Pellucid is a neurosymbolic reading instrument for high-stakes prose — software requirements, contracts, compliance policies. It detects vague terms, structural ambiguity, and untestable claims, then defends every call with a transparent panel of four specialists.
The system shall be user-friendly and shall respond quickly under all reasonable conditions. The interface should minimize user frustration by surfacing diagnostic information as needed.
Five ambiguities in twenty-eight words. Pellucid finds them in about two hundred milliseconds — and will defend each call.
The defect that travels furthest, costs the most, and looks the most innocent on first read.
In regulated engineering, the rework that breaks a programme rarely traces back to a faulty algorithm. It traces back to a sentence — a clause in a specification that two engineers, reading it on different days, took to mean different things. By the time anyone notices, the misreading has cascaded through design reviews, vendor quotes, test plans, and shipped code.
Industry studies of requirements defects consistently land in the same uncomfortable neighbourhood: roughly forty per cent of project rework in safety-regulated industries traces to ambiguous, incomplete, or untestable requirements written years earlier. The cost is not in the typing. The cost is in the compounding.
A spec that admits two readings will, at scale, produce two products.
Existing tools — Jama, DOORS, Polarion — manage requirements, but they do not read them. A vague-term checklist is not a reader. A linter is not a peer review. Pellucid is built to be the reader your specification never had: fast where speed is cheap, slow and deliberative where the stakes warrant it, and always willing to show its working.
Pellucid is not a single model with a confident answer. It is a fast classical reader that runs first, a panel of four specialists who deliberate next, and a calibrated aggregator that reconciles them.
The first reader is mechanical and fast. It carries a small dictionary of dangerous words and a parser that knows what a passive voice and an open-ended modal look like. It will not catch every kind of trouble — but the kinds it does catch, it catches in the time it takes to render the page.
Confidence is the most-abused number in machine learning. We treat it as an obligation. Every score is calibrated against a held-out set so a 0.7 means seventy per cent — not a model’s vibe. Every flagged span ships with the agent votes that produced it, the rule that triggered first, and a single line explaining what would resolve it.
DO-178C-aware lexicon. Catches “shall” drift across requirement levels.
IEC 62304 traceability. Flags ambiguity that breaks risk-based testing.
MIL-STD-498 conventions. Detects vague coverage and unbounded scope.
ISO 26262 awareness. Catches ASIL implications of soft requirements.
Reads model-risk policies, control language, and audit-grade contracts.
SOC 2, ISO 27001, internal-control wording reviewed for testability.
We charge for the cost of running the panel and curating the calibration set. Everything else — the rule layer, the rewrite flow, the audit trail — is bundled.
For an individual writer making sense of a single specification. Bring your own model key; rule layer and panel both included.
Open the editorFor a working group on a shared corpus. Custom domain ontologies, version-aware reviews, audit trails, and organisational calibration.
Start a teamFor a regulated programme. Self-hosted or VPC, custom-trained domain panels, signed audit exports, and on-call review.
Talk to usOpen the editor. Paste a paragraph. Read a sentence the way Pellucid does.