Decision log

Format: ID, date, decision, rationale. Newest last. Decisions are revisited only with a new entry superseding the old one.

Recorded decisions

ID	Date	Decision	Rationale
D001	2026-08-19	Research corpus lives in this dedicated public repo	Clean separation from the maintainer's personal site; public from day one so others can contribute; becomes the content source for the published site
D002	2026-08-19	Publish as a dedicated docs site on Vercel; primary candidate Fumadocs (Next.js), fallback Astro Starlight after a 1-day spike	Keeps the maintainer's familiar Next + MDX + Vercel workflow while adding sidebar navigation, TOC, search, and hierarchical routing that a reference guide needs
D003	2026-08-19	Featured verticals: Utilities & Energy, Banking & Financial Services, Manufacturing & Supply Chain, Public Sector	Maintainer domain depth (utilities), richest regulatory material (BFSI), the literal factory analogy (manufacturing), underserved audience (public sector)
D004	2026-08-19	Layer research runs pilot-first; pilot track is R03 Integration Fabric	R03 has the richest current state (SOA/API/EDI), the most concrete agentic change (MCP, gateways, agent-to-agent), and touches vendors, economics, and security, so it exercises the full template before fan-out
D005	2026-08-19	Dual license: CC BY-SA 4.0 for content, MIT for code	Content stays open and attributable with share-alike protection; site code stays maximally reusable
D006	2026-08-19	Working protocol: every phase and track passes a maintainer POV step and a technical challenge round before finalization	The repo serves readers, free of vendor bias and author bias; research supplies evidence, the maintainer supplies technical challenge, only what survives both is published
D007	2026-08-19	Cross-cutting concern register (C1 to C10) is mandatory in every track's findings	Identity, observability, traceability, grounding, impersonation, sovereignty, privacy, safety, cost, and resilience must be addressed (or justified N/A) at every layer, not in one chapter
D008	2026-08-19	Challenged Defaults is a mandatory section in every track; first named debate is graph databases vs topic modelling with contextual embeddings as the retriever	Anti-buzz mechanism: recommendations must beat the simplest credible alternative on evidence and economics
D009	2026-08-19	Third-party IP is not committed: extracted structure notes and citations only, with permission status flagged	The practitioner reference deck requires author permission before attribution or publication; vendor slides and unpublished drafts are referenced, not redistributed
D010	2026-08-19	The guide adopts the industry definition of an agent (a model using tools in a loop) and treats learning as an orthogonal axis, stated as: without a learning loop you have a fixed-policy agent, not a teammate	Evidence review found no mainstream definition requiring learning; relocating the authors' learning conviction to the maturity model's second axis is both honest and differentiating, since surveyed maturity models omit learning
D011	2026-08-19	Adoption sequencing is presented two-track: what the market does first (observed data) and what this guide recommends (authors' judgment), each labeled	Adoption data contradicted parts of the recommended order; presenting observation and recommendation separately preserves both truthfulness and practitioner guidance
D012	2026-08-19	Production case studies are referenced generically; system internals and in-progress research remain private until fully tested	The guide's credibility rests on published, verifiable evidence; proprietary implementation detail adds risk without adding verifiable value, and unproven research must not read as established practice
D013	2026-08-19	The deterministic boundary is stated as "models may inform, never decide": deterministic decision rules over verifiable credentials in the four irreversible zones, with model outputs as advisory inputs	Evidence showed three of the four zones already run probabilistic signals internally (identity risk scoring, payment fraud scoring, AI-assisted drafting) while authorization stays rule-bound; the agentic payment stack built independently by three networks is the convergent proof
D014	2026-08-19	Compliance posture is tiered: an evidence floor for every production agent, Article-12-grade instrumentation for the tier that could plausibly classify high-risk	The EU Digital Omnibus moved high-risk obligations to December 2027, so deadline-driven urgency is gone, but classification is now an evidence question, standards lead times run 12+ months, and sectoral regulators are moving faster than the horizontal law
D015	2026-08-19	Identity levels are assigned by capability surface, not by platform enrollment	Per-user (not per-agent) licensing removed the cost brake on presence identity, so control-plane registration no longer indicates the level of exposure; mailbox, meeting seat, and directory visibility are what create phishing and discoverability surface
D016	2026-08-19	Platform choice at the systems-of-record layer is governed by where the record and its permission model live, not by use-case category; agents reach records through tool servers wrapping governed APIs, and write access follows a trust progression rather than a policy switch	Four of five major vendors shipped external-agent paths with stated permission parity in H1 2026, collapsing the assumption that embedded means governed and external does not; the remaining differences are model control, cost meter, and customization ceiling
D017	2026-08-19	Practitioner patterns are published with their evidence status stated individually, including "design proposal, no published precedent" where that is the honest finding	Testing the four utilities patterns produced a split verdict: one has a measured result, one has shipped analogues, and two have no published production precedent anywhere; publishing them uniformly would have misrepresented all four
D018	2026-08-19	Rule promotion gates on counterexample survival and eval regression rather than frequency, lands promoted artifacts outside the model, and requires a demotion path	2026 research inverted the assumption that generation is easy and promotion is the control point: only a fifth of automatically learned policies were usable and automated gates erred in both directions, while roughly three quarters of policy statements depend on context that cannot be predefined
D019	2026-08-19	Identity tiers are assigned by capability surface, with the accountable-human sponsor recorded at ID2 access identity rather than ID3 presence identity	The platform architecture already carries a sponsor attribute on the agent identity itself, and agents are nameable and mentionable in collaboration surfaces without a user account, so requiring a mailbox to obtain accountability inverts the control
D020	2026-08-19	Customer-facing agents get a separate lane and a separate edge on a shared control plane, justified by legal exposure rather than by technology	Customer-facing failures are legally attributable, contractually binding and publicly visible while internal failures are contained; duplicating knowledge, identity, tools, evaluation and observability across the two is a defect rather than a safeguard
D021	2026-08-19	The founding metaphor is defended on the shape of the residual work, never on a headcount ratio or a timeline, and the widely circulated lights-out crew figure is corrected in the text	The residual task set has been stable in the human-factors literature since 1983, but the canonical automated-plant figure is misquoted in secondary sources, no supervision ratio for agents has ever been published, and the strongest administrative-data study finds task reorganisation rather than crew reduction
D022	2026-08-19	A4 and above carry an oversight-capacity gate expressed as a burst rate, added as a precondition on the autonomy axis rather than as a third axis	No credible human-to-agent supervision ratio has ever been published, so an autonomy ladder that assumes one rests on nothing; modelling oversight as a separate axis would imply enterprises can trade autonomy against oversight, which is the trade the gate exists to forbid
D023	2026-08-19	The master target state is published as three complete architectures keyed to size and regulatory intensity, with vendor gravity treated as an interface question rather than a structural one	Two enterprises of the same size and regulatory intensity build the same shape and wire it to different systems; publishing gravity variants would multiply maintenance while changing only the connector inventory
D024	2026-08-19	The economics model is two-level: cost per run reconciles to the invoice, cost per resolved outcome decides the business case, and the second includes supervision labour and the cost of wrong outcomes	Platforms meter per run and businesses decide per outcome; every documented AI-first reversal set a containment target, and supervision labour is absent from published business cases despite being a measured capacity constraint
D025	2026-08-19	Use-case prioritisation sorts on evaluability as a gate rather than as a weighted criterion, with valuable-but-unmeasurable use cases funded into a make-it-evaluable queue instead of being rejected or built	A weighted criterion lets a high-value unmeasurable use case outrank a measurable one, which removes the basis for the eval suite, the promotion gate and every reliability claim the maturity model rests on; the failure data attributes programme failure to the learning gap rather than to model quality
D026	2026-08-19	The roadmap is published as a nine-factor generator over a six-stage spine with entry and exit gates and no durations, resolved by a more-restrictive-wins conflict rule	Named variants cannot represent the factor combinations enterprises actually present; gates are defensible from evidence while durations are not, and an explicit conflict rule is what stops a stated risk appetite from cancelling a regulatory constraint
D027	2026-08-19	Supervision capacity is treated as a portfolio-level constraint, not a per-workload one	Workloads are approved individually and supervised collectively; five A3 workloads each consuming a third of one supervisor's burst capacity do not fit inside one supervisor, and the arithmetic is never performed because no single approval exceeds the limit
D028	2026-08-19	Every blueprint carries a mandatory honest-limits section naming what does not work, what has no published precedent, and what the guide refuses to claim	Blueprints are the most quotable artifact in a guide and the easiest to read as endorsement; the utilities blueprint has the longest limits section in the corpus and that length is itself the finding
D029	2026-08-19	Blueprints name agents by function and never by product, and state an A x L position for each	Product names date within a quarter and imply a recommendation the vendor phase has not yet made; an A x L position makes the autonomy and learning claim explicit and testable against the oversight gate
D030	2026-08-19	The vendor scorecard runs disqualifiers before scores, and publishes a section profile rather than a total	A weighted total lets a vendor compensate for an unacceptable answer with strength elsewhere, which is how platforms with no enforcement boundary win evaluations on ecosystem depth
D031	2026-08-19	Where the research cannot establish a vendor's answers to the question bank, the profile states the gap rather than inferring answers from adjacent products, even where that leaves a prominent vendor with the thinnest profile in the hub	The hub's value is that its claims are checkable; inferring an enterprise platform's identity, memory and metering posture from a published SDK would make the hub's least-sourced section indistinguishable from its best-sourced one
D032	2026-08-19	The site is a generated view of the repository markdown rather than a separate content tree, with `site/content/docs` git-ignored	Two copies of the corpus would diverge on the first correction; generation keeps the repository the single source of truth (D001) and makes the site's link graph, titles and ordering derivable rather than maintained
D033	2026-08-20	D001's public-repository assumption is suspended while publication-held paths remain tracked; the canonical repository stays private and the site publishes only the generated public view	A site filter cannot remove a mechanism from public Git history. The repository can become public only after the hold is released, or a separate publishable-only mirror with no shared history is created
D034	2026-08-23	The multi-card retrieval experiments are released on the site as a research section, negative results included, and the objective-conditioning and consumer-dependent-diversity claims are withdrawn from the guide; the mechanism-level technique entry stays under the publication hold and the research pages are the implementation guidance	Five rounds on public corpora with human judgements settled the claims the guide had carried as hypotheses, and the guide's credibility rests on publishing what was tested rather than what was hoped; the research pages describe experiments on public data only, so D012's rule that system internals stay private is unchanged, and D033's suspension of the public-repository assumption stands until the hold is released or a publishable-only mirror is created
D035	2026-08-23	The maintainer will not pursue patent protection for the multi-card retrieval mechanisms; the publication hold is released (its register stays as the build's mechanism, empty), the technique entry publishes with a status note, and all programme material may be shared; D033's suspension of D001 ends, and making the canonical repository public is the maintainer's own action	Five rounds of experiments found no supported technical effect beyond the corpus-derived representation that prior art already covers, so there is nothing left to protect, and the guide's value lies in publishing the evidence, the reversals, and the approach it supports

ID: D001
Date: 2026-08-19
Decision: Research corpus lives in this dedicated public repo
Rationale: Clean separation from the maintainer's personal site; public from day one so others can contribute; becomes the content source for the published site

ID: D002
Date: 2026-08-19
Decision: Publish as a dedicated docs site on Vercel; primary candidate Fumadocs (Next.js), fallback Astro Starlight after a 1-day spike
Rationale: Keeps the maintainer's familiar Next + MDX + Vercel workflow while adding sidebar navigation, TOC, search, and hierarchical routing that a reference guide needs

ID: D003
Date: 2026-08-19
Decision: Featured verticals: Utilities & Energy, Banking & Financial Services, Manufacturing & Supply Chain, Public Sector
Rationale: Maintainer domain depth (utilities), richest regulatory material (BFSI), the literal factory analogy (manufacturing), underserved audience (public sector)

ID: D004
Date: 2026-08-19
Decision: Layer research runs pilot-first; pilot track is R03 Integration Fabric
Rationale: R03 has the richest current state (SOA/API/EDI), the most concrete agentic change (MCP, gateways, agent-to-agent), and touches vendors, economics, and security, so it exercises the full template before fan-out

ID: D005
Date: 2026-08-19
Decision: Dual license: CC BY-SA 4.0 for content, MIT for code
Rationale: Content stays open and attributable with share-alike protection; site code stays maximally reusable

ID: D006
Date: 2026-08-19
Decision: Working protocol: every phase and track passes a maintainer POV step and a technical challenge round before finalization
Rationale: The repo serves readers, free of vendor bias and author bias; research supplies evidence, the maintainer supplies technical challenge, only what survives both is published

ID: D007
Date: 2026-08-19
Decision: Cross-cutting concern register (C1 to C10) is mandatory in every track's findings
Rationale: Identity, observability, traceability, grounding, impersonation, sovereignty, privacy, safety, cost, and resilience must be addressed (or justified N/A) at every layer, not in one chapter

ID: D008
Date: 2026-08-19
Decision: Challenged Defaults is a mandatory section in every track; first named debate is graph databases vs topic modelling with contextual embeddings as the retriever
Rationale: Anti-buzz mechanism: recommendations must beat the simplest credible alternative on evidence and economics

ID: D009
Date: 2026-08-19
Decision: Third-party IP is not committed: extracted structure notes and citations only, with permission status flagged
Rationale: The practitioner reference deck requires author permission before attribution or publication; vendor slides and unpublished drafts are referenced, not redistributed

ID: D010
Date: 2026-08-19
Decision: The guide adopts the industry definition of an agent (a model using tools in a loop) and treats learning as an orthogonal axis, stated as: without a learning loop you have a fixed-policy agent, not a teammate
Rationale: Evidence review found no mainstream definition requiring learning; relocating the authors' learning conviction to the maturity model's second axis is both honest and differentiating, since surveyed maturity models omit learning

ID: D011
Date: 2026-08-19
Decision: Adoption sequencing is presented two-track: what the market does first (observed data) and what this guide recommends (authors' judgment), each labeled
Rationale: Adoption data contradicted parts of the recommended order; presenting observation and recommendation separately preserves both truthfulness and practitioner guidance

ID: D012
Date: 2026-08-19
Decision: Production case studies are referenced generically; system internals and in-progress research remain private until fully tested
Rationale: The guide's credibility rests on published, verifiable evidence; proprietary implementation detail adds risk without adding verifiable value, and unproven research must not read as established practice

ID: D013
Date: 2026-08-19
Decision: The deterministic boundary is stated as "models may inform, never decide": deterministic decision rules over verifiable credentials in the four irreversible zones, with model outputs as advisory inputs
Rationale: Evidence showed three of the four zones already run probabilistic signals internally (identity risk scoring, payment fraud scoring, AI-assisted drafting) while authorization stays rule-bound; the agentic payment stack built independently by three networks is the convergent proof

ID: D014
Date: 2026-08-19
Decision: Compliance posture is tiered: an evidence floor for every production agent, Article-12-grade instrumentation for the tier that could plausibly classify high-risk
Rationale: The EU Digital Omnibus moved high-risk obligations to December 2027, so deadline-driven urgency is gone, but classification is now an evidence question, standards lead times run 12+ months, and sectoral regulators are moving faster than the horizontal law

ID: D015
Date: 2026-08-19
Decision: Identity levels are assigned by capability surface, not by platform enrollment
Rationale: Per-user (not per-agent) licensing removed the cost brake on presence identity, so control-plane registration no longer indicates the level of exposure; mailbox, meeting seat, and directory visibility are what create phishing and discoverability surface

ID: D016
Date: 2026-08-19
Decision: Platform choice at the systems-of-record layer is governed by where the record and its permission model live, not by use-case category; agents reach records through tool servers wrapping governed APIs, and write access follows a trust progression rather than a policy switch
Rationale: Four of five major vendors shipped external-agent paths with stated permission parity in H1 2026, collapsing the assumption that embedded means governed and external does not; the remaining differences are model control, cost meter, and customization ceiling

ID: D017
Date: 2026-08-19
Decision: Practitioner patterns are published with their evidence status stated individually, including "design proposal, no published precedent" where that is the honest finding
Rationale: Testing the four utilities patterns produced a split verdict: one has a measured result, one has shipped analogues, and two have no published production precedent anywhere; publishing them uniformly would have misrepresented all four

ID: D018
Date: 2026-08-19
Decision: Rule promotion gates on counterexample survival and eval regression rather than frequency, lands promoted artifacts outside the model, and requires a demotion path
Rationale: 2026 research inverted the assumption that generation is easy and promotion is the control point: only a fifth of automatically learned policies were usable and automated gates erred in both directions, while roughly three quarters of policy statements depend on context that cannot be predefined

ID: D019
Date: 2026-08-19
Decision: Identity tiers are assigned by capability surface, with the accountable-human sponsor recorded at ID2 access identity rather than ID3 presence identity
Rationale: The platform architecture already carries a sponsor attribute on the agent identity itself, and agents are nameable and mentionable in collaboration surfaces without a user account, so requiring a mailbox to obtain accountability inverts the control

ID: D020
Date: 2026-08-19
Decision: Customer-facing agents get a separate lane and a separate edge on a shared control plane, justified by legal exposure rather than by technology
Rationale: Customer-facing failures are legally attributable, contractually binding and publicly visible while internal failures are contained; duplicating knowledge, identity, tools, evaluation and observability across the two is a defect rather than a safeguard

ID: D021
Date: 2026-08-19
Decision: The founding metaphor is defended on the shape of the residual work, never on a headcount ratio or a timeline, and the widely circulated lights-out crew figure is corrected in the text
Rationale: The residual task set has been stable in the human-factors literature since 1983, but the canonical automated-plant figure is misquoted in secondary sources, no supervision ratio for agents has ever been published, and the strongest administrative-data study finds task reorganisation rather than crew reduction

ID: D022
Date: 2026-08-19
Decision: A4 and above carry an oversight-capacity gate expressed as a burst rate, added as a precondition on the autonomy axis rather than as a third axis
Rationale: No credible human-to-agent supervision ratio has ever been published, so an autonomy ladder that assumes one rests on nothing; modelling oversight as a separate axis would imply enterprises can trade autonomy against oversight, which is the trade the gate exists to forbid

ID: D023
Date: 2026-08-19
Decision: The master target state is published as three complete architectures keyed to size and regulatory intensity, with vendor gravity treated as an interface question rather than a structural one
Rationale: Two enterprises of the same size and regulatory intensity build the same shape and wire it to different systems; publishing gravity variants would multiply maintenance while changing only the connector inventory

ID: D024
Date: 2026-08-19
Decision: The economics model is two-level: cost per run reconciles to the invoice, cost per resolved outcome decides the business case, and the second includes supervision labour and the cost of wrong outcomes
Rationale: Platforms meter per run and businesses decide per outcome; every documented AI-first reversal set a containment target, and supervision labour is absent from published business cases despite being a measured capacity constraint

ID: D025
Date: 2026-08-19
Decision: Use-case prioritisation sorts on evaluability as a gate rather than as a weighted criterion, with valuable-but-unmeasurable use cases funded into a make-it-evaluable queue instead of being rejected or built
Rationale: A weighted criterion lets a high-value unmeasurable use case outrank a measurable one, which removes the basis for the eval suite, the promotion gate and every reliability claim the maturity model rests on; the failure data attributes programme failure to the learning gap rather than to model quality

ID: D026
Date: 2026-08-19
Decision: The roadmap is published as a nine-factor generator over a six-stage spine with entry and exit gates and no durations, resolved by a more-restrictive-wins conflict rule
Rationale: Named variants cannot represent the factor combinations enterprises actually present; gates are defensible from evidence while durations are not, and an explicit conflict rule is what stops a stated risk appetite from cancelling a regulatory constraint

ID: D027
Date: 2026-08-19
Decision: Supervision capacity is treated as a portfolio-level constraint, not a per-workload one
Rationale: Workloads are approved individually and supervised collectively; five A3 workloads each consuming a third of one supervisor's burst capacity do not fit inside one supervisor, and the arithmetic is never performed because no single approval exceeds the limit

ID: D028
Date: 2026-08-19
Decision: Every blueprint carries a mandatory honest-limits section naming what does not work, what has no published precedent, and what the guide refuses to claim
Rationale: Blueprints are the most quotable artifact in a guide and the easiest to read as endorsement; the utilities blueprint has the longest limits section in the corpus and that length is itself the finding

ID: D029
Date: 2026-08-19
Decision: Blueprints name agents by function and never by product, and state an A x L position for each
Rationale: Product names date within a quarter and imply a recommendation the vendor phase has not yet made; an A x L position makes the autonomy and learning claim explicit and testable against the oversight gate

ID: D030
Date: 2026-08-19
Decision: The vendor scorecard runs disqualifiers before scores, and publishes a section profile rather than a total
Rationale: A weighted total lets a vendor compensate for an unacceptable answer with strength elsewhere, which is how platforms with no enforcement boundary win evaluations on ecosystem depth

ID: D031
Date: 2026-08-19
Decision: Where the research cannot establish a vendor's answers to the question bank, the profile states the gap rather than inferring answers from adjacent products, even where that leaves a prominent vendor with the thinnest profile in the hub
Rationale: The hub's value is that its claims are checkable; inferring an enterprise platform's identity, memory and metering posture from a published SDK would make the hub's least-sourced section indistinguishable from its best-sourced one

ID: D032
Date: 2026-08-19
Decision: The site is a generated view of the repository markdown rather than a separate content tree, with site/content/docs git-ignored
Rationale: Two copies of the corpus would diverge on the first correction; generation keeps the repository the single source of truth (D001) and makes the site's link graph, titles and ordering derivable rather than maintained

ID: D033
Date: 2026-08-20
Decision: D001's public-repository assumption is suspended while publication-held paths remain tracked; the canonical repository stays private and the site publishes only the generated public view
Rationale: A site filter cannot remove a mechanism from public Git history. The repository can become public only after the hold is released, or a separate publishable-only mirror with no shared history is created

ID: D034
Date: 2026-08-23
Decision: The multi-card retrieval experiments are released on the site as a research section, negative results included, and the objective-conditioning and consumer-dependent-diversity claims are withdrawn from the guide; the mechanism-level technique entry stays under the publication hold and the research pages are the implementation guidance
Rationale: Five rounds on public corpora with human judgements settled the claims the guide had carried as hypotheses, and the guide's credibility rests on publishing what was tested rather than what was hoped; the research pages describe experiments on public data only, so D012's rule that system internals stay private is unchanged, and D033's suspension of the public-repository assumption stands until the hold is released or a publishable-only mirror is created

ID: D035
Date: 2026-08-23
Decision: The maintainer will not pursue patent protection for the multi-card retrieval mechanisms; the publication hold is released (its register stays as the build's mechanism, empty), the technique entry publishes with a status note, and all programme material may be shared; D033's suspension of D001 ends, and making the canonical repository public is the maintainer's own action
Rationale: Five rounds of experiments found no supported technical effect beyond the corpus-derived representation that prior art already covers, so there is nothing left to protect, and the guide's value lies in publishing the evidence, the reversals, and the approach it supports

Source: DECISIONS.md in the evidence repository behind this site.

Recorded decisions

On this page