Human-AI Accountability Redesign

Sources#

• AI-Augmented Human Resource Management? Insights from German companies
• Research: Why You Shouldn’t Treat AI Agents Like Employees
• The Human-AI Substitution Principle: When will you be replaced by AI in your organization?
• Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews
• When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration

Summary#

Five-pillar prescription from Kropp et al. (HBR May 2026) for redesigning organizational structure as agentic AI scales. The framing problem (AI as employee vs tool) is real but downstream — the underlying issue is that work, roles, and governance built for human pace and human accountability don't accommodate agents. Layering AI on existing workflows compounds errors and diffuses ownership. Companies that capture value redesign work; those that don't see review rigor decline and ownership fragment.

Why redesign is forced#

As AI takes execution, human roles concentrate on supervision, judgment, relationship building, and managing ambiguity. The shift is going unnamed in most workplaces. Oversight capacity does not expand automatically when output does — a manager whose team produced 5 documents/week can't oversee an AI that produces 50 without redesigning the unit.

The MSM/agentic-misalignment world (Agentic Misalignment (AM)) makes this sharper: agents that operate with weak human oversight per action and can take consequential moves are exactly where accountability redesign matters most.

The five fronts#

1. Sphere of accountability + span of control#

Oversight capacity does not expand with output volume. Redesign team sizes and reporting structure so that oversight remains tractable. The unit of accountability should match the unit the human can actually review.

2. Redesign roles + clarify expectations#

• Explicitly state oversight responsibility over AI systems in job descriptions.
• Set realistic expectations for velocity and volume of work with AI in the loop.
• Name the persistent human skills that remain critical — what the human is irreplaceable for.

Connects to Engineer PM Convergence: the persistent human skills are taste, judgment, ambiguity tolerance, customer-facing skills.

3. Reset performance management#

Reward quality of oversight and effective orchestration of AI, not just speed and output. Output goes up regardless; what differentiates is whether the human added taste/judgment/error-catching. Reviewing the agent's work is the new value-add — performance reviews need to measure it.

4. Treat AI as software with clear human accountability#

The blunt prescription: agents are software automation. They cannot be held accountable. Outputs need a salient responsible human — "When AI contributes to an outcome, it should be made salient to the responsible humans that are accountable for it." Especially critical in regulated environments.

Three subfronts:

• Decision rights — what the agent does autonomously vs requires explicit human approval. (Echoes Claude Code Auto mode design — classifier auto-approves safe, blocks risky.)
• Escalation — what triggers review, who intervenes, who bears the cost of delay or error.
• Consequences — when the agent fails, what happens next; accountable humans monitor + improve agent performance over time.

5. Design the agentic unit for the workflow, not the human role#

The "AI as 1:1 employee" framing assumes bounded roles + finite human capacity + delegation hierarchy. AI shares none of these limits. A single agent can operate across many workflows; multiple agents can reshape one job. Defaulting to one-agent-per-human-role pushes companies toward like-for-like replacement and underestimates redesign opportunity.

Better: pick the agentic unit — broader functional capability used across a team or process — that the workflow actually wants.

Pillar 4 at 70,000-applicant scale (Jabarian & Henkel, July 2026)#

The five pillars are a prescription; a pre-registered hiring field experiment is the closest thing the wiki has to one of them running in production. The design is pillar 4 verbatim — agents are software, decisions need a salient accountable human:

• Decision rights. The AI voice agent conducts the interview. Every hiring decision in every arm is made by a human recruiter, who reviews transcript, audio, and test scores. The automated stage is information collection; the judgment stage is untouched.
• Salience. The agent discloses its artificial identity at the start of the call (firm compliance, explicitly to avoid deception) and states that a human recruiter will review the interview and make the decision, not the AI. The accountable human is made salient to the subject, not only to the organization.
• The agentic unit is the workflow, not the role (pillar 5). The firm did not build "an AI recruiter." It automated one stage of a pipeline and left the rest intact — which is why the effect is cleanly attributable and why recruiters kept a coherent job.

The outcomes say this split works: 12% more job offers, ~18% more job starts and one-month retention, no productivity decline in hired workers. And it makes the cost of the split visible, which no prescription had done: the human evaluation stage became the queue. Median interview→offer time went 2.62 → 7.24 days because recruiters must now review conversations they did not conduct, and end-to-end time-to-hire rose from 20 to 24 days despite scheduling getting faster. Span of control (pillar 1) is the binding constraint the moment the collection stage speeds up — exactly as this page predicts, now with a number.

One finding cuts against the framing worry rather than for it. Where employee framing diffuses accountability and reduces error-catching, here the unambiguous tool framing coincides with recruiters increasing their reliance on independent evidence: they score AI-conducted interviews higher (1.90 → 2.01) yet weight the interview score significantly less in the offer decision, shifting weight onto the standardized language test. That is a decision-rights structure producing more scrutiny of the machine's output, not less — though framing was not manipulated here, so it is consistent evidence, not a test.

Accountability as a price, not a constraint (Banerjee & Singh, July 2026)#

The five pillars are a prescription; Banerjee & Singh's HAT substitution model (arXiv 2607.20781, practitioner-opinion — a formal model with no empirical data) supplies the same content as an economic decomposition, and the mapping is exact enough to be useful as vocabulary even though nothing in it is measured.

The model risk-adjusts every agent's cost as C̃ = C + λR, and decomposes AI risk into R' = ω₁R^(rel) + ω₂R^(comp) + ω₃R^(rep). Its §5.2 then maps those three components one-to-one onto accountability dimensions: technical accountability (did the system function correctly) → reliability risk; social accountability (reputational consequences borne by identifiable agents) → reputational risk; legal-regulatory accountability (liability when outcomes cause harm) → compliance risk. Because an AI agent "cannot bear legal liability, suffer reputational damage, or be held technically culpable," deploying it into an accountability-intensive role raises all three at once — the organizational form of the responsibility gap (Matthias 2004). The consequence the paper draws is pillar 4 restated as arithmetic: substitution fails to reduce risk-adjusted cost whenever λ(ω₁R^(rel) + ω₂R^(comp) + ω₃R^(rep) − R) ≥ C − C', even when the AI is nominally cheaper. Accountability binds "not by assumption but as an equilibrium consequence of the risk structure."

Two things this reframing is genuinely good for:

• It turns governance investments into parameters rather than obstacles. Auditability lowers R^(rel); regulatory engagement lowers R^(comp); transparent stakeholder communication lowers R^(rep); and organizational risk sensitivity λ is itself a lever a firm can set explicitly instead of leaving to implicit norms. Each moves the substitution boundary in a stated direction. That is a cleaner way to say "accountability redesign is a complement" than the prescription manages on its own.
• It predicts where the pillars will and won't hold. Large ω₂ (healthcare, finance, defense, legal services) sustains hybrid human-AI structures without any minimum-human-fraction mandate — the paper's distinction between hybrids that persist because regulation forces them and hybrids that persist because full automation is genuinely not cost-beneficial.

The limit of it, and where this page's evidence bites. Pillar 1 says oversight capacity does not expand when output does. The model assumes the opposite as its Assumption 1(iii): the AI coordination multiplier r'_0 — glossed in its own Table 1 as "integration effort, workflow orchestration, monitoring, human oversight requirements" — escalates no faster than the human one, which is what produces its flattening and wider-spans prediction (P3). The paper admits r'_0 "is not directly observable and should be calibrated as a scenario parameter," bracketed between 0 and r_0, never above. The randomized measurement in the section above puts it above: automating one stage made the human stage 2.8× slower and the end-to-end process longer. Flatter hierarchies with wider spans are what the model predicts; a lengthening review queue is what the vault measures.

Accountability someone else can compel: German co-determination (Kalff & Simbeck, July 2026)#

All five pillars are things management installs. Kalff & Simbeck's German HR study (arXiv 2607.13839, empirical, N=410 plus 14 expert interviews and three group discussions with works-council advisers) documents the same content as something a counterparty can compel — and then documents firms routing around it.

The mechanism. Under the Works Constitution Act (Betriebsverfassungsgesetz) §87(1) no. 6, any technology suitable for monitoring employee performance or behaviour is subject to works-council co-determination: the council holds veto rights, not consultation rights. This is pillar 4's legal-regulatory accountability — the HAT model's ω₂ compliance-risk term — with two properties the HBR prescription does not have. It is statutory, so it does not depend on managerial willingness; and it is held by the workforce, not by a regulator or the firm's own compliance function. The advisers interviewed describe councils demanding strict oversight against algorithmic discrimination and opaque "black box" systems, up to contractual blocking clauses, and the paper reports it working as the prescription intends where it engages: works-council approval requirements "can help ensure that AI projects protect employee interests," and strong representation "often serves as a counterbalance to individually tailored, algorithmically determined career trajectories, especially when these are enforced as rigid targets" (COD1).

The failure mode no internal prescription can name: the accountability structure is evadable by relocating the work. Three channels appear in the interviews:

• Offshoring the function. "Globally active companies sometimes outsource HR functions to affiliates or global service centres. By doing so, these functions are removed from the jurisdiction of local works councils and the scope of strict EU or German legal requirements" — and per COD1 this is "increasingly common when organisations wish to avoid negotiations regarding sensitive AI-based analytics." Decision rights are not redesigned; they are moved to where the rights don't attach.
• Buying the capability that doesn't trigger it. Firms adopt "simpler chatbots or generative-text assistants to sidestep these compliance obligations," and one vendor engineers personal data out of its ML product for the same reason.
• Starving the process of information. The paper's sharpest governance finding: "the concept of AI often remains intentionally ambiguous, which affects information flows in co-determination and participatory processes." A veto is only as good as the description of the system it is exercised over, and the description is written by the party being vetoed. See AI Employee Framing.

The general lesson for this page: an accountability mechanism strong enough to be binding is also worth evading, and evasion shows up as jurisdictional and product-selection choices rather than as visible non-compliance. Pillar 4's "make the accountable human salient" has an unstated precondition — that someone with standing can see what the system actually is.

The perception gap (referenced)#

BCG Henderson Institute: 76% of executives believe employees feel enthusiastic about AI adoption; only 31% of individual contributors report the same. Asking employees to use AI to "do more" without redesigning roles and accountability widens this gap. Adoption follows managerial role-modeling, not enthusiasm campaigns and not anthropomorphization.

The audit side of redesign: six conditions for calling it augmentation (CIVIC-AI, September 2026)#

The five pillars are a prescription for how to redesign. They contain no test for whether a completed redesign was worth doing, and no clock. When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration supplies both — a CIVIC-AI 2026 workshop whitepaper (22 authors across NUS, Stanford, A*STAR, NTU, UT Austin and Singapore's Ministry of Manpower, 8pp, practitioner-opinion: it defines a term and runs no study of its own; every figure in it belongs to a third party). Its premise is this page's, stated as a unit-of-analysis argument: usage logs, investment figures, productivity changes and headcount "do not explicitly measure changes in the nature of work itself," because "aggregate outcomes such as headcounts can mask structural changes like team reorganisations, shifting role responsibilities, and the delegation of tasks." The proposed unit is the workflow, and recording one means recording three things — what the agent may generate, recommend, decide or execute; where humans review, interpret, authorise, override or intervene; and how unfamiliar cases, unreliable outputs and system failures are handled.

The six conditions, in two layers (Table 1, reconciled cell-for-cell against pdftotext -layout). Layer 1 — workflow integrity, assessable at design, deployment and periodic operational review: (1) durable net value, where the multi-stakeholder gain must survive "full accounting of quality, human review, exception handling, rework, recovery, and the cognitive burden shifted to workers" — failure mode, "apparent productivity reflects work shifted elsewhere in the workflow"; (2) meaningful human control, where people retain "the competence, time, information, and authority" to detect errors, constrain agent actions to their intended scope, override decisions and continue working during system unavailability — failure mode, approval becomes ceremonial and deskilling raises operational risk; (3) clear accountability and recovery, with decision authority, provenance, escalation and fallback "explicitly assigned to human operators" — failure mode, errors without owners. Layer 2 — human development and job quality, assessable only longitudinally: (4) deepening learning, (5) career pathways (entry points and upward progression preserved or expanded), (6) job purpose — failure mode, "AI does desirable tasks, while humans are relegated to undesirable ones."

The structural claim worth keeping is the layer asymmetry: "a workflow may satisfy Layer 1 at deployment but fail to remain augmentative if it erodes Layer 2 over time." Every one of the five pillars is a Layer 1 object, assessed once, at redesign; none is measured on a clock. On this reading a five-pillar redesign that passed review in year one can stop being augmentation in year three without any pillar changing, because the people the pillars assign authority to no longer have the competence the authority presumes.

Where the boundary goes, task by task. Three properties, offered as the operational form of conditions 2 and 3: verifiability ("can a qualified person inspect the output and identify failure?"), reversibility ("can an error or action be corrected before serious harm occurs?"), and stakes ("what follows if the decision is wrong, delayed, or difficult to contest?"). Delegation can increase when outputs are inspectable, errors recoverable and stakes bounded; human authority should stay stronger where a task "defines goals, interprets ambiguous evidence, establishes validity, or shapes consequential use." The three conditions are stated as interdependent, in a form that is this page's own failure catalogue compressed to three sentences: "Output gains lose value when verification consumes the saved time. Human approval offers weak protection when reviewers cannot identify failure. Accountability is incomplete when responsibility is assigned without the information needed to reconstruct a decision."

The gap it claims in the existing governance literature is the sharpest thing in the document, and it lands on pillar 4. Disclosure requirements, audit trails and sign-off protocols "implicitly assume human reviewers retain the capacity to use them, but are silent on whether workflow redesign itself erodes that capacity." Making the accountable human salient does nothing if that human can no longer tell a good output from a bad one. The whitepaper's answer is the only genuinely new lever it offers: competence "does not persist automatically. It requires organisations to deliberately assign workers enough substantive review work and enough exposure to AI failure to keep the verification skill current." That is a staffing rule rather than a training programme — route real failures at reviewers on purpose, and treat the substantive-review share of a role as a maintained quantity. Nothing in the document measures whether it works, and the vault has no case where it was tried (see Layered Supervision, whose human layer has exactly this input unspecified, and The Tragedy of the Cognitive Commons, which argues the same maintenance is a commons no single firm is incentivized to fund).

The proposed instrument (§7), a shared workflow record, is the audit artifact the six conditions would be graded on: the baseline process and intended objective; tasks assigned to the agent and its level of authority; human review, override and escalation points; exception handling, fallback and recovery procedures; significant errors and unapproved or unintended agent actions; total effort, including verification and repair; and effects on human capability, AI literacy, learning opportunities, progression, agency and satisfaction. Two design notes travel with it: it must be complemented by employee-level evidence because "formal adoption data may miss unofficial AI use" (the shadow-adoption problem the German HR survey above measures at 183 of 410 respondents), and thresholds must stay sector-specific, since "clinical processes, research workflows, public surveys, and client-onboarding systems differ in their requirements for quality, privacy, contestability, and risk."

Weight it as a definition, not as evidence. Nothing here is measured: the "worked example" (an AI-mediated social survey with a stage-by-stage division of labour) is a hypothetical design whose pilot is proposed rather than run, and the Singapore section grades third-party government statistics the authors did not collect (carried, with its numbers and its caveats, on Organizational Complements to AI). What the document is good for is the same thing The Tragedy of the Cognitive Commons is good for — a structure that makes an already-observed pattern legible, here by converting "did the redesign work" from a productivity number into six named, separately falsifiable claims with their failure modes attached. Its authorship is the second useful fact: a labour ministry's statisticians co-signing the claim that their own instruments cannot currently establish condition 1.

Connection to existing wiki themes#

• Engineer PM Convergence — the persistent human skills (taste, ambiguity tolerance) align with what this paper says human roles concentrate on. Reset PM around oversight quality is the workforce-side mirror of "engineers become PMs because the bottleneck shifts."
• AI Native Product Cadence — Cat Wu's account of how Anthropic redesigned product cadence is one concrete instance of this redesign for the engineering function. HBR's prescription is the cross-functional version.
• Claude Code Auto Mode — decision-rights design at the tool level. The classifier that auto-approves safe / blocks risky is exactly the "decision rights" subfront made concrete.
• Harness Shrinkage as Models Improve — what doesn't shrink is the human role at the boundary; this paper names what that role becomes.
• Agent Loop Pattern — loops increase agent output volume per human; the span-of-control redesign in this paper is the missing partner.
• Loop Engineering — designing self-prompting loops multiplies output per human further; the review-bandwidth ceiling Osmani names is the span-of-control problem this paper's redesign answers.

Connections#

• Procedural Value in AI Decisions — the constraint from outside the firm. Every pillar here locates the human-agent boundary by properties of the work (verifiability, reversibility, stakes); a preregistered conjoint on 1,919 US job seekers shows the people judged by the system rank the boundary differently — decision authority +0.272, an appeal they can invoke +0.156, an opt-out +0.129, and a system-level independent bias audit only +0.068 (human involvement − bias audit, +0.204, p<0.001). The split is between rights an individual can invoke and oversight they cannot, and it maps onto the regulatory instruments: the EU AI Act mandates human oversight, GDPR Art. 22 a right to contest, NYC Local Law 144 a bias audit plus a notice of an alternative the rules never oblige the employer to provide. A redesign can pass pillar 4 and still leave the decision subject nothing to invoke — and a nominal reviewer who defers to the recommendation gives them nothing either

• The Enablement–Regulation Axis — the legislative layer above the co-determination section, and it is thinner than that section assumes. German works-council veto rights are an already-enacted statute being exercised; the question is whether new legislative demand is forming to close the evasion channels documented above (offshoring the HR function, buying the capability that doesn't trigger §87(1) no. 6). Chueri & Törnberg's 33-parliament corpus says barely: workplace-AI governance — algorithmic management, platform-work safeguards, human oversight — is 13.0% of the regulation-and-restriction frame, itself 21.8% of response-frame mentions, or 155 mentions in three years across 33 legislatures, and it sits behind copyright and creative-sector protection (36.1%). The threat side is twice as loud as the remedy: job quality and workplace control is 19.5% of threat content (345 mentions) against those 155. Parliaments are diagnosing algorithmic management considerably more than they are legislating against it

• Controlled Variance: AI's Edge as Reduced Dispersion — pillars 4 and 5 implemented and randomized at 67,056 applicants: AI collects information, humans hold every decision right, the agent discloses its identity and names the human decider. It works on outcomes and relocates the bottleneck onto the human stage (evaluation time 2.62 → 7.24 days), which is the span-of-control pillar arriving as a bill

• Layered Supervision — the software-engineering instance of conditions 2 and 3: supervision distributed across preventive, executable and human layers, where the human layer's re-scoping to concurrent supervision assumes exactly the reviewer competence the CIVIC-AI conditions say has to be deliberately maintained

• The Automation–Optimism Link — the term collision this page's new section creates: "augmentation" there is a use mode an individual is in, here a property a redesigned workflow passes six tests to earn

• The Tragedy of the Cognitive Commons — the levels above the org chart: same redesign prescriptions, plus the argument that organization-level action alone cannot solve a profession-level free-rider problem

• AI-Native Organization — the shape the span-of-control pillar has to survive: at $100M+ scale, 72% of AI-heavy companies run on 1–4 management layers and first-line spans are widening (7+ reports, 21%→30%), which is fewer reviewers each covering more delegated output

• The Household Production Boundary — the five pillars presuppose an organization; ATLAS finds most AI consultation on medical, legal, financial, and government matters happening in households at night, where none of the escalation and decision-rights scaffolding exists

• Agent-Native Infrastructure — agent-to-agent infrastructure needs the accountability/identity primitives this redesign calls for

• Companion concept: AI Employee Framing

• Cost mechanism: AI Brain Fry

• Decision-rights instance: Claude Code Auto Mode

• Workforce shift: Engineer PM Convergence

• Output-side accelerant: Agent Loop Pattern

• Risk surface: Agentic Misalignment (AM)

• Product-side instance: AI Native Product Cadence

• Interface-side complement: Interaction Models — org redesign keeps humans accountable; interaction models keep humans in the loop at the interface level (TML's "humans get pushed out by the interface, not the work")

• Role-evolution complement: Compute Allocator — names the individual-IC role ("decide what's worth compute") whose oversight quality and decision rights this redesign governs

• Non-code-output surface: Cowork — deck/dossier/inbox output lacks compiler/test verification, so accountability redesign matters more, not less

• Interface-side mirror: Turn-Based Interface Bottleneck — the interface-level version of the same critique: don't make autonomy the goal and push the human to the margin

• Solo-founder application: Founder as Agent Orchestrator — the five-pillar redesign framework collapses to one person but the accountability problems don't disappear; the founder retains decision rights and exposure

• Decision-rights substrate: MCP and Computer Use — the action surface that needs governing; "what does the agent do autonomously via MCP/computer use vs. what requires explicit human approval" is the concrete form of the decision-rights subfront

• Organizational complement: Organizational Complements to AI — accountability/span-of-control redesign is one of the workflow complements AI's value depends on; capability alone doesn't deliver it, the org has to rebuild the review-and-responsibility structure around delegated agents. Also the home of the HAT substitution model whose three AI-risk components this page's pillar 4 maps onto one-to-one

• Benchmark instrument: Configurable Human Participation — HAS-Bench's Human Agency Scale (A1–A5) and Control channel operationalize the decision-rights / escalation subfront as a measured variable; control-only authorization reaches 100% safety on protected actions where clarification/feedback (51/54%) cannot — the empirical case that decision rights can't be delegated down to lower-authority channels

Derived#

• Orchestration vs Employee Framing: Reconciling the Founder's Playbook with HBR's Accountability Evidence — applies this paper's five-pillar framework to the solo-founder context; shows where the framework collapses cleanly and where the accountability work does not disappear

Sources#

• Research: Why You Shouldn’t Treat AI Agents Like Employees — HBR, May 2026
• Working paper: https://emmawiles.github.io/storage/ai_employee.pdf
• Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews — Jabarian & Henkel, Voice AI in Firms (arXiv 2607.28222, 2026-07-30; empirical, pre-registered RCT): §2.4 (the AI discloses its identity and names the human decider; humans evaluate in every arm), §3.1–3.2 (outcomes), §6.2 (signal discounting), §7.1 (evaluation-stage queue). Parse warnings and full treatment at Controlled Variance: AI's Edge as Reduced Dispersion.
• The Human-AI Substitution Principle: When will you be replaced by AI in your organization? — Banerjee & Singh, arXiv 2607.20781 (2026-07-22; practitioner-opinion, formal model, no empirical data): §3.3 eq. 8 (the three-component AI risk decomposition), §5.2 (accountability mapped onto reliability / reputational / compliance risk, and the Matthias responsibility-gap framing), §4.4.1 Theorem 10 (the inequality under which a nominally cheaper AI still fails), §5.5.3 (the three governance levers λ, (ω₁,ω₂,ω₃), n_k), §5.5.2 (endogenous vs constraint-driven hybrids), §3.4 Assumption 1(iii) and §5.7 (the unobservable r'_0, bracketed 0 ≤ r'_0 ≤ r_0). Full treatment and the P1–P7 ledger at Organizational Complements to AI
• AI-Augmented Human Resource Management? Insights from German companies — Kalff & Simbeck, arXiv 2607.13839 (2026-07-15 / v2 07-20; empirical, mixed methods): §3 (the three group discussions with works- and staff-council advisers, and BetrVG §87(1) no. 6), §4.2 (works-council veto rights and blocking clauses; the externalisation-to-global-service-centres channel; the capability-substitution channel; the DEV3 no-personal-data quote), §5 RQ2 (the intentional ambiguity of the AI label degrading co-determination information flows). Glyph-level repair at ingest — docling emitted every digit and DOI/URL label as glyph names, restored 1:1 and re-verified against the PDF, so all quoted figures trace through that repair; Table 1 is row-shifted and cited nowhere, Table 2 verified clean. Full treatment at Organizational Complements to AI
• When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration — Wu, Ziems et al. (22 authors; NUS, Stanford, A*STAR, NTU, UT Austin, U. Washington, Singapore Ministry of Manpower), When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration, CIVIC-AI 2026 workshop whitepaper, arXiv 2609.12482, 2026-09-11, 8pp, practitioner-opinion (no original measurement: §5's worked example is a proposed pilot, §6 grades third-party Singapore statistics). Cited here for §2 (the capacity gap it claims in OECD/WEF/DOL/IMDA governance frameworks), §3 (the workflow as unit of analysis; the competence-maintenance staffing rule), §4 and Table 1 (the six conditions, two layers, failure modes; verifiability/reversibility/stakes), and §7 (the workflow record). Tables 1 and 2 reconciled cell-for-cell against pdftotext -f 3 -l 6 -layout — exact, no collapse, no row shift. Singapore figures and their handling at Organizational Complements to AI