The arena map is a qualitative product comparison, not a completed same-task benchmark. Compare Paciva with Copilot, Fyxer/Lindy and existing email/calendar workflows on screening, cross-tool completion and approval/recovery. Record actual completion, corrections, time and cost on the same inputs. No paired results are published in this brief; missing detail is not evidence that a competitor lacks a capability.

They can compete. Microsoft already connects Copilot to services including Gmail; Google supports personalized spam preferences; OpenAI and Anthropic sell applications with memory. Their distribution is a real advantage. Paciva’s hypothesis is that continuity across tools, user-controlled execution and accumulated feedback earn preference. The next beta must test that against existing alternatives.

If a tracked rival ships into the wedge mid-raise, the pre-drafted response states which arena cells flip, and which do not: per-user calibration lag, Gatekeeper's adversarial depth, and the three-rung economics.

The intended advantages are cross-tool context, reliable integrations, user-controlled actions, useful local execution and personalization from explicit feedback. None establishes exclusive capability or a guaranteed switching cost. A hypothetical thousand users with fifty thousand messages each is a scale scenario, not an existing corpus. Adoption and retention must establish whether these mechanisms compound in practice.

Notion announced that its Mail inbox will close on September 22, 2026. That is a product decision, not proof that inbox interfaces or email agents are dead. Paciva must demonstrate its own adoption: an earlier beta had limited usage, and another cohort is underway. The relevant test is whether users repeatedly delegate useful work.

Category reports and seat arithmetic describe different scopes and must not be nested as TAM/SAM/SOM. The disclosed occupational counts are 11.13M management plus 13.42M sales roles, totaling 24.55M. At the proposed $3,600 annual founding subscription, that produces an illustrative $88.38B ceiling, not validated demand. It does not enumerate the whole knowledge-worker/developer buyer set. The 50K to 200K seat illustration yields $180M to $720M annually; it is not a 3 to 5 year acquisition forecast. Pricing, conversion, retention and support economics still need validation.

Routing work to code or eligible local models can reduce cloud costs. Cloud inference and multi-provider failover still incur variable costs, and metered credits do not guarantee margin. The next cohort must measure successful workflows, retries, local/cloud mix, utilization, and onboarding/support costs. P50/P95 costs and realized gross margin are not published here. The $2,500 full-month M5 MacBook Pro capacity estimate is a founder value assumption, not Paciva revenue, hardware purchase price or a measured cost saving.

Public primitives, one chain: 1.5 hours a day of graymail triage, times 235 working days, times $36.30 an hour from BLS series. That is $12,796 per employee per year, or $3.2M a year at 250 people. Current BLS prints, $37.53 average hourly earnings and $46.60 fully loaded, sit above the rate in the chain, which makes the figure conservative and we say so. Volume context from Radicati: 361.6B emails a day in 2024 growing to 424.2B by 2028. The no-double-count rule applies: graymail hours, app-toggle hours, and meeting waste overlap, so each stat argues its own slide and they are never summed.

The first buyers are early-adopter knowledge workers and developers. The new beta tests which recurring workflows activate them, how often they return, and whether they convert to paid use. An earlier product variant saw limited usage in its first beta. The disclosed management/sales counts sum to 24.55M; those occupations are a comparison frame, not the full buyer definition or evidence of willingness to pay. Pricing has not been tested.

Pricing is untested. The proposed ladder is a capped free-subscription Founding Pass, $300/month founding seats grandfathered for life, and $3,000/month list for later individual cohorts; inference is metered separately. Grandfathered seats do not migrate to list.

The founder estimate values one M5 MacBook Pro at full utilization for a month at $2,500. An illustrative 24 hours saved at $36.30/hour adds $871.20, for $3,371.20 gross monthly value only when the benefits are distinct. Useful workload, utilization, hardware availability and additional charges change the customer outcome. This is the list-price rationale, not tested willingness to pay or guaranteed savings.

By mid-August 2026, OpenAI’s median researcher ranked by agent usage consumed more than $600/day of inference valued at API prices. OpenAI research, September 6, 2026. This internal observation illustrates intensive inference use. Paciva meters inference separately, so it does not establish the proposed subscription’s value, customer savings or the M5 capacity estimate, and adds nothing to the founder value calculation.

Illustrative invoice: $300 founding subscription plus the credits actually consumed; at list, $3,000 plus consumed credits. Credit rates and spending limits must be confirmed with the offer. Partner referral/revenue-share terms are not published. Partners may implement; Paciva still owns product reliability and must budget support.

Acquisition, compliance, compute and hiring. Acquisition reaches early-adopter knowledge workers and developers and tests conversion. Compliance funds privacy, security, legal and integration-authorization work. Compute provides local/cloud capacity and measures useful workload economics. Hiring adds capacity against product and operating bottlenecks.

The raise supports the next system and the beta-to-paid transition. Allocation percentages, runway, hiring counts and final SAFE terms are not published in this brief. Another beta is underway; the first earlier-variant cohort saw limited usage. The next financing depends on demonstrated adoption and economics, not instrumentation alone.

Building the system forced 41 concepts assessed, with 34 or more reported implemented across 9 interdependent clusters: scoring, screening, trust, memory, planning, gating, voice, briefing, and orchestration. These are internal portfolio classifications, not customer adoption or a patent grant. The defensibility is the interdependence: a competitor faces the system, not a feature list, and the system's behavior is tuned by data they do not have.

Benchmarks run CI-gated: a regression blocks the merge, so the numbers cannot drift quietly. Judge purity rules keep evaluation models separate from production models, and privacy gating keeps benchmark data inside the boundary. Every metric family is labeled: the classifier benchmark, FPR 1.50% to 0.29% and F1 0.90 to 0.94 against a named baseline, is a different family from the per-user personalization curve, 8.5% to 0.41%, and the deck never mixes them. The "70% gain" phrasing is pinned to its named baseline or dropped entirely.

The use-case catalog runs 120 entries: 63 professional, 51 personal, 6 both. The personal half is not garnish. BLS time-use data anchors a personal coordination tax that behaves exactly like the professional one: school forms, travel changes, appointments, renewals, the family logistics layer. The same loop that screens a cold pitch tracks form deadlines and rebooks around flight changes. One assistant, one memory, both lives, behind the explicit boundaries in A25. Email is one channel among many; the unified surface spans LinkedIn, email, Slack, and the widening connector set visible in the product tour.

We are building the next system for paid users while running another beta cohort. The first cohort saw limited usage of an earlier variant. The examples show the intended workflows across eight modes; they do not certify general availability. Release scope and dates should be confirmed with the founders as the beta progresses.

The first beta used an earlier product variant and saw limited usage. A new cohort is underway; current cohort size, retention, paying-customer count and collected MRR are not published here. Internal classifier and personalization figures are separate from adoption evidence; sample sizes, dates and evaluation artifacts are not provided in this brief.

The approximately 24-hour figure is a per-action time-savings model, not a direct measurement of new-cohort usage. The curve drawings illustrate hypotheses and reported endpoints. For the new cohort, report activation as a completed useful workflow, repeat use, D30 retention with its denominator, reversals per action and waiver adoption among eligible actions. Screenshots use illustrative data and are not traction evidence.

The engineering system is the product's thesis applied to ourselves: autonomous implement-review-merge loops with governed gates, where agents write, and humans hold the review gate the way Pax's users hold the confirmation gate. Velocity is measured by accepted merges and review-gate pass rate, never raw PR count, because ungoverned output is not shipping. This is the meta-story on slide 10: the system that lets two people out-ship funded teams is the system we sell. Pax for Code, the productized version, is optionality the plan holds without pricing.

One boundary statement, used identically everywhere: detected contact details, identifiers, and credentials are stripped before the work reaches any external model, live today, and non-frontier workloads can run on local models instead. Injection defense runs three layers, with a containment benchmark row in the same CI-gated harness as everything else. The threat model maps to the OWASP LLM Top 10 and NIST AI 600-1. On the amplifier question: Pax screens inbound and gates outbound; it converses to qualify strangers under its owner's policy, discloses itself as an assistant, and the action-safety taxonomy in A24 keeps consequential sends behind confirmation. AI-disclosure readiness is tracked as a shipping requirement, not a legal afterthought.

If LinkedIn policy changes or authorization is lost, pause affected connector operations and remove those workflows from the offered scope until an authorized route is available. If the integration vendor is unavailable, preserve pending actions and avoid unapproved retries while evaluating an authorized replacement. If Gmail scopes tighten, disable affected capabilities and request only permissions that remain supported. Each case requires notifying affected users and reassessing value and pricing. The retained share of useful work without each channel has not been quantified in this brief.

Every action in the catalog is classified reversible, compensatable, cancellable, or irreversible, and the class decides the treatment: what approval it needs, whether step-up confirmation applies, whether it is waiver-eligible, what the recovery path is, and what audit receipt it writes. Reversible actions undo in one tap. Irreversible actions never run without explicit confirmation, and no waiver can change that. This is an export of the shipped tool registry and policy engine, which is the point: the gate is architecture, not policy, and containment numbers travel only with their benchmark row.

Personal memory and workspace memory are separate stores with provenance on every entry, and authority is a governed matrix: who can see, who can revoke, what the employer can never reach. Tenant isolation separates workspaces; offboarding severs workspace access while personal memory leaves with the person; export and deletion are user rights, not support tickets; and cross-context leakage is tested, not assumed. The matrix maps to the NIST Privacy Framework, the Trust Services Criteria, and GDPR purpose limitation. The one-sentence answer: your employer gets the work, you keep your life, and the boundary is enforced in the schema, not the handbook.

No, because we never add them. The two numbers answer different questions and we keep them apart on purpose. Graymail, $12,796 per employee per year, models hours at a published labor rate: 1.5 hours a day of inbound triage, 235 working days, $36.30 an hour, derived in A14. It is what reaches one person. The Collaboration Tax, about $12,800 per employee per year, is modeled and labeled as modeled: $6,824 is the modeled estimate from tool switching alone, four hours a week across 47 working weeks, reorienting after about 1,200 app toggles a day (Harvard Business Review, August 2022, Soroco telemetry across three Fortune 500s, priced at the same BLS rate); unnecessary meetings add about $3,739; and time spent untangling AI-generated work adds about $2,232, from the workslop research showing 41 percent of workers received it in the past month at roughly $186 per worker per month, with 42 percent trusting the sender less afterward (BetterUp Labs and Stanford Social Media Lab, via Harvard Business Review, 2025). It is what the organization pays to move work between people.

Both draw on the same finite day, so summing them would double-count and we do not do it. A separate survey offers context, not an independent bound: about 60 percent of the workday goes to communicating about work, searching for information, and switching apps rather than the job the person was hired to do (Asana, Anatomy of Work Index). Components can overlap within the coordination estimate as well; treat it as directional, not a proven loss or guaranteed recovery. The related finding we lean on for direction rather than size: roughly 95 percent of organizations saw no measurable return on generative AI, largely because it was pointed at producing more rather than removing the work (MIT Media Lab, 2025). The full write-up is the Collaboration Tax guide at paciva.ai.

Acquisition is a use of funds. The initial audience is early-adopter knowledge workers and developers. The new beta tests activation and repeated use before scaling direct, referral or implementer channels. Compare acquisition cost, paid conversion and onboarding/support load by channel. Partner distribution and personal expansion are hypotheses; no validated channel economics or automatic second subscription is assumed.

The screenshots are examples of product surfaces; the embedded check-in and screening flows are interactive demonstrations with illustrative data. They are not proof of customer usage or savings. Ask the founders for a walkthrough of the current system and next-beta availability.

Every dated market claim is re-verified within a week of any meeting; this market moves monthly. The do-not-use ledger is as important as the use ledger: the urban-legend "23 minutes to refocus," untraceable IDC search stats, and stale Copilot seat counts appear nowhere in this deck.

Because only half the system starts cold. Aggregate priors, the classifiers, heuristics, and defaults tuned by the whole corpus, arrive warm on day one, and the twelve-month backfill seeds your memory before your first check-in. Day zero starts with per-user calibration, and that is the honest concession: the first week asks for more confirmations than the tenth. The curves on slide 7 show that week compressing. Team pricing is set at GA; the individual ladder carries until then.

A wrapper's value disappears when the model improves. Ours grows: better models make the deterministic-first architecture cheaper to run and the judgment layer more valuable. The bulk of actions resolve in code before any model call, inference runs behind multi-provider failover across three labs, and the moat lives in per-user feedback data no model ships with. The model is a component. The company is the memory, the policy engine, the trust graph, and the corpus.