The cheapest feedback-loop install is four sentences, and only one of them costs anything
The question
"What is the minimal instrumentation that flips a 'fuzzy' client decision into one with a real-time feedback loop — i.e., MAC's cheapest possible feedback-loop install?"
Follow-up #5 from the 2026-06-12 conviction brief, which ended by conceding that real-time observable feedback is the one lever that breaks miscalibrated AI-to-human confidence transfer — and then never priced the lever.
What we already know (from the vault)
- The question's own noun is stale. [[2026-06-12-agentic-targeting-conviction-calibrated-confidence]] framed MAC as an "instrumentation service that earns the right to transfer conviction" for clients. That framing died three weeks earlier: [[2026-05-11-mac-pivot-retainer-to-info-product]] killed retainer-MAC and Client Reporting outright, load-bearing reason being the phData conflict of interest, not time cost. MAC is a $350 one-time executable plugin the buyer installs and runs unattended ([[2026-05-14-mac-pricing-intent]]). So the install has to be self-installing — zero RDCO humans in the loop. That is a strictly harder constraint than the parent brief assumed, and it is what makes "cheapest" the right question rather than a cost-cutting one.
- The vault already holds the answer's spine, and it isn't a tool. [[operational-definitions]] (Deming via Wheeler & Chambers, Ch. 11.2) states the minimum specification for any countable thing: criterion + test + decision rule, and names the decision rule — the threshold converting an observation into yes/no — as the part every team skips. A "fuzzy" decision is, definitionally, one whose decision rule was never written. The fuzz is not missing data; it is a missing sentence.
- The four-layer filter treats instrumentation and feedback loop as separate layers ([[2026-04-30-rdco-thesis-targeting-systems-feedback-loops]]), and [[2026-04-30-amazon-wbr-metric-trees-business-stack-view-research]] maps them onto OODA: instrumentation = observe, feedback loop = the return path. Installing instrumentation without the return path buys a dashboard, not a loop. Most "we instrumented it" installs stop one layer short.
- Tooling is empirically not the constraint. [[2026-05-11-practicaldatamodeling-april-2026-survey-results]] — only 4.8% of 334 practitioners named tooling as their data-modeling bottleneck. Any answer to this question that begins with "install a tool" is answering a problem 95% of the market doesn't have.
- MAC already has a working instance of the loop, with production numbers. The customer-zero
/dqengagement ([[2026-05-14-mac-pricing-intent]]) ran 5 → 105 tests in ~6 weeks on one model, and the release gate held 8+ consecutive nightly cycles without human intervention. That last number, not the test count, is the feedback-loop evidence: the loop closed nightly, unattended. - Loop tightness is the active ingredient, not loop existence ([[2026-04-19-commoncog-much-ado-about-the-ooda-loop]] — the kernel that survives Chin's OODA takedown is that short feedback cycles beat careful planning under uncertainty).
What the web says
- The A/B-testing world's name for the missing sentence is the OEC. Kohavi's Overall Evaluation Criterion is explicitly "the experiment's decision rule" — a single agreed metric that converts a business objective into a pass/fail. The stated discipline is that trade-offs get made once, in advance, so the organization is aligned before results arrive (Analytics Toolkit glossary, quoting Kohavi et al.). Same three-part structure as Deming, arrived at independently.
- The OEC literature also names the hard part, and it is not the threshold. A good OEC "should not be short-term focused (e.g., clicks); to the contrary, it should include factors that predict long-term goals, such as predicted lifetime value and repeat visits" — i.e. it must be simultaneously short-term measurable and long-term predictive. Practitioners are told this "often requires multiple iterations to refine." Notably, the accessible glossary sources are prescriptive about the ideal and near-silent on how to pick one or how to know you picked wrong — the proxy-selection problem is acknowledged and unsolved in the public literature.
- Decision-log practice quantifies how little of a log is actually instrumentation. The canonical core template is seven fields — title, description, date, decision maker, rationale, alternatives considered, expected impact (Plane). Six are archival. Only "expected impact or outcome" creates a comparison a future review can score. Review dates are classified as "optional but high-value," which is precisely backwards: without one, the loop never closes.
- Leading-indicator guidance converges on latency, not count. "The most useful leading indicators respond to effort within days or weeks, giving teams a tight feedback loop between action and signal," and "choosing the right leading indicators is more important than tracking more of them" (KPI Tree). The practical rule circulating in the operations literature — a handful of lagging outcomes each paired with two or three leading drivers is enough to manage — is directional folklore, not a measured finding; treat it as a ceiling heuristic, not evidence.
- Nobody has published a rigorous minimum. Three searches across decision quality, minimum viable instrumentation, and OEC design returned marketing content, tutorials, and definitional glossaries. There is no empirical study of "what is the smallest instrumentation that flips an uninstrumented decision." That gap is itself the finding: the question is under-served, which is why it is available as positioning.
Convergences and contradictions
- Strong convergence across three unrelated traditions — SPC (Deming/Wheeler), online experimentation (Kohavi), and decision-log practice — on the same skeleton: name the thing, specify how you observe it, write the threshold in advance, and schedule the comparison. None of the three treats software as part of the minimum. This is a rare case where the vault's concept doc and the outside literature are structurally identical without citing each other.
- Contradiction with the parent brief's implied cost model. [[2026-06-12-agentic-targeting-conviction-calibrated-confidence]] positioned instrumentation as the service RDCO sells, implying the install is the expensive, defensible part. The evidence says three of the four parts are free writing. The expensive, non-templatable part is choosing the proxy observable — the one thing the OEC literature admits it cannot teach.
- Unresolved tension the vault should not paper over: MAC's proven loop (nightly
/dqgate, 8+ cycles) works in a domain that was already instrumentable — rows exist, comparisons are computable. It is evidence that MAC ships a working loop; it is not evidence that MAC flips a genuinely fuzzy decision. Presenting the 140,937-ghost-row find as proof of fuzzy-decision instrumentation would be exactly the overreach the 2026-05-07 MAC calibration correction already caught once.
Synthesis for RDCO
The minimal install is four sentences and a cron entry. (1) A written decision rule — the threshold that converts an observation into yes/no, written before the decision, because the fuzz lives in the unwritten threshold rather than in missing data. (2) A pre-registered expected outcome with a number and a date — the one decision-log field that is instrumentation rather than archaeology. (3) A proxy observable whose latency is shorter than the decision's cadence — this is the only part that costs anything. (4) A scheduled scoring event that compares (2) to what happened and logs the delta, automated if the cadence is nightly, calendared if it is monthly. Parts 1, 2 and 4 are specification work: minutes, not dollars. Part 3 is judgment, and judgment is the whole product.
That reframes what "real-time" means and makes it cheap. Real-time is not sub-second; it is faster than the decision recurs. A nightly gate is real-time for a team shipping daily. This also produces the selection rule for which fuzzy decision to instrument first, and it is the opposite of the instinct: instrument the decision you make most often, not the biggest one. The Tesla-fleet analog in [[2026-04-24-targeting-system]] says the agentic targeting system beats the implicit one on reps-per-unit-time, not on stakes. A loop that resolves once a quarter delivers four data points a year — it will not out-learn a senior operator's priors inside the operator's tenure, so instrumenting it is a spend with no payback horizon. High-frequency, low-stakes decisions are where a cheap loop compounds; the founder's own /dq release gate is a nightly-cadence decision, which is exactly why it worked.
The commercial implication cuts against the current MAC drip plan. The day-1..5 give-give-give drip is still undrafted and still flagged as blocking the launch funnel ([[2026-05-04-mac-product-shape-decisions]]). This brief supplies its spine — but it also says the giveaway is three-quarters of the method. That is fine, and probably good: at a $350 no-brainer price point sold into the IC data engineer as a level-up, giving away the specification discipline and charging for the encoded proxy-selection judgment (which Scope × Basis cell actually catches the failure you have) is a coherent offer. What it kills is any future framing of MAC as "we install your feedback loop" — the install is not the scarce thing, and a buyer who reads the drip will know it. Sell the 3×6 matrix as pre-encoded proxy selection by a principal engineer, which is the honest description of what the $350 buys and the only part that survives the analysis above.
One risk to name explicitly, because it inverts the whole value proposition. A cheap loop pointed at the wrong proxy is worse than no loop. It manufactures conviction fast — and the parent brief's confidence-alignment finding says humans absorb a confident signal's calibration without any improvement in their actual decision capability. A wrong-proxy loop is a machine for producing well-calibrated-feeling error, at speed. So the minimal install has a mandatory fifth element that is not instrumentation at all: a review condition on the proxy itself ("if the proxy and the outcome disagree N times, the proxy is wrong, not the decision"). [[operational-definitions]] already demands a change log for exactly this reason. Any MAC drip or Sanity Check treatment that ships the four-part kit without the proxy kill-switch is shipping the fast-wrong-conviction machine, and RDCO would be doing at $350 the thing it plans to criticize vendors for doing at enterprise prices.
Why this is in the vault
This is the content spec for the MAC day-1..5 email drip, which [[2026-05-04-mac-product-shape-decisions]] still lists as undrafted and blocking the launch funnel — and it resolves what the drip can give away (the four-part specification) versus what the $350 buys (pre-encoded proxy selection). It also retires the "MAC as instrumentation service" language that [[2026-06-12-agentic-targeting-conviction-calibrated-confidence]] left standing after [[2026-05-11-mac-pivot-retainer-to-info-product]] had already killed the retainer shape.
Open follow-ups
- Is there any measured evidence for the "reps-per-unit-time beats stakes" selection rule, or is it an analogy imported from the Tesla-fleet frame? If it is only an analogy, RDCO should stop stating it as a rule in external material.
- What is the actual failure rate of proxy selection — how often does a chosen leading indicator diverge from the outcome it was picked to predict, and after how many cycles is that detectable? The OEC literature admits the gap; someone in causal inference may have measured it.
- Does the four-part kit survive contact with a genuinely fuzzy business decision (hire / client-take / pricing), or does part 3 become unsolvable outside instrumentable technical domains? A worked non-data example would test the generalization claim before Sanity Check makes it.
- If three of the four parts are free, what is the smallest artifact that could deliver them at scale — a one-page template, a
/decision-ruleskill, a Sanity Check "Before Monday Noon" close — and does giving it away strengthen or cannibalize the $350 offer? - Should the proxy kill-switch ("if proxy and outcome disagree N times, retire the proxy") be a shipped MAC feature rather than a caveat — i.e. is calibration-protective UX the productizable differentiator the 2026-06-12 brief already flagged as an open question and nobody has since answered?
Related
- [[2026-06-12-agentic-targeting-conviction-calibrated-confidence]]
- [[2026-04-24-targeting-system]]
- [[operational-definitions]]
- [[2026-05-11-mac-pivot-retainer-to-info-product]]
- [[2026-05-14-mac-pricing-intent]]
- [[2026-05-04-mac-product-shape-decisions]]
- [[2026-04-30-rdco-thesis-targeting-systems-feedback-loops]]
- [[2026-04-30-amazon-wbr-metric-trees-business-stack-view-research]]
- [[2026-05-11-practicaldatamodeling-april-2026-survey-results]]
- [[2026-04-19-commoncog-much-ado-about-the-ooda-loop]]
- [[2026-04-15-commoncog-process-behaviour-charts]]
Sources
- Vault: [[2026-06-12-agentic-targeting-conviction-calibrated-confidence]] —
~/rdco-vault/06-reference/research/2026-06-12-agentic-targeting-conviction-calibrated-confidence.md(parent brief; real-time feedback as the one lever, confidence-alignment contamination) - Vault: [[operational-definitions]] —
~/rdco-vault/06-reference/concepts/operational-definitions.md(criterion + test + decision rule; the skipped third part; change-log requirement) - Vault: [[2026-04-24-targeting-system]] —
~/rdco-vault/06-reference/concepts/2026-04-24-targeting-system.md(implicit vs agentic split; Tesla-fleet reps argument; "MAC narrows the fuzz") - Vault: [[2026-05-11-mac-pivot-retainer-to-info-product]] —
~/rdco-vault/01-projects/mac/2026-05-11-mac-pivot-retainer-to-info-product.md(retainer + Client Reporting killed; phData conflict) - Vault: [[2026-05-14-mac-pricing-intent]] —
~/rdco-vault/01-projects/mac/2026-05-14-mac-pricing-intent.md($350 low-ticket; customer-zero numbers; 8+ nightly gate cycles) - Vault: [[2026-05-04-mac-product-shape-decisions]] —
~/rdco-vault/01-projects/mac/2026-05-04-mac-product-shape-decisions.md(executable-course shape; undrafted day-1..5 drip blocking launch) - Vault: [[2026-04-30-rdco-thesis-targeting-systems-feedback-loops]] —
~/rdco-vault/06-reference/2026-04-30-rdco-thesis-targeting-systems-feedback-loops.md(four-layer filter) - Vault: [[2026-04-30-amazon-wbr-metric-trees-business-stack-view-research]] —
~/rdco-vault/06-reference/2026-04-30-amazon-wbr-metric-trees-business-stack-view-research.md(four layers → OODA mapping; decision-trace pattern) - Vault: [[2026-05-11-practicaldatamodeling-april-2026-survey-results]] —
~/rdco-vault/06-reference/2026-05-11-practicaldatamodeling-april-2026-survey-results.md(4.8% name tooling as the bottleneck, n=334) - Vault: [[2026-04-19-commoncog-much-ado-about-the-ooda-loop]] —
~/rdco-vault/06-reference/2026-04-19-commoncog-much-ado-about-the-ooda-loop.md(loop tightness as the portable kernel) - Web: Overall Evaluation Criterion (OEC) glossary, quoting Kohavi et al. — https://www.analytics-toolkit.com/glossary/overall-evaluation-criterion/
- Web: Decision log — what it is, why teams use it, and template (seven core fields) — https://plane.so/blog/decision-log-what-it-is-why-teams-use-it-and-template
- Web: Leading vs Lagging Indicators (response-time window; quality over quantity) — https://kpitree.co/guides/core-concepts/leading-vs-lagging-indicators