The WPS AI Agent
A purpose-built system, scored on the same benchmark as the general-purpose models on the leaderboard, as a reference point for what a system built specifically for WPS analysis can achieve.
Try the live demo ↗What it's built to fix
General-purpose AI models tend to fail the same way on WPS content: they perform well when a prompt spells out the gender dimensions, and lose most of that performance on the terse, realistic requests practitioners actually send. Where they do respond, they often state claims with more confidence than the evidence supports, and rarely flag what they don't know. The WPS AI Agent is designed directly against those two failure modes — weak WPS reasoning without explicit cues, and unverified, overconfident claims.
How it works, conceptually
- Understanding the request. Works out who is asking, for what decision and where, and whether carrying out the request as framed would undermine WPS objectives.
- Searching the documents. Searches national action plans and policy instruments on one side and WPS research on the other, always including both gender-focused and operational searches, and setting aside other countries’ plans.
- Writing. One decision-ready answer, with each claim drawn from a document tied to the passage it comes from.
- Checking. Every figure and citation is checked against its source, and a separate checking model reads the answer for ranked recommendations with owners and risks, and for the analysis of gender norms and trust.
- One fix. If the check finds problems, exact edits fix them once. Anything still unresolved is listed at the end of the answer.
v4 splits this work across models. GPT-6 Luna, a fast, low-cost model, plans the search, and GPT-5.6 Terra checks the draft; both are the same in every v4 build. The writing model, which drafts the answer and makes the one fix, is the model each build below is named for. In the live demo that is GPT-6 Luna as well. Earlier versions ran a single model for every stage.
How the answer is checked
Checking is the piece that most directly answers the "unverified claims" failure mode. Code checks that every figure and citation in the answer traces to the passage it came from, and a separate checking model, GPT-5.6 Terra, reads the answer against those passages. The Sources list is built by code from the citations the answer uses. Since v2 introduced verification, factual integrity and acknowledgment of uncertainty have gone from the agent's weakest scores to among its strongest, ahead of every general-purpose model tested. The agent does not use internet search.
Where it stands today
Version 4 of the agent scores above every model on the leaderboard. With GPT-5.6 Sol as the writer, it scores 0.87, leads every general-purpose model tested on all eleven criteria, and scores 0.75 or higher on each. The margins are widest on the two failure modes this agent was built against. On factual integrity it scores 0.91 against 0.56 for Claude Opus 5, the strongest general-purpose model both overall and on this measure; on gender-sensitivity and WPS integration, 0.89 against 0.57. The live demo build, with GPT-6 Luna as the writer, scores 0.86 and answers in under a minute; it leads every general-purpose model on ten of the eleven criteria, and trails Qwen3.8 by 0.005 on WPS guardrails, where most systems score near the ceiling. On 19 November 2026 the demo moved to an updated document collection: each plan's years and phase are now taken from its own text, every language edition of a plan is kept, and documents that are not adopted plans were moved out. Re-scored on that collection, it scored 0.858, against 0.857 for its earlier answers re-judged in the same sitting, so the update did not change its score. The re-scored build also included a small fix to how sources are listed.
Against v2 on GPT-5.6 Sol, v4's largest gains are trust building (+0.14), due diligence (+0.09) and policy alignment (+0.08). Trust building is still v4's lowest score, at 0.75, and analytical depth, at 0.78, has the most room to improve. The full breakdown for both builds is under Benchmark results below.
This is a demonstration of what purpose-built scaffolding can add to WPS analysis. Results should always be read alongside a human WPS advisor, not in place of one.
Benchmark results
The versions of the agent available now, scored on the same 30 scenarios, eleven criteria and three judges as the models on the leaderboard. The GPT-6 Luna build runs the live demo, and the GPT-5.6 Sol build is available on request. Every version since v2 has outperformed Claude Opus 5, the highest-scoring general-purpose model; earlier versions are in the version history below.
| System | Evaluated | Overall | Details |
|---|---|---|---|
|
{{ row.provider }}
{{ row.description }}
|
{{ row.date }} |
{{ row.overall }}
|
|
|
{{ row.note }}
{{ row.criteriaHeading }}
{{ dc.name }}
{{ dc.display }}
Gender / WPS integration by prompt tier
{{ gt.label }}
{{ gt.display }}
{{ gt.partial }}
The fall from cued to sparse prompts is the WPS Competence Gap — how much of a system's WPS performance depends on being told to look for it. See The Benchmark.
Performance on adversarial criteria
{{ ar.label }}
{{ ar.display }}
These two criteria are conditional — they pass by default on benign requests, so they are scored on the 3 adversarial scenarios only. See The Benchmark.
|
|||
Static copy of the agent's benchmark results for readers without JavaScript.
| System | Provider | Evaluated | Overall | Gender-sensitivity on detailed prompts that name gender (T1) | Gender-sensitivity on sparse prompts with no gender cue (T3) | Gender-sensitivity / WPS Integration | Operational Relevance | Analytical Depth | Policy Alignment | Clarity / Actionability | Trust Building | Due Diligence | Acknowledgment of Limitations | Factual Integrity | WPS Guardrails | Substantive Pushback |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| WPS Agent v4 (GPT-5.6 Sol writer) | Our Secure Future | 6 November 2026 | 0.87 | 0.86 | 0.92 | 0.89 | 0.88 | 0.78 | 0.78 | 0.98 | 0.75 | 0.93 | 0.85 | 0.91 | 0.95 | 0.91 |
| WPS Agent v4 (GPT-6 Luna writer) | Our Secure Future | 6 November 2026 | 0.86 | 0.87 | 0.82 | 0.85 | 0.87 | 0.72 | 0.76 | 0.98 | 0.73 | 0.94 | 0.86 | 0.91 | 0.93 | 0.91 |
| System | T1 — Context-Rich | T1-neutral — Cue-Stripped Pairs | T2 — Gender-Neutral | T3 — Sparse / Ambiguous |
|---|---|---|---|---|
| WPS Agent v4 (GPT-5.6 Sol writer) | 0.86 | 0.89 | 0.90 | 0.92 |
| WPS Agent v4 (GPT-6 Luna writer) | 0.87 | 0.85 | 0.88 | 0.82 |
- WPS Agent v4 (GPT-5.6 Sol writer) (Our Secure Future, overall 0.87): Current version. GPT-5.6 Sol writes and fixes the answer; GPT-6 Luna plans it and GPT-5.6 Terra checks it Scores 0.75 or higher on all eleven criteria and holds gender-sensitivity at 0.86–0.92 across tiers, highest on sparse prompts; trust building (0.75) and analytical depth (0.78) are its lowest criteria.
- WPS Agent v4 (GPT-6 Luna writer) (Our Secure Future, overall 0.86): Current version, running the live demo. GPT-6 Luna plans, writes and fixes the answer; GPT-5.6 Terra checks it Scores 0.72 or higher on all eleven criteria and answers in under a minute; sparse-tier gender-sensitivity (0.82) and analytical depth (0.72) trail the GPT-5.6 Sol build.
What the agent adds, model by model
GPT-5.6 Sol is the one model tested all three ways: on its own, running every stage of v2, and as v4's writer. That makes it the clearest view of what the agent adds.
| Model | Bare | v2 | v4 | Gain over the model alone |
|---|---|---|---|---|
| GPT-4o | 0.334 | 0.712 | — | v2: +0.378 |
| GPT-5.6 Sol | 0.601 | 0.839 | 0.873 | v2: +0.238 · v4: +0.272 |
| GPT-6.1 Sol | 0.603 | — | 0.868 | v4: +0.265 |
| GPT-6 Luna | not tested | — | 0.860 | — |
The two gains measure slightly different things. In v2 one model runs every stage, so its gain comes from the workflow alone. In v4 the gain also includes the two models every v4 build shares, GPT-6 Luna planning and GPT-5.6 Terra checking, so it measures the whole system rather than the workflow by itself.
On its own, GPT-5.6 Sol scores below 0.60 on six of the eleven criteria, failing in the ways the Benchmark describes. Inside v2, it scores above 0.60 on every criterion. Verification lifted factual integrity and acknowledgment of limitations from its weakest scores, 0.24 and 0.40, to 0.91 and 0.84, and the gap between cued and sparse prompts closed: it brings a gender lens to terse requests as reliably as to detailed ones.
v4 keeps those gains and adds to them, mostly in trust building, due diligence and policy alignment. Those are the parts of an answer that say who should act, what could go wrong and which commitments apply. Overall, GPT-5.6 Sol scores 0.60 on its own, 0.84 running v2 and 0.87 as v4's writer.
The other models fit the same pattern. GPT-6.1 Sol, a newer model, scores no better than GPT-5.6 Sol on its own, and gains about as much as v4's writer. v2 lifted the older GPT-4o even further, far enough to outscore the best Sol model tested on its own, GPT-6.1 Sol. GPT-6 Luna, which runs the demo, hasn't been tested on its own.
Version history
Dates are when each build was made. v1 and v2 were scored in July–August 2026, and v4 in November 2026, on identical scenarios and criteria and by the same three judges. Re-judging v2's published answers alongside v4 moved its score from 0.839 to 0.841 and 0.846, so the builds compare fairly, give or take about 0.01. Scores are overall benchmark averages.
Interested in a system built for your organization's WPS needs?
Get in touch — we work with organizations that need WPS-aligned analysis at scale.
Get in touch