Selected case study / Project communication

Communication and Specification Assistant

A governed path from project conversation to source-linked draft requirements.

Role-tagged messages enter an immutable conversation record. Documented language forms become versioned requirements, conflicts stay visible, and only authorized approvals can produce cited draft clauses.

Open source Read evaluation

Actual local interface / synthetic conversation data

Problem

Project decisions are easily separated from their source message, current status, conflict history, and approval authority.

System

Immutable message IDs, deterministic extraction, requirement lifecycle, conflict detection, role gates, SQLite audit events, and cited draft clauses.

Evidence

Five synthetic conversations with 35 messages, plus 33 manually labeled language stress cases that retain two known misses.

Boundary

Documented phrase coverage only. No open-domain chat, authenticated identity, professional specification, or stakeholder acceptance evidence.

System

Approval changes state; conversation alone does not

The ledger separates what was said, what was extracted, what conflicts, and what has been approved. A clause is generated only from an active requirement with a permitted approval event and a source-message reference.

Communication and specification system map showing role-tagged messages, extraction, requirement state, conflict handling, approval gates, audit events, and source-linked draft clauses — Generated from the workflow regression and language stress artifacts. Supplied roles, deterministic language coverage, and mandatory professional review remain visible boundaries.

Measured behavior

The stress set exposes where documented language coverage stops

Perfect workflow-fixture results show deterministic regression consistency. The separate language set is the more useful boundary check because it includes paraphrases, negative controls, and retained failures.

Requirement F10.958Across 33 manually labeled stress cases

Requirement recall0.920Two number-word requirements remain missed

Exact-case accuracy0.939Direct, paraphrase, and negative-control cases

Negative-control accuracy1.000Questions, history, rejection, and ignored references

Exact-case accuracy comparison for direct language, paraphrases, and negative controls with two retained misses — Manually labeled coverage comparison. The paraphrase group retains failures rather than hiding them behind the aggregate score.

Local project communication interface with role-tagged messages, requirement ledger, conflicts, approvals, and source-linked draft clauses — Actual local interface using bundled synthetic roles, messages, and project requirements.

Engineering decisions

Governance is modeled as data, not prompt wording

The implementation keeps provenance and state transitions inspectable instead of asking a language model to imply authority or silently reconcile conflicting instructions.

Immutable message identity

Every extracted value and draft clause points back to a stable message ID, sequence, role, and timestamp.

Versioned requirement state

Proposed, approved, superseded, and conflicting requirements remain distinct records rather than one mutable summary.

Fail-closed approval gates

Category-specific role permissions reject unauthorized approvals and record the denial without changing the clause set.

Auditable abstention

Questions, rejected values, historical references, unsupported chatter, and unresolved conflicts remain outside the approved draft.

Retained failure

Number words fall outside the current extraction grammar.

The fixed stress set misses "four thousand two hundred square metres" and "a dozen consultation rooms." These failures are checked in because the project does not claim open-domain understanding or silently infer unsupported values.

Inspect retained failures

Run locally

Replay the workflow and language audit

The evaluator rebuilds the workflow regression, source-linked trace, stress metrics, retained-failure report, and visual comparison from bundled synthetic fixtures.

python projects/project-specification-copilot/evaluate_specification.py python -m pytest tests/test_project_specification_copilot.py streamlit run projects/project-specification-copilot/app.py