The Alpha Plan, Explorable

Five spikes falsify the riskiest assumptions before any code, each with pass criteria frozen in advance:

• Context-draft review: will beta customers correct an AI-drafted playbook after seeing sample results?
• STT and readiness: does a real clip return insights in under 20 seconds, with accuracy floors per noise slice? Now also tests phone voice-isolation modes and background-voice suppression.
• Native auth round trip: the mobile sign-in flow end to end, including the attendee upgrade path.
• Attendee QR field test: will strangers at a busy booth scan and consent in under 25 seconds?
• Legal feasibility: voice enrollment and remote-call capture pressure-tested per pilot jurisdiction.

The smallest complete learning system in trusted clients' hands: platform core, lead intelligence with context, capture on Android and iOS, attendee QR join and sharing, base and layered insights, human review, the feedback loop, a staff admin seed, and exactly two reports: the rep's daily brief and the exhibitor's post-event wrap. Both report templates are mandatory by founder decision.

One real event with counsel-approved consent configuration, event-day readiness drills, a reviewed Lead Loss Rate, and output the customer accepts. The post-event wrap doubles as the pilot's acceptance artifact.

Prove value after the show: follow-up meeting intelligence, lead export (CSV and structured), CRM handoff, the weekly pipeline digest.

Capture extends to remote meetings: a visible bot joins Meet, Zoom, or Teams calls behind its own consent flow, after the vendor bake-off and legal clearance.

The complete sellable product: themes and ROI, full staff admin, billing, the commercial trust pack, public app stores, dynamic ad-hoc reporting, and the organizer outcome report. GTM leads with organizers as the primary client, bundling the app for their exhibitors.

Density- and demand-gated: organizer renewal forecasting and at-risk views, Company and Solo modes, seat billing, live streaming, Arabic market launch, custom report composition, marketplace work.

One table holds every material decision with a status: locked adjudicated with reasons recorded, default a working assumption not yet contested, open genuinely undecided. Every decision carries the condition that would reopen it. Anything can be challenged, but only with evidence of what changed.

Substantial decisions get an Architecture Decision Record explaining the alternatives and why one won. Recent ones: ADR-016 restored opt-in voice self-enrollment with nine controls; ADR-017 set the two-milestone structure; ADR-018 made billing provider-neutral with a first Western gateway; ADR-019 scoped remote-call capture; ADR-020 named the product Alpha.

When two documents disagree on a topic, docs/README.md names the winner. Everything a losing document says on that topic is history, not instruction. This is what caught and corrected the voice-enrollment flip-flop.

A founder decision becomes a dated directive note in docs/plans/. The planning agent executes it: register rows, ADRs, and reconciliation of every affected document, then marks the note executed. Nine directives have run this loop so far, from voice enrollment through the reporting module.

Sign-in (passwordless, redesigned from V2's passwords), invitations, join codes, notifications, profile and security, privacy center (new), offline status (new), accessibility settings.

Event and booth setup, capture-policy configuration, live dashboard, pipeline, conversation feed, topic insights, review queue (new), knowledge sources (new), playbook draft review, billing, export.

The capture loop: readiness check, consent and start, active capture, attendee QR status (new), save or discard, review, follow-ups, private self-coaching, end-of-day, capture recovery (new).

Almost entirely new: the wireframes never drew the person being recorded. QR join, consent decisions, live status, summary claim, the participant's own summary, privacy and deletion.

All redesigned around the privacy boundary: anonymised aggregates with minimum thresholds, no named exhibitor negatives, explicit not-enough-data states. Your organizer-interests document will be reconciled here.

Our own back office, all new: support, pipeline operations, consent exceptions, cost reconciliation, billing operations, recovery jobs, incident posture.

Eight independent AI reviewers each read the entire plan through one professional lens: architecture, legal, product, economics, security, operations, go-to-market, and AI quality. Each returned ranked findings citing exact documents. They found 44 issues, 7 critical, including gaps no single reader would catch, like the missing plan for actually booking the pilot event.

Three more agents argued the big go/no-go question from fixed stances: a Skeptic (challenge the premise), a Pragmatist (what ships), and a Critic (what breaks). Their unanimous push-back changed the plan: run the cheap evidence spikes before writing code, and restage scope across milestones. The founders then amended the verdict: stage everything, cut nothing.

Every finding got one of three dispositions, recorded in a reconciliation document: a fix, a decision-register row, or a written rejection with reasons. Nothing was silently dropped. The same discipline now applies to every review round.