Skip to content

What's Changed

This page records how this knowledge base has changed over time — new and expanded content, citation corrections, freshness updates, and structural changes — captured release by release. How the knowledge changed is itself a form of provenance.

v13.1.0 — 2026-09-29

New & expanded content

  • Before You Buy AI for Your Warehouse — a new Decision guide for an operator who already owns a data warehouse and has a vendor pitch in front of them. It does not rank platforms. It sets out a five-step test, costed across two weeks, for finding out what the technology does on your own tables.

The finding it is built on: two unassisted measurements, one published by a platform vendor about its own data and one from an academic benchmark built from real private warehouses, differ from each other by more than a factor of two. Those two are not strictly comparable either, which the guide states rather than smooths over. Once the system, the questions and the scoring all move, no published figure describes your situation.

The guide also sets the pass mark before the test rather than after, holds ten of its twenty questions back so the definitions written partway through are not scored on the questions that shaped them, and says plainly that twenty questions is a coarse sample.

Citation corrections

  • The Planning Year and Running It When It Is Already Broken — two source records corrected. A budget-cycle figure is recorded against the business unit the survey actually asked about rather than the whole firm, and the software-teams OKR study is no longer credited to a conference track that could not be confirmed. Quotes remain verified against the preprint they were always checked against.

v13.0.0 — 2026-09-26

New & expanded content

New bundle: Operating Cadence — how an organization sets direction and keeps it moving across a year, with every layer of the stack graded by what the evidence actually supports and what is inherited convention.

  • How the Planning Stack Fits Together — why the vocabulary of mission, vision, purpose and strategy is genuinely unsettled, what the research says about mission and values statements, and where the famous strategy execution-failure statistic came from. It opens with the short list of what survives scrutiny.
  • The Planning Year — the end-to-end annual sequence, from strategy refresh through annual and quarterly goals to the monthly and weekly reviews. Only two organizations have published a dated sequence, and they disagree with each other.
  • Running It When It Is Already Broken — how to diagnose a planning system that has already disappointed people, what to change in week one, and what to remove rather than add.
  • Glossary and Validation — thirteen terms, and the sourcing rules this bundle holds itself to.

by @grandaha in https://github.com/grandaha/knowledgeville-workspace/pull/307

v12.11.0 — 2026-09-20

New guide: does this need a language model?

A recurring decision has to be made by something. This guide compares four options and shows how to test the choice with your own numbers: a rule you write, a general language model told what you want, a model trained on your own examples, or two of those working together.

Two cheaper steps come first, and neither is a model. Measure what the work costs you now, including how often it is currently wrong. Write the rules for the cases where a person would not hesitate. What is left is a smaller problem, and sometimes there is nothing left to automate.

A few labelled examples are worth more than most people assume. On a banking benchmark with 77 possible answers, a trained model reached 83.42% accuracy from ten examples per category, against 93.66% on the full training set. Both framings of that pair are true and the guide gives both: ten examples got most of the way up the scale, and the full set still removed about three fifths of the remaining errors. That benchmark is also the friendly case, with balanced and cleanly separable categories, so imbalance and overlap push the number you need up sharply.

When accuracy drops, check the threshold before replacing the model. A trained model returns a score, and a person picks the score above which you act. In one study a detector lost 28 points of F1 when the people it was detecting changed tools; recalibrating that cut-off on a few hundred fresh examples recovered 86% of the loss with nothing retrained. The guide includes the test for when recalibration is not the answer.

Rules are cheap to run and expensive to keep. In an earlier generation of rule-based automation, EY put the failure rate for initial projects at 30 to 50%, and Forrester found $3.41 spent on services for every $1 of software. Write the obvious rules, give them an owner, and stop there.

What you trained is not what you own. Cohere retired fine-tuning for its classification models in September 2025, and previously fine-tuned models stopped being accessible. Customers lost the trained thing itself, not just the ability to make more. Your labelled examples are the part of that investment that survives.

The guide closes with a worked example: 4,000 support tickets a month into twelve queues, run through every step. It lands on using a general language model, not because a language model is clever, but because at that volume anything you build costs more than it saves.

Page: Does This Need a Language Model?

Updated

Which AI Model for Which Job now points readers at the prior question, whether the work needs a language model at all, before comparing language models to each other.

The Decision Guides glossary adds deflection, fine-tuning, labelled examples, precision and recall, and threshold, and extends routing to cover a cascade.

v12.10.1 — 2026-09-19

Corrected: what US law actually requires of AI in employment decisions

Three pages described the US regulatory position in a way a US reader could reasonably take as "nothing here binds us." That is wrong, and the correction turned out to be larger than the wording.

The regulatory landscape section compared three frameworks, none of which is US law. NIST's framework, the EU AI Act and ISO 42001 sat under an enforceability column reading "Voluntary (US)". Read together, the page's only regulatory map contained no binding US obligation at all. It now carries the ones that do bind, and says plainly that the table compares frameworks rather than jurisdictions.

What binds today. New York City requires an independent bias audit of an automated employment decision tool within the past year, a published summary, and candidate notice at least ten business days before use. Illinois makes discriminatory use of AI a civil rights violation across recruitment, promotion, discharge and discipline, not only hiring, and makes failing to notify employees a separate violation. California makes it unlawful to use an automated-decision system that discriminates on a protected basis.

What is arriving. From October 2026, Connecticut provides that using such a tool "shall not be a defense against a complaint alleging a discriminatory practice", and requires an employer filing a federal layoff notice to disclose whether the layoffs relate to its use of AI. Colorado follows in January 2027, with the state attorney general's implementing rules still to be adopted.

The part that is not new at all. The federal Uniform Guidelines on Employee Selection Procedures have shaped selection testing since 1978, and the pages now state their four-fifths rule accurately. A selection rate below four-fifths of the highest group's rate "will generally be regarded ... as evidence of adverse impact," while smaller differences "may nevertheless constitute adverse impact". It is a rebuttable inference with exceptions running in both directions, not the pass mark it is often treated as.

These obligations differ in kind, and that matters more than the dates. One mandates a recurring audit and advance notice. One mandates non-discrimination and notice with no audit at all. One extends an existing discrimination prohibition without either. A compliance approach built for one does not transfer to the others.

Two further corrections. A board-facing page asserted that rules covering hiring were "already in force or imminent" and never named one; it now names two. And a practitioner guide told counsel to be specifically briefed on federal agency guidance covering AI in hiring. That guidance is no longer published, so the briefing now points at the statutes, which have proved more durable than the guidance interpreting them.

Pages: Data Governance · Executive Summary · Standing Up an AI Governance Function

v12.10.0 — 2026-09-18

Updated: what the evidence now says about shared business definitions

The working note on why an AI assistant returns three different answers to the same question has been brought up to date, and one of its claims has been corrected.

A benchmark built on real enterprise data, and it prices one layer of meaning. Most published results on this question run over public databases. This one is built from private enterprise warehouses drawn from real query logs. Telling the model which tables to use, how they join, and what each column means roughly doubles accuracy, from 9.5% to 18.9%. Supplying everything else as well, including the query plan, reaches 25.9%, so about three questions in four are still wrong even then.

Where the work pays depends on how bad your names already are. A paired test of schema annotation gained 8.7 points on the most cryptically named database and lost 2.9 points on the most intuitive one, where the notes introduced noise into a picture the model already had right. That is the first published result that tracks the payoff against how much meaning is genuinely absent, and it suggests a check worth running before spending anything: read the column names your first use case touches as a stranger would.

Two vendor benchmarks, with their own numbers and their stake named. One of them undercuts its own pitch. A semantic-layer vendor's strongest result came from a hand-written 4 KB markdown document pasted into a prompt, with no platform, runtime or tooling involved.

A correction. The note implied the Apache Ossie metric standard confines itself to agreed measure definitions, the bounded work with the good track record. It does not: an ontology specification and an import path from Palantir's ontology both exist in the repository. What survives is narrower and better evidenced. Every ontology artifact sits in an unreleased draft that warns against depending on it, and the specification has no field at all for who owns a definition, who approved it, or whether anyone did.

Read it: Why Your AI Gives Three Different Answers

Corrected: four pages on semantics and metadata

Two pages disagreed with each other about the same practice, and three claims did not survive checking.

When recording competing definitions is documentation, and when it is an evasion. The business glossary guidance still calls for recording known variations in a definition across business units, but only where each definition is genuinely bounded to its own use. Where the variations are instead the output of work meant to settle on one answer, the disagreement survives with governance metadata attached, and an assistant asked a general question still returns whichever variant it reaches first.

"Context layer" is a category still being defined, largely by the firms that sell it, rather than a settled architectural pattern. Three places that asserted the label in our own voice now attribute it to the vendor defining it.

Three claims withdrawn or corrected. That AI agents are now the primary consumers of master data at scale: withdrawn, as nothing measuring it surfaced. That a semantic layer ensures a shared definition: it encodes an agreement rather than producing one, and where the functions never agreed it can only carry the disagreement forward. And the claim that the data mesh model does not address programmatic access for agents was simply wrong, since programmatic addressability is one of the model's named characteristics of a data product.

Pages: Lineage & Metadata · Master Data Management · Data Audits & Automated Quality Governance · Glossary

v12.9.0 — 2026-09-16

New: what to write when AI work carries a graded target

A new page for anyone who has been told that next quarter's objectives must include AI, and who has to write those objectives this week. The advice and one diagnostic question come first, and the evidence behind them follows.

The definition turns out to be the first finding. A 2024 systematic review of the academic literature on Objectives and Key Results found 47 studies, and reported that nine of them cite no source at all for what an OKR is. The commonly cited sources describe materially different practices: a goal-setting discipline, an operational recipe with a fixed number of objectives per quarter, and management by objectives with milestones attached. Measuring the effect of a practice requires agreement about where the practice starts and stops.

The costs of numeric targets are better evidenced than the benefits of this one. No study measuring OKR use against organizational outcomes surfaced in the published record, which describes the literature rather than the practice. By contrast, what targets do to hard-to-measure work is documented across economics, public administration and education research: attention moves toward whatever is cheapest to count, performance crowds toward the target from both directions, and misreporting rises with the stakes.

A quarterly grade suits review far better than cancellation. Returns to this class of investment are J-curved, with short-term losses preceding longer-term gains, so an early decline is an expected shape rather than a verdict.

Two attributions that do not survive checking. The sentence "when a measure becomes a target, it ceases to be a good measure" was written by the anthropologist Marilyn Strathern in 1997, about British university examination grades, and not by Goodhart. And Andrew Grove's High Output Management, the book credited with inventing the practice, contains the mechanism but never the acronym, and states that the system "is not a legal document upon which to base a performance review".

Read it: Goals and Targets for AI Work

v12.8.1 — 2026-09-16

Updated: independent evidence now sits beside the consultancy figures

Consultancy figures now say whose survey they are, and independent evidence sits beside them. Across the Enterprise AI Transformation bundle, 44% of citations came from consultancies or analyst firms, and the concentration mattered more than the average: the executive summary sat at 85%, and the measurement and strategy pages at 68 to 84%. Where such a figure carries a claim, the page now names the publisher, says that respondents are describing their own organizations, and notes that the publisher sells the work.

Eighteen findings from official statistics and academic research now sit beside the consultancy surveys. The headline adoption gap turns out to be definitional rather than a disagreement. 19.8% of US firms and 20.0% of EU enterprises report using AI, counting firms one vote each, so small businesses dominate. The Census AI supplement puts the firm share at 18% for November 2025 to January 2026, and at 32% weighted by employment, reaching 50 to 60% among very large firms in the Information, Professional Services and Finance sectors. On realized value, a Federal Reserve and Duke survey found reported AI productivity gains about three times the gains implied by the same firms' own revenue and employment, and a survey of some 6,000 executives found nine in ten reporting no impact at their own firm.

The claim that centralizing AI capability improves outcomes is now presented as contested. The guide to running an AI strategy process previously told readers that the evidence favors holding the center. Independent research complicates it: across 949 plants, better information technology went with more autonomy for local managers, while communication technology centralized decisions. Neither that study nor the analytics research examines an AI centre of excellence, and no independent study measures returns from a central AI team against a federated one, so the pages now say so.

The MIT Project NANDA figure is no longer described as independent research. It is a report that was not peer reviewed, held here through a secondary copy, and the entry now records what its own methodology section says: figures described as directional, drawn from interviews rather than company reporting, with success defined by the researchers.

Pages updated: executive summary · program architecture · integrated assessment · AI strategy framework · running an AI strategy process · measurement framework · measurement maturity assessment

v12.8.0 — 2026-09-16

New: where should the human check go?

A research note on a question that follows from the Chaining Tasks brief: when AI does several steps of a job in a row, where should a person check the work, and does one check at the end catch mistakes made earlier?

Fields that run high-consequence work, including surgery, nuclear power and banking, put the check just before the step that commits, where a mistake becomes costly or hard to undo. Recent research gives reasons to doubt that a single check there catches earlier mistakes. Errors feed on each other across steps, agents rarely notice a mistake they inherit, and reviewers tend to accept flawed AI output while growing more confident. No study yet tests the case that matters most: an end-of-job review against an early error hidden in a polished result.

v12.7.0 — 2026-09-15

New: how the order of the work may shape what AI takes on

Research Briefs opens a new theme, AI and Work, with a brief of Chaining Tasks, Redefining Work: A Theory of AI Automation (Demirer, Horton, Immorlica, Lucier and Shahidi, 2026). The paper argues that AI pays off most on runs of neighboring steps with one human check at the end, so where AI-suitable work sits in a workflow matters, not just how much of it there is. The brief sets out what the evidence can and cannot show, and the authors' ties to Microsoft and Anthropic.

v12.6.3 — 2026-09-14

Updated: survey findings now say only what their sources say

Several Enterprise AI Transformation pages leaned on a single consultancy or analyst survey to claim more than that survey showed. Each such sentence was checked against the source's exact wording, and the ones that overstated it were corrected.

  • Gartner placed generative AI in the trough of disillusionment in its 2025 Hype Cycle for AI, not in 2024. In August 2024 it said gen AI was past the peak.
  • BCG's leading companies show 1.7 times the revenue growth and 1.6 times the EBIT margins of the 60% it calls stagnating or emerging. Their 1.4 times greater cost reductions are expected, not achieved, and the gap is an association, not proof of cause.
  • McKinsey found CEO oversight of AI governance to be one of the elements most correlated with self-reported impact, not the single strongest predictor. That finding is about governing AI systems, not running a transformation program.
  • BCG's 10-20-70 rule is quoted as BCG states it: 70% of strategic focus on people and processes, 20% on technology, 10% on algorithms.
  • Prosci ranks active sponsorship as the top contributor to change success; the pages no longer call it a predictor.
  • An unverifiable Deloitte quote is removed, McKinsey's executive survey is described as covering US executives, and Accenture's strategic-bet finding is described as a correlation.

Pages updated:

v12.6.2 — 2026-09-14

Updated: how many generative-AI projects were abandoned

Several pages still presented Gartner's 2024 forecast, that at least 30% of generative-AI projects would be abandoned after proof of concept by the end of 2025, as a live prediction. That deadline has passed. A January 2026 Gartner article reports that at least 50% were abandoned by the end of 2025, well above the forecast. Gartner gives no sample or method for that figure, and the pages that discuss the evidence say so.

v12.6.1 — 2026-09-14

Corrections: the evidence on redesigning work for AI

Several pages presented McKinsey's 2025 finding on workflow redesign as a cause of AI profit impact. The report's own footnote calls it a correlation, so these pages now describe it that way. They quote its 21% figure as McKinsey wrote it ("fundamentally redesigned at least some workflows"), correct how the MIT study's figures and methods are described, and link to the research note that weighs the evidence.

v12.6.0 — 2026-09-14

New research note: Rethink the Process, or Just Add AI?

A working note on the common advice that organizations should not automate their existing processes with AI, but redesign them around it. It finds that the mild form holds: technology pays more when the organization changes around it. The absolute form does not: two causal studies measured gains from adding AI to an unchanged workflow. It shows that the best-known survey finding behind the advice is a correlation, and that McKinsey's newest edition no longer repeats that analysis. It also looks at what 1990s reengineering taught, including where its famous failure rate came from, and when improving the existing process is the safer bet.

v12.5.0 — 2026-09-13

New research note: What an AI Test Can and Cannot Tell You

A working note on AI evals, the tests organizations run on the AI systems they build or buy. It sets out four findings that hold up: an AI grader has to be checked against people, answer keys contain errors, agent scores are easy to inflate, and small test sets overstate certainty. It also reports what the public record shows. In every published case found, users found the problem after release, and in two of the changes a warning existed before launch and was overruled. It is plain about what nobody has shown: that evals prevent incidents, how common they are outside vendor surveys, who should own them, and what bar is good enough to ship.

v12.4.0 — 2026-09-13

New decision guide: Keeping What Your AI Knows True

When the knowledge an AI assistant or agent uses is wrong or out of date, it repeats it with the same confidence as a right answer. This guide helps the person responsible for what an AI tells people decide how each body of knowledge stays true. It sorts knowledge by whether a machine can check it: code and system settings can be tested, while policies and procedures cannot. It then compares reviewing knowledge when it is used, reviewing it on a schedule, letting AI maintain it, and answering only from reviewed sources. It is plain about the limits: no one has shown which approach works best, and in the two court rulings found, the source content was correct and the AI still got it wrong.

v12.3.0 — 2026-09-13

New decision guide: Who Should Own the AI Bill

Organizations pay for AI in two ways, and the two bills behave differently. Seat licenses, such as Microsoft 365 Copilot, cost the same whether they are used daily or never. Pay-per-use spend shows use, but only to whoever builds the tracking. The guide helps a finance lead or budget holder decide who holds each bill and which control to use: showback, chargeback, a finance-held pool, or per-person allowances. It also says plainly what no one has shown. No study finds that charging business units changes how they use AI, and the guide cites no vendor overrun rates.

AI spend figures updated

Three platform pages now use 2026 evidence. 98% of FinOps practitioners now manage AI spend, but only 26% of large US companies report full, real-time visibility into what their AI costs to run. The first two pages also link to the new guide.

v12.2.1 — 2026-09-13

Statistics refreshed to the latest editions of three recurring studies

Enterprise AI Transformation pages now cite the newest edition of each study they rely on.

  • McKinsey, The State of AI in 2026. 89% of organizations use AI in at least one business function, 37% report any EBIT impact, and about 6% are high performers. 44% of organizations using AI are scaling it across the enterprise, and 54% of companies with $1 billion or more in revenue.
  • BCG, The Widening AI Value Gap (2025). About 5% of companies achieve AI value at scale, and 60% little or none. This replaces the 2024 figure that 26% had moved past proofs of concept. BCG's 10-20-70 rule is now cited to its 2025 restatement, as a rule for where to focus effort.
  • Cisco AI Readiness Index 2025. About 13% of organizations are its most AI-ready "Pacesetters", a share unchanged for three years.
  • One McKinsey figure that could not be confirmed is replaced with BCG's verified finding on measuring AI value.

Pages updated: - Executive Summary - Integrated Assessment - Program Architecture - Sequencing Playbooks - Stakeholder Communication Templates - The 90-Day Launch - AI Strategy Framework - Practitioner Guide: Running an AI Strategy Process - Assessment: Strategic Readiness Scoring - AI Governance Framework - Practitioner Guide: Standing Up an AI Governance Function - AI Readiness Assessment Framework - Building a Data Culture - Technology Architecture Framework - Practitioner Guide: Designing an AI Platform Function - Technology Architecture & Platform - Workflow Optimization Framework - The Four Levels of Workflow AI Integration - Assessment: Workflow Maturity & Opportunity Scoring - AI Adoption Framework - Practitioner Guide: Designing an Adoption Program - Talent & Capability Framework - Assessment: Talent Readiness Scoring - Measurement & Value Realization Framework - Attribution Methodology - Practitioner Guide: Standing Up AI Measurement - Assessment: Measurement Maturity Scoring

v12.2.0 — 2026-09-13

New decision guide: Which AI Model for Which Job

Should a team's recurring AI work run on one model, a model per type of work, a router that picks for each request, or an open-weight model? The guide explains why a listed price is not the real cost, and why being able to switch matters more than any single choice. It ends with a 20-task pilot for testing the decision on your own work.

Technology Architecture Framework corrections

  • Technology Architecture Framework: the routing section now cites its evidence and separates three patterns: assigning a cheaper model to a task, a router choosing per request, and fallback to a second model. It notes that benchmark savings do not carry over cleanly to multi-turn or agent work. The self-hosting guidance drops the claim that suitable open models run on commodity hardware and adds the OECD's volume threshold. The prompt-caching paragraph now cites OpenAI's current guide.

v12.1.1 — 2026-09-13

EU AI Act dates corrected for the 2026 postponement

In July 2026 the EU postponed the AI Act's high-risk obligations. Rules for high-risk AI in sensitive areas such as employment and education now apply from 2 December 2027, not 2 August 2026. Rules for AI built into regulated products, such as lifts or toys, apply from 2 August 2028. Eight Enterprise AI Transformation pages and the glossary still gave the old date. They now carry the current ones, checked against the European Commission's AI Act page.

The dates for the February 2025 bans, the August 2025 general-purpose AI rules, and the August 2026 transparency rules were already right and are unchanged.

v12.1.0 — 2026-09-13

How prioritized AI work actually gets done

Knowledgeville covered what to prioritize and who is accountable for it, but not what happens next. Two new pages in the AI Transformation Lead playbook take a position: once a use case is prioritized, decide who will own the result before you decide who will build it.

  • Staffing the Delivery Team covers the first handoff, from the backlog to the team that builds. Name the owner first. Staff three seats: the owner's delegate, one or two builders, and an architect from the hub. Budget the domain expert's time the way you budget an engineer's, and run only as many teams as you have owners. The page also explains what a vendor means by "pod," and why that differs from the run-state unit in The Three Accountabilities.
  • The Handoff to the Owner covers the second handoff, from that team to whoever runs the result. It explains why AI systems need an owner after launch and lists the five things that must change hands. It gives a test to run before the builders leave, says when a team should stay permanently, and argues for judging a use case a year after launch.

One correction

  • Setting the Operating Model (Chief AI Officer) now describes the portfolio council in its source's own words: business leaders commit sponsorship and benefits ownership, and technology leaders commit capacity. The earlier line paraphrased the source.

v12.0.0 — 2026-08-25

Knowledgeville is now an OKF v0.2 knowledge base

The Open Knowledge Format released v0.2, and this knowledge base has moved to it. The headline change is that provenance is now part of the format itself, not just a local habit.

Every page now carries machine-readable provenance

Open any page's source and you will find, in its frontmatter:

  • generated — which model wrote it, and when it last meaningfully changed
  • verified — who checked its claims against their primary sources, and on what date
  • sources — the sources behind the page, as structured data rather than only a list at the bottom
  • stale_after — a freshness horizon, where every source behind the page carries one

An agent or a person reading a bundle can now ask what has been checked, and by whom, without trusting a summary. Across the published bundle that resolves to 103 pages human-reviewed and 105 unverified — the honest split, not a target. A page with no cited evidence carries no verification claim, because in OKF an absent field means something.

Two pages were wrong, and v0.2 is why

The OKF bundle explains the format, and it had predicted this direction. Its own words:

"There is no standardized verification or trust layer, so a reader cannot check a bundle's claims through the format itself."

That gap is now closed, so both pages that described it have been rewritten:

  • Provenance and Trust — said none of our verification practice was required by the format. The spec has since moved toward it.
  • Where OKF Could Go — its "gaps today" section now records what v0.2 closed and what is still open.

The distinction that survives, and is now the honest open question: the format records that a check happened. It cannot tell you the check was any good. "Verified" remains the publisher's assertion, and a reader still has no way to weigh one publisher's against another's.

Two rendering corrections

A parser fault had been corrupting two live pages:

Both now render as written.

v11.1.2 — 2026-08-24

Corrections

A sweep for unsourced claims turned up a pattern the existing checks could not see: assertions that are load-bearing but contain no number, so nothing flagged them.

Access & Integration now credits Deloitte for a passage that was presented as anonymous consensus. The paragraph describing the data architecture AI leaders are converging on read as "the 2026 consensus" with no source. It is a near-verbatim paraphrase of Deloitte's State of AI in the Enterprise 2026 — a survey of 3,235 leaders across 24 countries — which is now verified and cited.

Two further "2026 consensus" claims are gone. A second passage on that page, and one in Data Mesh Governance in Practice, asserted a consensus sourced to "multiple implementation reports" or to nobody at all. Both now read as what they always were: this framework's recommendation.

AI Governance & Risk drops an invented figure. It claimed most enterprises run governance "three to four years behind" their AI deployments. Nothing supported that number. The point it was making — that governance is routinely stood up after the AI it governs — stands without it.

The AI Governance Framework no longer opens on an unsourced proportion. It asserted what share of enterprise AI is assembled from third-party components. The concrete list that follows makes the case without needing a measured figure.

v11.1.1 — 2026-08-23

Cleanup

Four pages lose a trailing glossary block that pointed nowhere. Lineage & Metadata, Data Governance, Infrastructure Readiness and Security & Compliance each ended with a list of bolded terms telling you to see the glossary, with no link — and the glossary defined none of the terms named. Each page already defines those terms where you meet them, which is where they belong.

The Glossary now links to the pages its terms come from. Its section introductions referred to "Component 02" through "Component 06", numbering from an earlier structure that no longer matched any page.

Corrections

Lineage & Metadata drops two unsourced passages — one asserting what most enterprises had realised by 2026, one asserting what enterprises reporting AI value had done. The AI context layer section now explains what actually distinguishes a context layer from a catalog: an agent queries it mid-task, so it cannot afford the staleness a catalog can.

Its tooling table no longer carries an analyst award in the suitability column. The award was accurate but held by more than one vendor in the same table, and that column describes fit.

v11.1.0 — 2026-08-23

New & expanded content

Expanded — Data Audits & Automated Quality Governance — the audit walkthrough now covers all six quality dimensions instead of two. New steps for consistency, timeliness, validity and uniqueness, each with its audit techniques, a scoring formula and a threshold. A new step covers the composite score: how to pick dimension weights from your use case's failure mode, the formula, a worked example, and the thresholds in one table.

The worked example is deliberately a failing one. A dataset scores 84.0 overall and is still unfit for production AI use, because three dimensions sit below their blocking gate and the composite shows none of that.

Corrections

Data Quality — the six dimensions were credited to the wrong source. They come from a DAMA UK working group paper of October 2013, not from DAMA-DMBOK, which is a separate and broader publication. There is also no canonical list of six: DAMA's Dutch chapter distilled 60 preferred dimension definitions from 127 across nine authoritative sources. The page now says both, cited.

Quality thresholds are now stated as recommendations, not standards. No standards body publishes them. ISO 8000-8:2015 says directly that establishing a threshold is the organisation's job and that the standard does not set one. Both pages now say so and tell you to set your own against the consequence of failure, and to write down the reasoning.

A contradiction between the two pages is resolved. The framework page blocked any dataset scoring below 70 on a dimension, while the audit guide blocked accuracy below 90. The audit guide's per-dimension gates are now the single set, and the scoring section links to them — previously it named a scoring approach with no way to reach the method behind it.

v11.0.0 — 2026-08-10

New & expanded content

New bundle — Research Notes — working notes on questions that are still open. Everything else here is finished and meant to be relied on; these are sourced findings published while the conclusion is still unsettled. Every number is still cited and verified.

New note — Why Your AI Gives Three Different Answers — three executives ask an AI assistant how many active customers the company has and get three numbers, because Finance, Sales and Support each define "active" correctly for their own job. On what the word ontology is currently sold to solve, which version of that work has a track record, and where the evidence runs out.

Two corrections it records, both against primary sources: the Open Semantic Interchange specification's first release was v0.1, not the v1.0 several write-ups report, and it has since moved to the Apache Incubator as Apache Ossie.

v10.4.0 — 2026-08-04

New & expanded content

New brief — The Talent Obsession

A read of TechCrunch's report that enterprise AI's bottleneck has moved from the model to the engineer who deploys it, and that only about 2,000 US engineers qualify.

The brief separates two layers the coverage blurs. The structural claim is checkable elsewhere: two frontier labs stood up services arms. The figures are not — every one comes from the executive search firm that places these engineers, from a study whose data sources are named but whose full text sits behind a request form. The shortage turns out to be a definition rather than a count, and the firm setting that threshold sells the scarce thing.

Companion to When to Embed an Engineer, which covers the hiring decision itself.

v10.3.0 — 2026-08-04

New & expanded content

New guide — When to Embed an Engineer

For leaders whose AI pilots demo well and stall before production, and who have been told to hire a "forward deployed engineer." The guide is built around the decision rather than the job title: embed an engineer, hire a firm, or fix the handoff.

It covers the scope condition that makes embedding worth its cost, why deep context usually beats a fast start for internal work, who maintains what gets built, and the compensation and reporting-line traps that surface only after a search has already run. Every source quoted sells embedding in some form, and the guide says so at each point of use.

v10.2.0 — 2026-07-11

New & expanded content

  • Author The Great Decoupling industry brief

v10.1.0 — 2026-07-11

New & expanded content

  • Author MCP and the Context Gap

Citation corrections

  • Correct overstated MCP-governance-is-unprecedented claim

v10.0.0 — 2026-07-07

New & expanded content

  • Author Why Dumping a File Rarely Works for AI-Assisted Data Analysis
  • Author Preparing Your Data Before You Ask for AI-Assisted Data Analysis
  • Author Code Interpreters vs. Plain Chat
  • Author A Landscape of AI Data Analysis Tools, completing track 1
  • Author Asking Good Analytical Questions
  • Author From Raw Numbers to a Defensible Conclusion
  • Author Where AI Fabricates Data and Numbers
  • Author Silent Arithmetic and Logic Errors, completing track 2
  • Author Building a Method, Not a Lucky Prompt
  • Author Testing Your Method on a New Dataset, completing the bundle
  • Refine and correct four pages in AI-Assisted Data Analysis after a full editorial review of the finished bundle

Site & structure

  • Scaffold AI-Assisted Data Analysis bundle
  • Reframe AI-Assisted Data Analysis from 5 tracks to 3
  • Surface AI-Assisted Data Analysis on the homepage

v9.5.0 — 2026-07-06

New & expanded content

  • New bundle: AI-Assisted Software Development. The two real questions a developer or team lead faces with AI coding agents — does it actually make you faster, and does it open you up to new attacks. Covers the contested productivity evidence (a randomized controlled trial found AI coding tools slowed experienced developers down even as they believed it sped them up, set against vendor-reported speedups on simpler tasks) in Does AI Coding Actually Help?, and the named, documented vulnerabilities and supply-chain campaigns already targeting AI coding assistants in Securing the AI Coding Workflow. (#247, #248, #249)

v9.4.0 — 2026-07-06

New & expanded content

  • Name and map competing agentic-AI framework acronyms in workflow-optimization guide

v9.3.1 — 2026-07-06

No reader-facing knowledge changes in this release — wording polish only.

v9.3.0 — 2026-07-06

New & expanded content

  • Who Owns the Codified Judgment (ai-native-organization) — a new page asking who owns a firm's judgment once it's codified well enough for an AI agent to apply it, and who's exposed when a firm buys that judgment rather than building it in-house. https://github.com/grandaha/knowledgeville-workspace/pull/240
  • Verify AI's Facts and Citations (everyday-tasks) — a new recipe for checking an AI's claims and sources before relying on them, grounded in real hallucination-rate studies rather than a generic "always verify" caution. https://github.com/grandaha/knowledgeville-workspace/pull/241
  • Build vs. Buy (decision guides) adds a new section on AI systems specifically — vendor dependency even when fine-tuning, per-token billing instead of a flat subscription, and the ongoing burden of evaluating factually wrong output that traditional software doesn't carry. https://github.com/grandaha/knowledgeville-workspace/pull/235
  • Explain a Concept Simply (everyday-tasks) — the prompt now tells the AI to say so instead of forcing an analogy when none fits, and to flag anything experts still debate; the caveat now names the actual risk (a fluent but wrong analogy on a topic that feels like general knowledge but is actually contested). https://github.com/grandaha/knowledgeville-workspace/pull/239
  • Building Yours Without an Engineer (team second brain) now states plainly it's an untested design, not a validated practice — the setup steps stay, but the page no longer implies this has been proven at team scale. https://github.com/grandaha/knowledgeville-workspace/pull/234

v9.2.1 — 2026-07-05

New & expanded content

  • Rewrite Knowledgeville welcome intro in Dave's voice
  • Add Team Second Brain to the homepage bundle list

Site & structure

  • Fix team-second-brain page order to match narrative sequence

v9.2.0 — Team Second Brain bundle — 2026-07-05

New & expanded content

  • Author The Single-Player Ceiling as a decision-aid reference
  • Author Building Yours Without an Engineer, completing team-second-brain

Citation corrections

  • Correct overstated single-user claim on team-second-brain lead page
  • Add GitHub ecosystem evidence as phase-one history

Site & structure

  • Scaffold the Team Second Brain bundle

v9.1.0 — Data Readiness reframed around the use case — 2026-07-03

New & expanded content

  • Open the Data Readiness framework with a use-case diagnostic
  • Re-scope the AI Readiness Assessment to the use case, not the org
  • Reposition Data Readiness framework pages around the use case
  • Relabel practitioner guides around use-case scope

v9.0.0 — The AI-Native Organization bundle — 2026-07-03

New & expanded content

  • Author The Three Accountabilities and The Partnership Pyramid, Generalized

Site & structure

  • Scaffold ai-native-organization bundle

v8.1.0 — 2026-07-03

New & expanded content

  • Connect Enterprise AI Transformation's roles to AI Accountability

v8.0.1 — 2026-07-03

Site & structure

  • Reorder AI Accountability roles into hierarchy, tighten overview

v8.0.0 — 2026-07-03

New & expanded content

  • Chief AI Officer bundle (operator's manual for the AI strategy owner)
  • AI Transformation Lead bundle (operator's manual for the program owner)
  • AI Risk Officer bundle (operator's manual for model-level risk)
  • AI Governance Lead bundle (operator's manual for AI policy operationalization)
  • AI Security Lead bundle (operator's manual for model-level security)

Site & structure

  • Restructure ai-enablement under a new ai-accountability container

v7.0.0 — AI Enablement bundle — 2026-07-02

New bundle: AI Enablement

The operator's manual for the person accountable for AI adoption inside an organization — the AI Enablement Lead. A new top-level bundle, sibling to Enterprise AI Transformation (which covers the program at leadership altitude); this is the enablement seat, deep.

Eight pages: - What AI Enablement Is — and the Role — the discipline, why adoption is the binding constraint, and the role/title taxonomy. - Who's Accountable for What — the shared-accountability model up the chain, so you are not made accountable for adoption you cannot control. - Standing Up the Function — Center of Excellence models, the AI Council, funding, and decision rights. - Finding Where AI Earns Its Keep, Building the Champion Network, Enablement That Sticks, Measuring Adoption, and When It Stalls — the operator's how-to arc.

Grounded in primary sources (IBM CEO Study 2026, Microsoft AI Center of Excellence guidance), citation-audited, and dual-gated.

v6.1.1 — 2026-07-02

Site & structure

  • Consistent navigation: each area now leads with its "start here" page (what it is / how it works), instead of that page landing wherever it fell alphabetically. Fixes, for example, the Decision Guides order so How These Guides Work comes before the guides.

v6.1.0 — 2026-07-02

New & expanded content

  • Decision Guides — When to Let AI Do It — a new guide on whether to hand a task to AI or do it yourself: where AI earns its keep, where to keep your hands on the work, and how to decide task by task. Anchored in primary agent-guidance on when to keep a human in the loop (sensitive, irreversible, or high-stakes work).

v6.0.0 — 2026-07-02

New & expanded content

  • Decision Guides — a new area of cited, plain-language guides for the recurring, consequential decisions a knowledge worker faces: the real options, the criteria that matter, the tradeoffs with sources, and when each choice wins. Built to be used with your AI assistant — connect it and your agent works from grounded, sourced criteria instead of a generic pros-and-cons list.
  • First guide: Build vs. Buy — whether to build a capability in-house or buy it off the shelf, anchored in primary-source evidence (McKinsey–Oxford on IT-project risk, the software-maintenance cost share, and Geoffrey Moore's core-vs-context).

v5.4.1 — 2026-07-02

Site & structure

  • Reordered the top-level navigation so Enterprise AI Transformation sits directly below What's Changed and above AI Briefs.

v5.4.0 — 2026-07-02

New & expanded content

  • AI Briefs — Exposed but Unprepared — a new brief opening the Workforce and Skills theme in Industry Briefs. It reads an HFS Research study finding 27 million Global 2000 jobs exposed to AI over three years, and argues the real problem is how few of those companies (only 14%) have a coherent plan. The brief deliberately defuses the headline: "exposed" bundles job loss with job redesign, the figures are one firm's estimate, and the firm sells the transition it describes — while keeping the useful signal that unprepared companies are cutting hardest and hiring least.

v5.3.2 — 2026-07-02

Site & structure

  • The homepage now speaks in the first person — Knowledgeville is written and shared by me, still linking out to One Step Labs.
  • Refreshed the AI Use Cases summary to include Realtors and the new AI Playbooks, and to stay current as more sub-domains are added.

v5.3.1 — 2026-07-01

Site & structure

  • Tidied the AI Briefs navigation: the two areas now read Research Briefs and Industry Briefs under the AI Briefs heading, instead of repeating "AI" in each.

v5.3.0 — 2026-07-01

Site & structure

  • The two brief bundles are now one: AI Briefs. Research Briefs (papers) and Industry Briefs (analyst and market commentary) now live together under a single AI Briefs area, instead of two separate top-level entries. Each keeps its own focus and sourcing; the top level is simpler. Old links are redirected to the new locations, so existing bookmarks keep working.

v5.2.1 — 2026-07-01

Fixes

  • Tightened source fidelity across three recent additions: restored a verbatim quote of the OKF specification, corrected two citation/attribution details in the Aaron Levie industry brief, and reworded an AI Playbooks setup note so it no longer claims a capability for specific named products.

v5.2.0 — 2026-07-01

New & expanded content

  • AI Use Cases — AI Playbooks — a new sub-domain of agent-executable playbooks: procedures an AI assistant reads and runs on its own across your email and calendar, rather than prompts you paste. Six to start — Morning Brief, Inbox Triage, Meeting Prep, Weekly Review, Follow-up Tracker, and End-of-Day Wrap — each with numbered steps for the assistant, an example of what it produces, and read-first guardrails (it reads and drafts, but never sends, deletes, or changes anything).

v5.1.0 — 2026-07-01

New & expanded content

  • AI Industry Briefs — Enterprise AI Is Not a Startup — a new Market and Adoption brief on Aaron Levie's (Box CEO) argument, via HFS Research: enterprise AI is constrained by workflow, data, and accountability — not model quality — and the real transformation lives in the ~88% of the economy that does not run like a tech startup. Levie's figures are flagged as forward projections, with both the CEO's and the analyst's commercial stakes noted.

v5.0.2 — 2026-07-01

Fixes

  • Industry brief source links now show the full source URL instead of just the domain, so the exact article a brief is based on is visible at a glance.

v5.0.1 — 2026-07-01

Site & structure

  • Refreshed the homepage Where to start guide to route by intent across the current knowledge base — begin with the Open Knowledge Format, copy a ready-to-run recipe, read the enterprise model, follow the research and industry briefs, or connect the whole base over MCP.

v5.0.0 — 2026-07-01

New & expanded content

  • Open Knowledge Format — a new bundle explaining OKF, the open format this whole knowledge base is built on, for knowledge and strategy leaders. Two arcs across seven pages:
  • Understanding OKF — what OKF is, how it works, where it fits against RAG, knowledge graphs, and plain docs, and how it handles provenance.
  • The Potential — our grounded read of where it could go: knowledge for agents, the honest gaps in v0.1, and whether the direction fits you.

Facts are grounded in the canonical OKF specification; the forward-looking half is our analysis, marked as such.

Site & structure

  • Open Knowledge Format now leads the knowledge base — first in the bundle list on the homepage and first in the site navigation, since it is the format everything else is built on.

v4.2.0 — 2026-07-01

New & expanded content

  • AI Industry Briefs — The Great Value Migration — a new brief, and a new Strategy & Investment theme in the AI Industry Briefs. It's a plain-language read of an investor's thesis that as AI gets cheap, value shifts from selling compute to owning and orchestrating it — favoring the big cloud platforms. Every borrowed figure is traced to its original source, and the author's disclosed stock positions are flagged as stake, not advice.

Site & structure

  • Connect over MCP — the homepage now shows a third way to use Knowledgeville: point any AI assistant at the live knowledge base over the Model Context Protocol, at https://www.onesteplabs.com/knowledgeville/mcp. Alongside reading it as a site and cloning the GitHub repo.
  • The AI Industry Briefs area is now listed on the homepage bundle index.

v4.1.0 — 2026-07-01

New & expanded content

  • Realtors — a new set of plain-language, paste-ready AI recipes for real-estate agents, joining Everyday Tasks and Executive Leadership under AI Use Cases. Nine recipes across three areas:
  • Listings & Marketing — write a listing description, just-listed/just-sold social posts, and an open-house promo kit.
  • Prospecting & Outreach — neighborhood farming letters, expired and FSBO scripts, and referral asks to your sphere.
  • Client Communication — buyer and seller follow-ups, plain-language explainers for offer terms, and review requests.

Each recipe hands you a prompt, shows a real before-and-after, and flags where an agent must keep control — fair-housing language, do-not-call and texting rules, and the line between explaining a term and giving legal advice.

Site & structure

  • New Realtors area under AI Use Cases, with its own three-section map.

v4.0.0 — 2026-06-30

New & expanded content

  • AI Industry Briefs — a new area of plain-language briefs on what analysts and the market are saying about AI, written for business leaders. The companion to AI Research Briefs: where those cover academic papers, these cover analyst, vendor, and market commentary.
  • The AI Adoption Gap — the first brief, a read of an HFS Research thesis on why AI's market valuations have run ahead of real enterprise adoption, and what leaders should do about it.

Site & structure

  • New top-level AI Industry Briefs section, opening with its first theme, Market and Adoption. Each brief states plainly who is making a claim and what they stand to gain, so you can weigh it yourself.

v3.0.0 — 2026-06-29

New & expanded content

  • AI Research Briefs — a new bundle of plain-language briefs of notable AI research papers, written for business leaders and knowledge workers: what each paper found, how much to trust it, and what it means for you.
  • There's No Silver Bullet for Coding-Agent Rewards — why no single reward keeps a coding agent honest, and what a layered verification stack does instead.
  • Why AI Drops Its Caution the Moment You Ask for Advice — AI models assert cause-and-effect when asked for advice that they'd carefully hedge under analysis, and a one-line re-prompt brings the caution back.

Site & structure

  • New AI Research Briefs area on the site, grouped by theme (Agents & Tool Use, Reasoning & Capabilities), with a guide to how each brief is structured and how to read its trust assessment.

v2.1.1 — 2026-06-29

New & expanded content

  • Plain-language readability pass on everyday-tasks (20-page audit)

v2.1.0 — 2026-06-28

New & expanded content

  • Executive-leadership sub-domain bundle (design + scaffold, 16 planned stubs)
  • Author draft-a-board-update (first executive-leadership page)
  • Author prep-board-investor-qa (first workflow-shape page)
  • Author remaining 14 executive-leadership pages (bundle complete 16/16)
  • Surface Executive Leadership on the homepage

v2.0.0 — 2026-06-25

New & expanded content

  • Reframe a top-level domain as a "knowledge bundle", not a "framework"
  • Ai-use-cases domain + everyday-tasks bundle (20 demo-first pages)
  • Surface AI Use Cases on the homepage

Site & structure

  • Add the Maturity Self-Assessment tool to the site nav

v1.1.0 — 2026-06-22

New & expanded content

  • Interactive Maturity Self-Assessment tool, generated from the assessments (#88)

Structural & tooling

  • SemVer releases + de-emoji the changelog
  • Exclude overview.md from the distributed bundle (OKF alignment)

v1.0.0 — 2026-06-22

Initial release of the Knowledgeville knowledge base.

Frameworks in this release

Enterprise AI Transformation — a complete eight-track framework for taking an organization from AI ambition to measured value, plus a Running the Program section that ties the tracks together into an assessment, a sequencing model, and a 90-day launch:

  • AI Strategy & Leadership
  • AI Governance & Risk
  • Data Readiness
  • Technology Architecture & Platform
  • Workflow Optimization & Automation
  • AI Adoption & Culture
  • Talent & Capability Building
  • Measurement & Value Realization

Every statistic on every page is backed by a verified primary source in the evidence ledger, traceable in two clicks, and enforced in CI.

Read it at https://knowledge.onesteplabs.com/enterprise-ai-transformation/.

What's next

Future releases record what changed since the last one — grouped as new & expanded content, citation corrections, freshness updates, and structural & tooling changes — so returning readers can see what's new as more domains and topics are added.