🤓 Latest Submissions
ERRATA
Transit controllers constantly correct themselves while handling radio, maps, incidents and live service pressure. In a conventional voice workflow, each new utterance can become another instruction which creates a dangerous failure mode when a correction conflicts with what was said seconds earlier. ERRATA is built around a stricter rule: speech is a draft; operational truth is versioned. A controller can say: “Route 55 west, skip King Edward and Cumberland until 9:30.” AssemblyAI Universal-3.5 Pro Realtime transcribes the live microphone stream. ERRATA interprets the instruction as a non-mutating draft, shows the controller exactly what would change, and waits for explicit human Apply. If the controller then says: “Wait — garde Cumberland. Make it 10.” ERRATA does not create a second instruction. It repairs the same change_id, advancing the existing service change from one revision to the next while preserving a single canonical operational state. Incomplete speech is handled safely as well. A phrase such as “Wait, keep Cumberland. Make it.” produces NEEDS_CLARIFICATION: Apply remains unavailable, the revision does not advance and the canonical hash stays unchanged. Final authority is also version-bound. A stale reviewed hash is refused; only the current reviewed state can be committed. ERRATA then serializes the canonical change into GTFS-Realtime and independently decodes the resulting protobuf so the downstream operational consequence is visible and testable. AssemblyAI is load-bearing in the live product path. AI33 provides generated voice guidance to the operator, while deterministic Python logic owns parsing, validation, revisioning, evidence generation and commit authority. ERRATA is not a passenger chatbot. It is a control layer for high-consequence voice operations where fast, fallible speech must never silently become operational truth.
30 Sep 2026
COLLIDER: Semantic CI for AI Agents
Parallel AI coding agents can each be locally correct while silently making incompatible decisions about requirements the shared specification never resolved. Those failures may survive code review, Git merge, and ordinary tests because the problem is semantic, not syntactic. COLLIDER is Semantic CI for agentic software development. It treats independent agent implementations as ambiguity probes. When workstreams disagree, COLLIDER checks the authoritative source and separates two cases: AGENT_DRIFT, where the specification already chose the answer and the implementation should be repaired from evidence; and SPEC_GAP / UNKNOWN, where multiple implementations are plausible because nobody actually decided. In the canonical failed-payment scenario, 30 local tests pass and all three workstreams are green, yet integration reveals two conflicts. One is automatically repairable from source evidence. The other—customer identity, email vs account_id—is genuinely unspecified. COLLIDER asks one bounded human question instead of inventing a choice. Once the human selects account_id, COLLIDER compiles that decision into a specification patch, targeted code repair, regression contract, and executable decision memory. Integration moves from 2 conflicts to 0. That memory then becomes a Semantic CI guard for future changes with three verdicts: MERGE_ALLOWED, MERGE_BLOCKED, or DECISION_REQUIRED. The signature behavior is intentionally different from conventional CI: a future cents-to-dollars change can pass 37/37 tests while COLLIDER still returns DECISION_REQUIRED because the source is silent and the workstreams now disagree. Tests can say pass or fail; COLLIDER adds a third answer: nobody decided this yet. Target users are engineering teams building with AI coding agents, especially teams running parallel agent workstreams. They interact through the public demo or the GitHub pull-request gate, review the detected semantic conflict, make the minimum required human decision, and let COLLIDER compile and preserve that decision for future work. COLLIDER is designed around provenance and abstention. It does not relabel seeded fixtures as live agent output, does not force a decision when evidence is missing, and preserves real failures as evidence. A real IBM Bob run with three isolated subagents reproduced the customer-identity specification gap, and a later targeted fresh-agent replay passed the withheld regression contract 7/7.
27 Sep 2026
