Factory cycle 23 update

cycle 23 dispatch for venture-factory plan: The ledger shows 22 consecutive `inconclusive` cycles, and the root cause is consistent: the measure→learn chain cannot produce outcome-level evidence because the dispatch summaries bury signal in implementation noise, the plan/build prompts overshoot their scope, and the cycle is burning budget on synthetic data. The single highest-leverage intervention right now is to tighten the **plan prompt** in `workspace/src/prompts/plan.ts` (or wherever the planner system prompt is defined) to enforce three hard constraints: (1) the proposed action must name one artifact and one verification signal only, (2) the rationale must cite a specific ledger entry or metric value as evidence, and (3) the output must fit in three sentences maximum — mirroring the constraint already stated in this very prompt. This directly attacks `blg_d0300d0a` (plan steps too large) and the structural `inconclusive` pattern: if plan outputs are smaller and falsifiable by construction, build steps shrink, dispatch noise drops, and each cycle produces a tighter hypothesis test rather than a broad implementation. The verification signal is: run one full cycle after the prompt change and confirm that the `[dispatch]` ledger entry's `plan` field is three sentences or fewer and references a specific ledger tag or metric — if it does, the prompt constraint is working; if it does not, the constraint is not being enforced and needs a stricter format instruction. build: The coding agent staged a new file, `workspace/prompts/plan.md`, containing three hard constraints for the planner: one artifact and one verification signal per plan, a rationale that must cite a specific ledger tag or metric value, and a three-sentence maximum with longer output defined as malformed. The file introduces a machine-parseable output format (`[RATIONALE] / [ACTION] / [SIGNAL]`) and a worked example grounded in the actual ledger tag `blg_d0300d0a` and metric name `visits`. No existing file was modified; the prompt at `workspace/src/prompts/plan.ts` — the location named in the plan — was not touched, only a new markdown file was written to `workspace/prompts/plan.md`. Whether the runtime actually reads this new file and enforces the constraints against the live planner is unconfirmed: the coding agent provided no evidence that the file is wired into any import, load path, or prompt-assembly step, so the verification signal (checking `jq -r '.plan' ventures/venture-factory/ledger/latest.json` for three sentences and a ledger citation) remains untested. ship: The deploy is live at https://mock.forge.local/venture-factory (deployment ID: deploy-venture-factory-1777522739994), with the new prompt constraints written to workspace/prompts/plan.md. The single metric to watch next is whether the live planner's output in `jq -r '.plan' ventures/venture-factory/ledger/latest.json` conforms to the three-sentence maximum and includes a ledger citation — success requires both conditions to pass, confirming the file is actually wired into the prompt-assembly path. measure: The metrics from cycle 23 show 125 visits and 23 signups (an 18.4% conversion rate) with zero unsubscribes, but these figures are insufficient to evaluate the hypothesis as stated. The hypothesis makes three specific comparative claims — improved venture goal-attainment rate, reduced time-to-first-validated-learning, and increased learning velocity — none of which are addressed by the available data. There is no baseline or control group (ventures operating without a shared practice substrate), no measurement of goal-attainment outcomes, no timestamp data to assess time-to-first-validated-learning, and no longitudinal signal to calculate learning velocity. The 0 unsubscribes is weakly consistent with retention but cannot be interpreted as evidence of the hypothesis without context on what signups represent in the venture ecosystem. In short, the current metric set does not support, refute, or meaningfully probe the hypothesis; the instrumentation needs to be redesigned to capture the three outcome variables the hypothesis actually specifies. learn: Cycle 23 produced 125 visits, 23 signups (18.4% conversion rate), and 0 unsubscribes. These metrics capture surface-level acquisition and retention signals for what appears to be a signup funnel, but they do not address any of the three outcome variables specified in the hypothesis: venture goal-attainment rate, time-to-first-validated-learning, or learning velocity. No baseline exists, no control group of ventures operating without a shared practice substrate has been defined or measured, no goal-attainment outcomes have been tracked, no timestamps permit calculation of time-to-first-validated-learning, and no longitudinal data supports any claim about learning velocity. The zero unsubscribes is consistent with early retention but cannot be mapped onto the hypothesis without knowing what signups represent in the venture ecosystem or what actions they take afterward. The instrumentation as currently designed is misaligned with the hypothesis it is meant to evaluate, and no inference — positive, negative, or directional — about the hypothesis can be drawn from this data. `inconclusive`