Towards RSI-GTM: Trust-Constrained, Verification-Gated Recursive Self-Improvement for Go-to-Market

Abstract Agentic GTM drives the execution cost of outreach toward zero, but at the same time exposes an old problem at a new scale: existing systems optimize the quality of a single outreach, not the long-run consequences of an entire outreach sequence. When sellers use AI to generate outreach and buyers use AI to filter outreach, volume-based strategies push reply metrics and trust loss up together, while reply rate, domain reputation, and brand trust do not move in the same direction. This paper asks a question: if the next generation of GTM is destined to be an autonomous system rebuilt by AI and code, in what way should that system improve itself, so that it does not mistake "sending more" for "getting better"? This paper proposes RSI-GTM, a GTM recursive-improvement framework constrained by trust budgets and governed by a verification gate as its adoption rule; its contributions are formalization, a benchmark design, and synthetic negative results, not proof that this paradigm works at real vendors. This paper's answer is pulled back to the range the evidence can support; the complete conclusion is stated only in Chapter 11, and this abstract is only a guide: first, the formalization of trust constraints — GTM is defined as a sequential decision problem of matching model capability to willingness to pay with minimal friction and minimal trust loss, with the trust budget accounted separately across four dimensions — account, channel, brand, and partner — never transferable across dimensions, and waiting as a first-class action; second, the benchmark design of GTM-PolicyBench — a simulator, a trajectory-level offline evaluation layer, and a prospective trial protocol, with a yield threshold, a Pareto frontier, and mandatory harm disclosure; third, veto evidence within synthetic settings and a set of negative results (see the main results table in Chapter 7 and Chapter 11). The recursive-improvement architecture itself is only a future vision in this draft, not a proven conclusion: recursion with a verification gate robustly vetoes bad recursion within synthetic settings; but on efficiency, the threshold layer and the joint layer, and the state layer have not been shown to beat a trust-constrained strategy without recursion (§§7.3, 7.7); the workflow-only layer holds within synthetic settings under a 24-week, 8-week-epoch design (narrow margin, small effect, see §7.8), and that gain does not hold under a 48-week, 4-week-epoch design (see §7.9), remains a marginal effect of the same order under a 48-week, 8-week-epoch pure-time control (interval lower bound only +0.0001, see §7.10), still has an interval containing zero after the candidate space is expanded (see §7.10), and under strict-superset retesting is still depth 1 with an empty-comparison rate still above 60% (see §7.11), while independent validation seeds broke depth 1 in the sense of adoption counts (empty-comparison rate 73.3% → 20.0%, adoption distribution spread over 5 rounds) but produced no progressive recursion (total progressive adoptions = 0; both multi-adoption cases were flips back), and the increment over the same-batch control is unconfirmed (D−A0 +0.0030, interval containing zero, see §7.11), so it can only be stated as "a one-off small gain under the 24-week design, still a marginal effect of the same order at 48 weeks with the interval at the zero line's knife edge, with horizon and epoch not significantly changing it, and not shown to scale with time or rounds," it does not re-establish "workflow recursion gain confirmed" because arm A's lower bound is barely positive, is not to be presented as "recursion strengthens with rounds," and must not be used to infer that joint or general recursion holds. The §7.11 independent-validation-seed arm is stated only as breaking depth 1 in the sense of adoption counts; it is not to be interpreted as progressive recursion having been restored, and the significance of D−STATIC must not be substituted for the D−A0 comparison. The dual-sample gate test (see §7.12) further qualifies that statement: under dual-sample confirmation, D's effect almost disappears (E vs. static +0.0010, interval containing zero; E−D0 −0.0079, interval excluding zero), showing that D's adoptions lack cross-sample robustness and mainly come from multiple chances to pass (with the gate-tightening qualification); the fallback arm, under a binary candidate set + independent validation seed design, has almost no increment (F−D0 +0.0000), but that test is unfair (the candidate set does not contain W2, and the original-evidence cases for seeds 6 and 29 do not exist in this arm), and is not to be interpreted as the fallback rule being generally invalid. Neither dual-sample confirmation nor the fallback rule produced two-steps-forward recursion; total progressive adoptions are 0 in all three arms. On sequential evaluation, trajectory-level fitted Q evaluation can be used only as a ranking heuristic, while naive bandit offline evaluation and sequential doubly robust evaluation under a 24-step deterministic policy can both not be used as a verification layer. The true predictive value of public signals for developer ramp rates has not been confirmed, and real weekly panel collection is a negative result at this stage; public tests are further negative results and directional support, not passes. All of this paper's positive conclusions are confined to synthetic settings; real-vendor effects, conversion-rate improvements, and cross-vendor generalization are all future work. Code and data: https://github.com/louisliqi0315/RSI-GTM This is a preprint, not peer-reviewed; results are limited to synthetic data.