measured after a run
Three checks run on a completed analysis, and each writes its own report.
Grouping
Scored twice
Once on how well separated the clusters are, and once by a judge that reads the passages in a cluster and answers, per passage, whether it belongs. The report carries the score and the individual verdicts, so a low score resolves to the passages that produced it.
Summaries
Judged field by field
Description, examples and reasons are each judged on their own rather than as one summary, so a weak examples field cannot hide behind a good description.
Labels
Compared by character span
A run is compared against the hand-labelled reference set by position rather than by text. A produced passage counts as covering a reference passage at thirty per cent overlap, and the combat function and the capability-or-enabler label are then compared separately.
And then a person reads it
Every run also writes a workbook in review format, laid out for a human grader to mark up. The final reading of a capability map is a person’s.
Method
Structured analytic techniques and a capability map
Structured analytic techniques are procedures that externalise an analyst's reasoning so another reader can inspect it. The CIA tradecraft primer and the successor literature describe several dozen of them, and what they share is a fixed procedure applied the same way each time rather than a judgement formed privately.
A FORCE capability map belongs to that family. Intelligence analysis methods differ in what they are for; this one is for reading a released document set against a fixed taxonomy.
What makes it inspectable
The frame is external to the documents and does not adapt to them
Six combat functions and two item classes, fixed before a document arrives
Every entry carries the passage it was read from
An empty cell records silence rather than absence
The output of all eight stages is delivered with the result
Limits
What a FORCE analysis does not establish
Four things sit outside the method, and stating them is part of it.
FORCE gives no recommendations for action. The frame orders the evidence; the judgement stays with the reader.
Nothing is topped up from outside the released set, and no gap is filled by inference. An assessment is a statement about your document set, not about the world.
An empty cell records that the set is silent on a combat function, not that a capability is absent.
A passage is reported as what a document says, not as a fact about what exists. No figure appears in a result that is not in a source document.
The rule the rest of the method serves
The evaluation of the results and the strategic decision-making always remain with the customer.
Sources
The doctrine and the literature this rests on
The combat functions follow AJP-01 Allied Joint Doctrine (NATO, 2022), and the capability vocabulary follows the NATO Defence Planning Process (2024). Structured analytic techniques are catalogued in A Tradecraft Primer (US Central Intelligence Agency, 2009). The paired-comparison ranking follows Bradley and Terry, Rank Analysis of Incomplete Block Designs (Biometrika, 1952); the grouping uses UMAP (McInnes, Healy and Melville, 2018) and HDBSCAN (Campello, Moulavi and Sander, 2013). The frames this method records but does not rank are set out under Frames, and the register behind the Doctrine Explorer holds 6,128 publications from 33 countries and NATO.
One step
Send us the question, not the documents.
Describe the decision that has to be taken and roughly what you could release. Within two working days you get an assessment of what an analysis could and could not establish.