Reference and stewardship

Use High, Medium or Low confidence in a specified proposition, not in an entire programme. For Qualified or Rejected historical claims, state confidence in the editorial narrowing or rejection and its supported boundary; never attach it implicitly to the rejected headline. “Unknown” identifies an unestablished local quantity. Accepted claims survive the editorial challenge; accepted-with-conditions claims require their stated prerequisites. Qualified claims have narrower scope. Advanced methods require additional capability. Experimental ideas need capped validation. Unresolved questions stay visible. Rejected ideas do not reappear as recommendations.

Evidence indicates a conclusion only within its method and population. Evidence suggests a direction when material uncertainty remains. Early evidence may justify testing, not a broad promise. Source inheritance does not create independent corroboration. Architectural feasibility and high confidence in an accounting identity do not establish commercial ROI.

Booking.com and experiment-linked machine learning. The problem was turning model development into business improvements. The supplied research describes a primary practitioner paper about successful deployed ML applications and their experimental validation. The useful mechanism is connecting each model to an outcome and a decision. The reported outcome is production precedent across roughly 150 successful applications, not a denominator of all attempts or a pooled profit return. Evidence is a primary technical account, with selection and scale limits. Learn to evaluate the decision effect; do not infer that modern agent orchestration, a small firm or every ML project will achieve the same result. Booking.com KDD paper. Source treatment: OS §11; RT §§7,50–51. Publication: 2019 (year precision); Tier 2 practitioner evidence; not independent commercial evidence.

Unilever and product-image production. The problem was the cost and speed of imagery production. The company described a digital-twin and AI workflow that reduced production time and cost. The mechanism is a changed asset-production process, not an autonomous commercial strategy. The company-reported outcome illustrates a potential expense or throughput benefit; it does not isolate pure model causality, demonstrate incremental sales or provide a transferable return estimate. Evidence is a company case. Learn to compare accepted assets and actual purchases; do not generalise it to media, retention or an integrated GTM platform. Unilever announcement. Source treatment: OS §11; RT §50. Publication: 18 March 2025; Tier 3 company-reported evidence; not independent.

These references are carried forward from the supplied research and Red Team review. Editorial & Evidence QA checked the existing primary pages and publication metadata on 9 October 2026. This was a bounded citation check, not independent validation of commercial outcomes or a fresh audit of the full Booking.com methods. Their stated scope is preserved. The excluded Uber amount, provider ROAS claims and universal uplift ranges do not enter any investment model.

For a justified machine-executed action, preserve the link from decision/version to approval, exact action, platform receipt, exposure, outcome and evidence. Bind approval to the material action fields. Recheck current permission, destination, state and cumulative limits before execution. An indeterminate platform result requires a state query and reconciliation before retry; a successful API response is not necessarily a delivered customer outcome. Keep manual or native fallback and a named restart authority. [SYS §§15,20–22,25; AGT §24 and contract pack]

For a justified predictor, record objective, target, horizon, population, training/holdout periods, leakage tests, feature provenance, calibration, missingness, relevant subgroup performance, version, owner, drift trigger and fallback. For uplift, add identification and treatment-policy evaluation. For generated summaries, verify source fidelity and missed/contradictory evidence. These are different evaluations.

The specialist sources retain detailed API schemas, identity patterns, statistical methods and security controls. Consult SYS §§23–28 and AGT’s Implementation Contract Pack only after the chosen Play requires those capabilities. The five working artefacts remain the operational record of what was actually configured.

The following remain unresolved locally: additional AI profit, coordination value, five-year ownership cost, measurement power, cohort forecasting quality, available treatments, budget mobility, reviewer capacity, native platform adequacy, customer acceptance, memory reuse and the value of greater autonomy. The Editorial Synthesis Report’s issue register records why each matters, required validation and the consequence of error.

Unresolved Design Choice: choose a CRM/export approach when one workflow can be supported economically; choose shared analytical services when several funded workflows need consistent joins and economics. Choose embedded execution with shared standards as the initial organisation; consider a shared team only for recurring coordination demand. Choose rules and cohorts when treatment evidence is sparse; consider individual treatment models only when the evidence and extra value support them. No universal architecture or team structure resolves these context-dependent choices.

Additional Research Required: before financial commitment, obtain actual data, costs, licensed-tool tests and authority evidence. Before using an excluded public number, verify its original method and cost basis. A missing input remains unknown; it is not an invitation to insert a market-average uplift.

Update a claim, its linked Play, tool and evidence record together. Keep the old version and reason for change. Reopen a rejected recommendation only with explicit evidence and an editorial verdict, not a renamed feature.

An advanced GTM organisation knows which customer problems it can change, values them honestly, gives people authority to act, and learns within the limits of its evidence. Its AI performs useful, bounded work. Its operators can stop, recover and simplify the system. Its advantage, where it exists, comes from better interventions and the ability to repeat them economically.

Build the necessary system. Prove the additional value. Automate what earns automation. Retain what improves the next decision. Repeat.