The beta. You break it, we owe you.
Testers get the desk early, and everything is free to anyone who signs up while the beta runs — during the beta only. The paid plans continue after the beta ends (see pricing). In return: your findings — where the analysis read you wrong, where a document confused you, where a tag did not fit. The comprehension rule applies to testers first: if you did not understand a screen, that is our defect, not yours.
The walk-through desk keeps everything on your device. The real analysis sends what you write about your case to our AI providers to produce your result — the privacy notice names each one. The honest limits: the weighing engine, real document delivery and support SLAs land with the beta — this page states intent, not a promise of dates.
All four stages unlocked
While the beta runs, signed-in testers unlock Stage 01 (Analyse), Stage 02 (Resolve & Evidence Map), Stage 03 (Claim & Pre-Action Documents), and Stage 04 (ACAS & Hearing Preparation) without charge. When the beta ends, normal one-off stage pricing resumes.
Real disputes or stress-test scenarios
Run a live consumer, employment or housing dispute you are actually dealing with, or throw complex multi-party facts at the desk to test where statutory pinpointing, time-limit warnings or letter drafts need sharpening.
Every finding logged against the engine
When you flag a missing statutory hook, an over-eager claim head or an unclear paragraph, we turn that finding into a deterministic regression case in our verification harness so the same defect cannot quietly return.
We do not publish vague assurances about AI capability. Every change to the legal engine is measured against real England & Wales tribunal decisions and stress-test corpora before it ships. Here is the progression that brought the desk to its current beta build:
- Phase 01 · Deterministic Statute Book & Precondition Gates. We separated creative language generation from statutory verification. Every cited Act and section is checked against our England & Wales statute index, and structural preconditions — such as the two-year qualifying service threshold under ERA 1996 s.108 versus Day-1 automatic unfair dismissal exceptions — are enforced deterministically.
- Phase 02 · Pre-Output Deliberate IRAC Reasoning (
qwen.qwen3-next-80b-a3b). We upgraded the primary analysis, document drafting and three-round letter campaign engines to perform structured Issue–Rule–Application–Conclusion (IRAC) verification before emitting a single claim head. The engine scans every party, contract layer and statutory threshold (including CCA 1974 s.75 £100–£30,000 limits, s.140A unfair relationship fallbacks, CRA 2015 s.49 service standards, and Housing Act 2004 deposit rules). - Phase 03 · Standing Preservation, Multi-Respondent Joinder & April 2026 Bands. Where a regulator such as the ICO or HSE investigates systemically rather than awarding individual compensation, the desk now explicitly pairs the regulatory report with the direct County Court or Employment Tribunal route that preserves your personal remedy. Where an employer and a client site or agency split responsibility, it surfaces Employment Tribunal Rule 34 (Selkent) joinder options. Discrimination injury-to-feelings bands are locked to the April 2026 presidential Vento guidance.
Curated Benchmark Suite
31 of 32 cases passed all positive and negative constraints simultaneously, recalling 70 of 71 required statutory groups (98.6% group recall) and 31 of 34 governing Acts (91.2% Act recall) with 0 of 35 forbidden-statute violations — including complex five-issue agency worker, disability, whistleblowing and data protection disputes.
Category-Trap Suite
29 of 29 synthetic trap cases passed every constraint, recalling 83 of 83 required statutory groups (100.0% group recall) with 0 of 60 negative-constraint violations across B2B-vs-consumer boundaries, credit card cash-advance traps, tenancy deposit caps, and short-service dismissal rules.
Employment Tribunal Corpus
Tested against 238 real Employment Tribunal decisions: the engine recalled 544 of 626 tribunal-cited statutory groups (86.9% group recall) and surfaced 481 of 528 tribunal-cited Acts (91.1% Act recall across user-facing output; 74.1% in primary detected heads alone), satisfying every cited group simultaneously in 168 of 238 cases.
Housing Tribunal Corpus
Tested against 101 real First-tier Tribunal (Property Chamber) decisions: the engine recalled 232 of 270 tribunal-cited statutory groups (85.9% group recall) and surfaced 222 of 255 tribunal-cited Acts (87.1% Act recall across user-facing output; 64.7% in primary detected heads), passing all groups simultaneously in 73 of 101 cases.
Six-question scorecard on 104 official tribunal decisions with the judge’s verdict stripped (tests/verdict-cases-2026-09-28.json):
- Q1 · Right statutory instruments surfaced: 195 of 218 tribunal-cited instruments named across primary claims and options to pursue (89.4% instrument recall; 158 of 218 = 72.5% in primary heads alone), with 26 of 58 (44.8%) exact section matches.
- Q2 & Q3 · Remedy heads & April 2026 Vento bands: 18 of 18 real-award cases received the matching remedy head; 11 of 13 successful discrimination cases surfaced the injury-to-feelings head; 101 of 101 valuations matched the April 2026 presidential guidance with 0 stale band warnings.
- Q4, Q5 & Q6 · Verbatim fact grounding & why-not reasons: Across 148 asserted rulebook claims covering 383 statutory elements, 383 of 383 (100.0%) carried a verbatim quote from the claimant’s own text. Across 824 candidate provisions that were inspected and not asserted, 824 of 824 (100.0%) carried an explicit stated reason why.
Stratified synthetic provision-seeded holdout (140-case stratified subset of tests/complex-cases-holdout-2026-08-24.json — 38 consumer, 51 employment, 51 housing, where 92.5% of target provisions are absent from prompts by construction):
Across all 140 stratified holdout cases (222 seeded statutory provisions), the engine surfaced the target Act in 160 of 222 instances (72.1% overall Act recall): 63 of 70 (90.0%) in employment, 64 of 84 (76.2%) in consumer, and 33 of 68 (48.5%) in specialist housing schedules.
Identical-input repeatability (k=3 and k=5 live handler replays across the 32 curated accuracy cases, tests/live-replay-qwen-k5-2026-11-10.json):
Replaying the exact same case inputs 3 times (k=3, 96 runs) and 5 times (k=5, 160 runs) against the live handler produced a 96.9% to 100.0% case pass-rate spread (mean 99.0% at k=3 and 99.4% at k=5, with 4 of the 5 full replicates scoring 32 of 32 = 100.0%). Across all 5 identical-input runs, 31 of 32 cases (96.9%) returned the exact same pass/fail verdict and the exact same mustInclude statutory group hits (99.8% pairwise Jaccard at k=3; 99.9% at k=5; 354 of 355 group checks passed; 0 of 175 mustExclude violations), while 24 of 32 cases (75.0%) returned the exact same assertedClaims set across all 5 runs (85.5% pairwise Jaccard).
Where we want testers to push hardest: Identifying the governing Act and grounding claim elements in verbatim quotes are now strong across all three domains, while exact sub-section pinpointing on specialist housing schedules (such as Leasehold Reform and Building Safety sub-provisions) and procedural time-bar edge cases on historic accounts are our active engineering frontier. If you have a case in those lanes, we especially want you in the beta.