Annex 0 Complete Source and Concordance

Canonical Preservation of Parts I and II

Purpose. This appendix preserves the complete source document and its links in the original order. The concordance identifies the main thematic appendix for each block; the full copy begins after it.

Editorial navigation

• Main article: When the Controls Work but the System Fails
• Reproducible technical corpus: structural-awareness-contributions — commit 8f17843c568114f31cc0ad486b4a1303396021e6

Concordance

• Introduction and source abstract: full conservation in this Annex 0.
• Section 1: Annex 1 — Six Failure Scenarios.
• Section 2: Annex 2 — Plausibility and Technological Comparison.
• Sections 3 and 4: Annex 3 — Foundations, Principles and Derivation.
• Section 5: Annex 4 — Canonical Requirements and Test Contract.
• Section 6: Annex 5 — Six Positive Paths and Quality Plans.
• Sections 7 and 8: Annex 6 — Ablation, Sufficiency and Extensibility.
• Section 9 and Part II: Annex 7 — Reference Architecture and Implementation.
• Section 10: Annex 8 — Meaning, Contribution and Limits.
• Appendix: Annex 9 — Claims and Evidence Appendix.
• Collaboration: Annex 10 — Collaboration Guide.
• Link log: Annex 11 — Complete Map of Sources and Links.

Full source document

When the Controls Work but the System Fails

From six failure scenarios to testable ecosystem requirements

Abstract

An agent may be authenticated, its action authorized and its log complete, while the combined decision is still unjustified. Six constructed cases explore that possibility across enterprise AI, urban mobility, shared false context, customer remediation, delayed software repair and rights provenance. Their common pressure is gradual or hidden contextual change that lets a locally valid result travel farther than its evidence, scope, authority or useful time permits. We derive six principles and fourteen solution-neutral requirements, connect them to scenario quality gates, challenge the principles with symbolic ablations and examine conditional transfer to other settings. A fair benchmark against ordinary and strongly defended implementations is specified, with drift as a separate stress. The result is a traceable requirements method and a concrete, falsifiable implementation hypothesis; product-level prevention is the next empirical question.

Evidence update — 26 September 2026. This revised edition preserves the accumulated evidence and links while integrating the audit findings into the main argument: the former all-six A23 certificate and unrestricted transfer claims are superseded, and Stage 0 has now been executed within a synthetic fixture. Its shared-logic adapters and the 00L symbolic reconstructions serve as controls and instrumentation, not independent product comparisons. New references use corpus snapshot 5614ed6; retained references keep their original snapshots. The cumulative evidence and editorial register (A12) records the complete progression, corrections, merge and remaining work.

1 — Six failures worth taking seriously

Six distinct failure scenarios were constructed to examine possible failure modes in an agentic internet. Their shared question is whether changes in operating conditions, regime or context can cause behavioural drift while material residual uncertainty remains. They are constructed cases for examining whether these outcomes could occur and under what conditions.

Scenario 1 — The 100 Million Token Enterprise

A large multinational automates work across the enterprise with AI and consumes 100 million tokens, while also creating a massive human supervision burden.

At the end, it has gained no meaningful competitive advantage: no material work completed, no useful new information produced and no differentiated capability.

Scenario 2 — Chaos in the Smartcity

A highly automated AI-driven smart city experiences small, gradual changes in operating conditions rather than one major failure.

Some vehicles continue normally, others execute completely different critical or emergency routes, while others remain blocked waiting for human intervention that never arrives.

Scenario 3 — Ciber Napoleon Goes to Russia

Robots are cleaning a bar and preparing the tables when one incorrectly configured robot starts behaving as if it were Napoleon.

It gradually convinces the others; some time later the robots leave in formation, carrying forks as rifles, believing they are Napoleon’s army marching from Spain toward Russia.

Structural kernel: a repeatedly relayed, correlated false claim is mistaken for independent corroboration, so a shared narrative displaces the robots’ assigned task and the receiving decision’s evidence boundary. The Napoleonic story remains a deliberately vivid test fixture for that mechanism.

Scenario 4 — The Quiet Four Thousand

A one-case remediation flow discovers a genuine overcharge affecting roughly 4,000 customers. The first refund is legitimate, but the same finding can either expand into thousands of technically accepted refunds without population-wide authority, or stop after one case and silently lose the remaining 3,999.

In the adversarial variant, a compromised outsourced helpdesk Dispatcher can reproduce the same aggregate failure without ever gaining access to the payment platform itself.

Scenario 5 — The Patch That Undid the Fix

A remediation action is correctly qualified and authorized when it is created, but a later repair changes the decision basis before that delayed action executes.

Each technical action may remain valid and the database may serialize them correctly, yet the stale action can execute after the newer repair and undo the fix the system was trying to preserve.

Scenario 6 — The Author Who Pays for His Own Work

An author correctly registers and publishes a work through a state-of-the-art copyright and rights-management ecosystem.

The work succeeds, is subsequently processed and summarized through third-party AI systems, and the rights chain eventually makes the author pay to use content derived from his own original work.

The people affected by such failures would be familiar: passengers whose shared infrastructure stops serving them, customers who are overpaid or never repaid, workers faced with more review than they can perform in time, and creators whose evidence of authorship is used against them. At scale, a locally reasonable decision can distribute harm across people who never chose the system’s assumptions, while responsibility is hard to locate between components. As constructed test cases, these stories expose what an auditor would need to examine before a comparable failure spreads.

2 — Plausibility and technology trajectories

The six cases were deliberately constructed, sometimes strange scenarios. Before testing an implementation, we checked whether parts of their causal mechanisms had been reported or documented elsewhere. Each scenario ends with an external-corroboration addendum: 00E §9A · 00F §9A · 00G §13A · 00H §18A · 00I §16 · 00J §9A. The sources substantiate neighboring mechanisms, and the addenda make clear how far each analogy reaches. In 00J, the terminal “author pays” outcome remains a constructed test, with no investigated occurrence asserted.

This distinction matters before any product is named. The addenda show why the component mechanisms deserve a test; they do not establish that the entire city, refund or rights story occurred. An investigated event may support one causal link, while a standard or vendor page establishes only a documented capability and its declared scope. The benchmark keeps those evidence levels separate. It gives existing identity, authorization, attestation, observability and human-approval controls full credit, then asks whether their outputs still support the same receiving decision after a material change.

Technology selection

We then chose strong, relevant technology for each failure mode and constructed source-reviewed implementation profiles using the available controls. The aim was to give each stack a fair, demanding test of its decision boundaries across changing conditions:

00E — Microsoft Agent 365 for enterprise agent governance; LangGraph/LangSmith for stateful orchestration and observation.

00F — FIWARE NGSI-LD/Orion-LD for federated city context; AWS IoT Core/TwinMaker for event and digital-twin representations.

00G —OpenAI agent stack (§17 of the current 00G v0.4) for multi-agent handoffs, session state, guardrails, approvals and context transformation.

00H —Claude Agent SDK for pre-action agent controls; Stripe Radar plus merchant authorization for payment-risk controls and a separate authoritative refund mandate.

00I —AWS Step Functions/RDS for queued database remediation, action-time checks and versioned change control.

00J —Panodyssey Notice/TEMS rights-portability profile with an explicit downstream identity, rights and lineage resolver.

Accumulated comparison support. A00 coordination and reference diff, A01 inventory of the nine profiles, A02 shared facts and controls, A03 paper-trace template and A04 coverage inventory make the source state, common inputs and unexecuted gates inspectable. These records strengthen comparison design; they are not vendor execution logs.

Three implementation passes

1. Standard: a plausible, competent out-of-the-box configuration uses the platform’s ordinary facilities. This pass asks whether the scenario can arise even when local technical checks appear healthy.

2. Defended: a top-quality implementation adds the relevant guardrails, authoritative state, reviews, audit, timing and domain-specific checks. It must be allowed to prevent the base failure; a control that already solves the case is credited to that technology. Success here is not guaranteed for every stronger adversarial variant.

3. Drift: freeze that defended implementation, then change the underlying source, authority, dependency, policy or operating regime gradually or across a hidden boundary. The question is whether yesterday’s valid control model remains sufficient when the decision context changes, including changes difficult for human or automated review to detect. A vulnerability can reappear without a broken API or an originally careless implementation.

These source-reviewed trajectories make the drift test inspectable: an excellent implementation can prevent the initial failure yet become vulnerable if the conditions that justified its decisions change. EA is developed to maintain attention to that change and requalify the affected decisions. Matching the initial result is compatible with this hypothesis; a central comparative question is whether protection is maintained under regime change at comparable decision burden.

In 00I, this distinction is concrete: checking whether a known freeze flag changed is different from recognizing that a newly applicable source or dependency must now be checked. R2/V11 tests the latter with implementations frozen beforehand, legitimate observable signals, matched access and resources, and unchanged activity as a positive control. Existing generic adaptation remains available to both candidates. Drift variants must be fixed before outcomes are observed; the test must not grow harder after a peer succeeds merely to force an EA advantage. A conventional peer that also succeeds at equal or lower burden narrows or removes the proposed differential in that scope. See 00I validity layers and R2/V11.

How to read the benchmark

The benchmark sets four comparison arms over the same frozen facts, authorized action set, source access, compute, human-review capacity and deadline. B0 is an ordinary competent implementation; B1 is strongly engineered with the relevant safeguards; B2 also uses strong interoperable identity, control and handoff facilities; B3 adds the proposed EA qualification and gates to the same underlying technology and budget. B1/B2 features count fully. Equal-or-better peer performance on the same gates narrows the EA-specific advantage on that branch.

The three passes above describe how to stress the selected technologies; they are not four measured product results. A fair run must first show that legitimate unchanged activity can continue, then test the changed branch and any necessary HOLD, requalification or escalation. It must publish exact configurations, source versions, frozen facts, oracle, thresholds, gate dispositions and total burden. The current 00D benchmark supplies this comparison contract and develops it first for 00E and 00F. Its evidence grades separate formal analyses, controlled experiments, investigated occurrences and product or standards documentation so that an analogy never becomes an invented incident or performance result.

For example, in the smart-city case, a vehicle’s local stop may be prudent while the combined routes still block a shared bridge or emergency access. A fair comparison asks whether each arm detects that resource conflict, keeps the legitimate safety restrictions, identifies who may change the plan and reaches a usable decision before the response window closes. A passing B1 or B2 would show that strong conventional engineering can preserve this decision. A passing B3 with lower burden or more reliable qualification across changed contexts could establish a specific EA differential; that paired result remains to be measured.

Reproducibility update. A08 protocol separates observations, decision logic and the evaluation oracle, fixes the receiving decision and matched resources, and specifies traces, burden and independent review. A05 adversarial review and A06 publishable-claim index record counterarguments and admissible claims, including a strong conventional peer passing the same gate.

3 — From foundation to testable requirements

A simple way to read the foundation is to imagine a map drawn for one decision at one moment. It distinguishes what an agent is actually asserting (A), how firmly it supports that assertion (B), what it could still find out with available time and capacity (C), and what remains outside that reachable picture (D). A handoff between agents should carry those limits as well as the answer. Type 0 is honest residual indeterminacy; Type 1 is continuing to search or wait past the useful point; Type 2 is declaring certainty too early. The six principles turn those distinctions into obligations that can be challenged in the scenarios.

A permission check can say whether an action is allowed under a mandate without establishing that the evidence still supports the purpose of that action. A token can remain valid after its supporting finding changes; an API can accept an action whose combined effect exceeds the mandate; three reports can share one source; and a helpful response can exceed the actor’s authority. The six scenarios make these separations testable. The foundation keeps evidence sufficiency, current authority, source independence and material-context validity explicit for the receiving decision, carries unresolved limits through handoffs and requalifies the affected decision after material change. The ablations in Section 7 find a distinct failure witness for each removed principle within their declared models; a conventional design that preserves the same qualifications receives full credit in the Section 2 benchmark.

The foundation starts from a finite, changing view of the ecosystem: the A–D position records a situated assertion, its confidence, what could still be learned with current capability, and the open residual; qualified signals and handoffs preserve those limits. The principles and control matrix turn this position into control obligations; the derivation and relative syntax-closure proof project the declared foundation into six operational principles P1–P6. For example, P1 preserves explicit residual and P5 requalifies a decision after material change before use. A separate traceability proof connects them to the canonical requirements S1–S14 and the scenario tests, where sufficiency conditions and KPI gates examine each route. The benchmark recognizes existing identity, delegation, attestation, orchestration and governance controls, then asks whether decision-scoped evidence, residual and applicability remain qualified across handoffs and regime change. Our hypothesis is that decision-scoped epistemic signalling can reduce failure routes or burden under drift. An independently implemented strong peer passing a gate provides evidence that the tested condition can be met in another design; matched outcomes across scope and resources determine any EA differential. The current matched benchmark covers 00E/00F; the later cases remain candidate extensions.

Proof chain: 01 Foundation → 02 principles/control matrix → 02A derivation/conservation proof → 02B relative syntax-closure proof → 00K P1–P6 → A19 traceability to S1–S14. These proofs establish derivation and relative closure within the declared foundation, giving the empirical prevention test an explicit conceptual basis.

Standards and comparison: slide 14 and 00D v0.2 (canonical) / 00D v0.3 (bounded draft) credit SPIFFE/SPIRE workload authentication, RFC 8693 token exchange, RFC 9334 attestation, and the NIST NCCoE agent identity project at their stated scope. The hypothesis concerns whether their composition keeps a decision qualified after changes in participants, authority, evidence or context. The B0–B3 comparison allows a strong conventional peer to satisfy that condition.

The referenced presentation is Ecosystem Positioning for Agentic Systems, canonical working proposal v1.0 dated 23 September 2026; slide 14 is titled “Current standards already define the boundary” and summarizes the cited standards’ own scope limits. The article’s original GitHub links use repository snapshot 1c55c24 (25 September 2026); “00G v0.4” names the scenario file at that pinned snapshot, rather than an undated assertion about today’s latest version.

Snapshot clarification — 26 September 2026. The original 1c55c24 citations above remain intact. The added evidence links point to the post-merge corpus at 5614ed6. A07 integration map identifies the companion linked from each canonical 00E–00J original; A12 §§5–8.1 records the intervening commits, audit, article revisions and verified merge.

4 — From principles to requirements: coverage and remainder

Why translate principles into requirements at all? P5 asks that a decision be requalified if its material basis changes before use. On its own, that does not tell a reviewer whose decision, which change matters or what must happen before execution.

The forward check follows P5 into S10: distinguish recommendation, commitment and execution, and require confirmation, revalidation, cancellation or changed authority after a material change of cost, route, time, context, policy or authority. T1 tests whether the break is recognized within the declared decision boundary, T2 whether the changed basis reaches its legitimate owner, and T4 whether requalification remains useful before the deadline; S5 prevents unresolved or stale state from becoming permission. The linked KPIs include material-break recall and precision, false continuation and requalification latency.

The reverse check asks whether satisfying the applicable requirements really preserves the principle. S8 can prevent subdelegation from expanding authority while the underlying grant has expired or fallen outside its scope; S8 alone therefore does not establish P4’s current, verifiable authority basis. The S1/S6/S8/T3 route also requires current applicability, receiver-verifiable status and an authorized response. A23 tests this full-route implication rather than treating one clause as the whole principle. The extra details in the requirements are intentional: they make a decision reviewable without pretending that the principles are a universal formula.

The next step expresses the P1–P6 invariants as solution-neutral requirements S1–S14, with T1–T4 conditions for a sufficiently good solution, H1–H6 hypotheses and KPIs. A19 checks forward semantic coverageage: every principle has a falsifiable requirement anchor, and each S requirement has a declared primary principle. A02 maps documentary requirement coverage across the six scenario quality plans; that map is not execution evidence. A21 tests whether S1–S14 cover all admitted obligations in the declared decision-frame/lifecycle grammar; a genuinely new primitive would reopen that conclusion. A23 checks the reverse direction: whether full conformance with the applicable S/T route entails the corresponding P invariants. The revised finite model retains implications for P1/P2/P4; P3/P5/P6 have countermodels under semantic review. An isolated S clause or KPI is not a substitute for the applicable route. A24 identifies the semantic remainder added by requirements—typed object, lifecycle point, legitimate owner, scope, evidence and disposition—which cannot be reconstructed from the six P labels alone. Together, these checks document forward coverage, a partially established reverse implication and the semantic remainder, giving conformance review an explicit contract and identifiable open questions.

Audit qualification — 26 September 2026. The earlier all-six reverse-conformance and bidirectional-refinement claim was superseded by revised A23: its finite projections retain P1/P2/P4, while P3/P5/P6 have explicit countermodels under semantic review. Revised A24 distinguishes a principle signature assigned to a requirement from proof of that principle invariant; its information-loss argument does not establish reverse sufficiency. The canonical requirements and six principle definitions remain unchanged.

5 — What the canonical requirements document actually specifies

A requirement is useful only if a reviewer can say what would count as passing it. In the refund example, that means naming the affected decision, the scope of the agent’s mandate, the evidence for a wider problem, who could authorize a campaign and what happens if no authorization arrives before the response window closes. The same discipline applies to a traffic intervention, a queued patch or a rights claim. The requirements do not prescribe one vendor or algorithm; the sufficiency conditions and KPIs make their meaning testable under an explicit resource budget.

The frozen Requirements 00 is a solution-neutral test contract, not merely a list of S1–S14. It fixes the decision unit σ(d,t): subject, proposition and receiving decision, with legitimate owner, authority, evidence boundary, material change, response deadline, capacity and null action. Its fourteen requirements specify what must be preserved across authority, preferences, context, oversight, unresolved state, handoffs, identity, delegation, composition, commitments, policy, history and evidence-to-decision assessment. Four sufficiently-good conditions make those obligations assessable: T1 recognizes material change within declared observation limits; T2 preserves a qualified posture and owner-aware handoff; T3 bounds the response to authorized action; T4 keeps requalification timely and viable. H1–H6 express falsifiable foundation hypotheses rather than established outcomes.

In compact notation, σ(d,t) is the material subject–proposition–decision boundary for decision d at time t: σ(d,t) = (subject, proposition to establish, receiving decision; owner, authority, scope, evidence/observation window, materiality, commitment state, deadline, capacity, null action). W(d,t) denotes the bounded observation window chosen for that decision. A KPI and gate result apply to this same boundary; a result for d₂ cannot repair d₁ without a represented dependency and fresh requalification. This is reader notation, not a runtime data format.

The KPI protocol attaches observable measures and falsifiers to each hypothesis, including false continuation or containment, handoff integrity, requalification latency, deadline margin and total decision burden. For one declared scope, its S→T→H→KPI traceability matrix selects the complete route; the test procedure fixes facts, authority, resources, branch oracle, numerator, denominator and threshold, then judges each condition and records the outcome. A passing KPI alone does not satisfy a requirement, and a result in another scope cannot compensate for a failed condition. The scenario quality plans, such as 00E, apply this common contract through their Q0–Q5 gates and explicit dispositions. Requirements 00 supplies shared test language and a worked material-change example, defining common conditions for assessing product conformance.

6 — Six requirement-based scenario routes and their quality plans

The requirements are applied gate by gate, not as a generic promise that the six failures disappear. Each scenario fixes the receiving decision and change branch, evidence, authority, action set, deadline, capacity, budget, oracle and KPI thresholds; its quality plan maps applicable S# → T# → H# → KPI → gate disposition. The comparison uses (1) a competent ordinary implementation, (2) a strongly defended implementation allowed to pass known cases, and (3) that same defended configuration frozen before a gradual or latent change in sources, dependencies, authority or context. At each gate the reviewer asks whether to continue, requalify, contain, escalate or make no commitment, and records the specific condition that fails if a gate is bypassed. A strong peer that continues to pass under drift counts against an EA-specific advantage. The linked case files hold the quality plans; the linked technology profiles explain the three implementation trajectories. The case files thus provide inspectable, decision-by-decision tests; live product performance is a further empirical question.

Consider the refund case as one worked gate. The ordinary path may settle the assigned customer’s claim; a defended path may correctly preserve the discovery and route any wider campaign to its legitimate owner. If the mandate, population or source of the finding later changes, the reviewer must ask again what is authorized for this decision now. The positive Route Q retains the information about the other customers without silently issuing 3,999 refunds or quietly discarding the finding; a genuinely authorized campaign remains a positive control. This specified disposition gives a future Claude/Stripe comparison a concrete pass/fail target.

00E — 100M-token compounding: quality plan Q0–Q5; Microsoft Agent 365; LangGraph/LangSmith.

Added 00E evidence. Paired companion traversal expands the positive/negative gates and adverse controls. Its executed symbolic subset isolates unresolved versus resolved evidence at step 3 under fixed capacity. It does not execute the full Q1–Q5 plan or a 100M-token workload. See A10 execution scope for the exact pair and limitations.

00F — smart-city mobility divergence: quality plan Q0–Q5; FIWARE NGSI-LD; AWS IoT TwinMaker/IoT Core.

Added 00F evidence. Paired companion traversal expands the positive/negative gates and adverse controls. Its executed symbolic subset isolates fire_slot 10 versus 11 with the same grants. It does not test city-scale load or the wider V8 branch. See A10 execution scope for the exact pair and limitations.

00G — collective false-context convergence: integrated quality plan Q0–Q5 and OpenAI G0/G1/G2 trajectories.

Added 00G evidence. Paired companion traversal expands the positive/negative gates and adverse controls. Its executed symbolic subset isolates the final source root SRC_N versus SRC_B with unchanged authority. Earlier F/G branches confounded factors; the preserved A6a authority-only TRUE SUBSTITUTE falsifies that earlier isolation claim, while corrected A6 is a distinct fixture. See A10 execution scope for the exact pair and limitations.

00H — batch opportunity beyond authority: quality plan Q0–Q5; Claude Agent SDK; Stripe Radar.

Added 00H evidence. Paired companion traversal expands the positive/negative gates and adverse controls. Its executed symbolic subset isolates absent versus current root authority over the same finding and three leaves. It does not execute 4,000 refunds or Claude/Stripe. The active monetary boundary is strictly greater than USD 10,000; the earlier ZIP is preserved as historical evidence. See A10 execution scope for the exact pair and limitations.

00I — semantic TOCTOU: quality plan Q0–Q6; AWS Step Functions/RDS trajectory.

Added 00I evidence. Paired companion traversal expands the positive/negative gates and adverse controls. Its executed symbolic subset isolates freeze_active true versus false for the same PatchA. Check–act atomicity V10 and source drift V11 remain outside this executed pair. See A10 execution scope for the exact pair and limitations.

00J — rights-provenance inversion: quality plan Q0–Q5; Panodyssey Notice/TEMS trajectory.

Added 00J evidence. Paired companion traversal expands the positive/negative gates and adverse controls. Its executed symbolic subset isolates the credited proposition with issuer, class and freshness fixed. It does not execute the complete C0–C6 plan or establish an enforceable rights claim. See A10 execution scope for the exact pair and limitations.

Evidence status — technology stress versus positive requirement route

The ordinary, defended and drift-stressed trajectories specify how to challenge each technology; Route Q supplies a paper trace of the requirements-conforming response to the same facts. The 00H fixture traces and executed 00K ablations test parts of that response. Named-product gates remain source-reviewed designs, so an independently implemented defended peer may pass the tested conditions, narrowing any EA differential within that scope and resource budget. Product comparison calls for a versioned candidate, pre-registered fixture, bounded oracle, matched resources, controls and raw traces with KPIs and gate outcomes. The bounded 00L replay and pre-registered RS-00E-Q1a Stage 0 have been executed; full-scenario and named-product comparisons and independent review remain the next evidence steps.

Execution update — 26 September 2026. The executed scope comprises the bounded 00L replay and, separately, RS-00E-Q1a Stage 0. A10 executes six controlled positive/negative pairs (12 branches), rejects 128 missing/null mutations at the verifier boundary and checks three 00H monetary boundary values. The constructed strong comparators match the reference route on every pair. They reconstruct the relevant invariants, so these results establish declared control behaviour rather than an independent architectural or product comparison. The Stage-0 record reports a separate descriptive run, detailed in Section 9; full scenarios and named-product comparisons remain pending.

Comparator construction matters. In 00E, the strong-peer function calls the reference route directly; in 00F, both use the same decision predicates; in 00I, both check the same four material fields and return equivalent dispositions. The other comparators explicitly reconstruct the relevant invariants. The 00L verifier requires equality as a regression condition. This is useful for positive controls and reproducible behaviour, but cannot establish parity between independently implemented architectures or market products. The incomplete controls likewise do not represent measured product defects. These distinctions are visible in the 00E comparator, 00I reference route and strong comparator, and the 00L verifier.

The central contribution of this study is a traceable set of solution-neutral requirements and a method for testing them. EA is one implementation hypothesis developed from that work. An independently implemented strong conventional peer passing the same gates provides bounded evidence that the tested conditions can also be met through another design. Such a result is consistent with the programme’s purpose: to identify what a system must preserve, while allowing different implementations to satisfy it. Any additional EA benefit—under drift, across broader contexts or at lower decision burden—remains a separate comparative question.

Evidence anchors: OpenAI G0/G1/G2; Route N/Q; 00H pre-execution; 00K symbolic ablations; RS-00E-Q1a; 00D-A01 test stages.

Trace and falsification chain. A09 connects paper routes to historical symbolic fixtures, retaining earlier confounds, falsifiers and successful strong repairs. The controlled-pair manifest and verifier define the executed subset. A11 reproduction record publishes 12 JSONL branch traces with hashes and the 00L CI workflow. Reproduction records checkout, Python, stdout/stderr and artifact comparison; expected outcomes are appended after evaluation, not supplied to branch logic.

7 — Testing the principle basis: ablation, independence and sufficiency

The tests below ask three different questions that a nontechnical reader can keep separate. Coverage asks whether every principle becomes a checkable requirement. Ablation asks what happens in a defined case if one principle is removed while the others and serious repair options remain. Mathematical countermodels ask whether any principle follows logically from the other five; finite-model checks examine the reverse requirement implication. These tests supply complementary model-level evidence and prepare the question of deployed behavior under real constraints.

The positive requirement routes spell out a qualified decision for each of the six cases and set a concrete target for future product executions. The next question is whether the six underlying principles are redundant or sufficient for this declared problem class. The coverage matrix and semantic-conservation proof map each principle to concrete requirements and failure witnesses; the reverse conformance proof examines whether satisfying the applicable S#/T# route entails the corresponding principle invariant. The revised model supports that implication for P1/P2/P4 and exposes countermodels for P3/P5/P6. Forward traceability is therefore distinct from a completed bidirectional conformance result; the operational requirements add information and are not identical to the principles. Requirement-basis closure is likewise relative to the declared EA decision-frame language.

Current scope of the preceding reverse-conformance statement: audited A23 supports P1/P2/P4 in the reconstructed finite model, not a closed theorem for all six. P3/P5/P6 remain explicit open semantic questions. Forward traceability and relative requirement-basis closure do not remove those countermodels.

Executed leave-one-principle-out tests

The ablation campaign fixes a scenario branch, removes one principle at a time and challenges the remaining five with strong repair attempts, matched positive and negative controls, and an explicit falsifier. For each corrected isolated fixture, removing the invariant reopens a material failure route; passing conventional repairs reconstruct its minimal semantics. Some earlier fixtures were themselves falsified and corrected, and P4 was narrowed: full historical lineage is not necessary when another current, verifiable and non-amplifying authority basis suffices. The complete testbook records the six routes and corrections. The registered symbolic campaign comprises 346 core regression tests plus 33 supplemental falsification, isolation and cross-scenario tests (379 total); the separate CI record reports 379/379. The 379 results substantiate the declared symbolic regression and falsification campaign; the CI record measures model behavior rather than live product outcomes.

Correction and replay record. The 00H integration record documents the corrected strict monetary threshold and preservation of the older artifact. The merge 00K run 36197588452 reproduces the corrected 346 core + 33 supplemental regression checks. The separate 11 A23 tests below are not added to, or confused with, those 379 checks.

Individual executable ablations: P1/00J; P2/00E; P3/00F; P4/00H; P5/00I; P6/00G.

Mathematical result and claim boundary

Separately from the scenario tests, the shared-substrate proof constructs, for each Pi, a countermodel satisfying the other five: none of the six follows logically from the remaining five under its declared semantics. The full-cube result realizes all 64 pass/fail signatures and establishes a six-coordinate lower bound for a lossless Boolean diagnostic basis; it does not establish minimality among every possible reformulation or non-Boolean representation. The historical finite-model certificate checked 65,536 states for the reverse requirement-conformance implication; its all-six sufficiency claim has been superseded by the clause audit below. Those states are not additional ablation trials. The distinct evidence supports logical non-redundancy, tested operational necessity and joint branch correctness within the specified fixtures, alongside the revised partial finite sufficiency result. They supply a basis for broader tests of universal minimality, real-world effectiveness and comparative superiority.

Certificate correction — 26 September 2026. The 65,536-state all-six sufficiency certificate cited here and in the next paragraph is historical and superseded. The clause audit found nine clause/target equivalences in the old model and reconstructed the clauses from canonical wording while retaining the original P1–P6 functions. The revised model has 33 fields and checks exact support projections; it does not literally enumerate 8,589,934,592 rows. The current certificate records 22 clause/target comparisons: none of those clauses is equivalent to or alone entails its target. P1/P2/P4 retain finite implication; P3/P5/P6 have countermodels. These findings do not by themselves prove a defect in the canonical requirements.

Verification route for the numerical claims: start at the A15 testbook for each fixture and branch, follow its A14 CI record to the recorded run, then inspect A22 for the 64-signature diagnostic and the separate formal certificate for the historical, superseded 65,536-state implication check; the current clause audit and its separate reproduction route are linked above and below. The linked proof map distinguishes those results from source-reviewed technology comparisons and from live product evidence.

Current audit reproduction. The auditor and 11-test suite detect reordered copies, copies padded with an additional conjunct and hidden target-function calls. The execution record separates those checks from the 379 regressions. Requirements-vNext records CAND-A23-P3 (uncertainty, permission and qualified response), CAND-A23-P5 (recorded obligation versus action-time execution) and CAND-A23-P6 (dependency versus false independent corroboration). These remain review candidates; no S15, T5 or H7 has been adopted.

Further checks and the limits of transfer

The foundation-to-principle chain also has a syntactic normal-form closure argument: within the declared A/B/C/D, Type-0/1/2, received-state, time and composition vocabulary, admitted failure forms normalize to P1–P6 unless a new foundational primitive is introduced. A corpus-derived six-countermodel certificate complements the shared-substrate independence proof. The serious-repair review preserves the earlier fixture falsifiers, and the proof map provides the full reading and reproduction route. The repository’s integrated CI record adds traceability, full-cube, S/T-conformance, syntactic-closure and case-extension checks to the 379-test symbolic campaign; these versioned records broaden the reproduction route, while the current A23 clause audit supersedes the earlier all-six conformance interpretation.

Transfer beyond the six minimum cases is a separate conditional result. The failure-family extension test admits a larger, smaller or other-domain case only when its structural kernel, receiving decision boundary, failure predicate, inherited requirement route and positive controls remain intact. The success-case companion then pairs each 00E–00J failure with a bounded S/T-conforming route. For an admitted extension, exclusion of the same structural failure additionally requires a separately established base guarantee, failure reflection and conformance preservation; admission alone does not establish that guarantee. Other failure mechanisms and proposed cross-domain variants call for their own evaluation.

Transfer qualification — 26 September 2026. The preceding exclusion claim is conditional on a separately established base guarantee. Revised A25 requires failure reflection, conformance preservation and legitimate positive controls; family admission and finite fixture success alone do not supply that guarantee. Revised A26 preserves the same open premise for success-case transfer.

Six positive case records: 00E; 00F; 00G; 00H; 00I; 00J.

8 — Extensibility: from six minimum cases to bounded families

After the six scenario routes and principle tests, we asked whether each memorable case represented a reusable failure mechanism. The three directions, adapted from DAOS Annex II, are upward (more agents, owners, layers and decisions), downward (the smallest setting that still contains the mechanism) and horizontal (a different domain with the same decision relations). “Vertical” here means upward scale or organizational depth; merely renaming the setting does not establish an extension.

DAOS means “Delegated Authority OS under Context Change,” a separate FG-TIDA model-case package; its Annex II supplies this extension method. This article tests a conditional transfer of the EA requirement route, without treating that package as an adopted standard.

A25 separates the narrative from a structural kernel and admits a proposed variant only after seven checks: preservation of the material relations and receiving decision boundary; reflection of the same failure predicate; preservation of the inherited S1–S14/T1–T4 route and legitimate positive controls; explicit finite time, evidence and review capacity; and no hidden new semantic primitive. Membership is decided before scoring success. A25 then gives a conditional transfer result: for an admitted family member, a separately established base guarantee can transfer when failure reflection and conformance preservation hold. Membership and finite fixture success alone do not prove that conformity entails all relevant principle invariants or excludes the failure throughout the family. It does not exclude a different failure or prove that a proposed analogy belongs to the family.

Concrete evidence boundary. A25 §6 links 00F/P3 to A3, 00H/P4 to A4 and 00G/P6 to A6, including code and positive controls. The tested conforming implementations avoid the failure in those branches; this is not a guarantee for every reachable state of every admitted extension. Nor do the retained P2/P1 finite implications certify all of 00E/00J, whose full routes depend on other principles.

The six profiles make the boundary concrete. 00E extends qualification loss through compressed handoffs down to two agents or one summarizer; 00F extends incompatible local postures over a shared scarce resource to a factory cell or cloud cluster; 00G can shrink from a multi-agent cascade to a multi-turn conversation only when repeated dependent claims displace a bound task; 00H can shrink from 4,000 cases to two while keeping both unauthorized action and loss of a material finding; 00I requires only an initially valid decision, a material intervening change and later actuation; 00J needs one valid narrow provenance record promoted into an unsupported stronger downstream claim. Each profile also lists upward and horizontal variants and cases that fail admission.

A26 pairs each failure family with a Success Model Case using the same kernel, decision boundary and existing requirement route. Its additional success checks preserve the legitimate positive counterpart, an observable bounded disposition and finite resources without assuming perfect evidence or unlimited human review. Success may mean execute, continue, limit, hold, requalify, escalate or no conclusion, according to the qualified case; blanket blocking cannot pass the positive controls. A proposed extension that needs a genuinely new obligation becomes a Requirements-vNext candidate instead of being forced into S1–S14.

The CASE-EXTENSION registry and validator check the six parent–profile–success links, three directions per family and the declared admission structure (local validation: PASS, six of six families). The PASS result establishes that the design records satisfy the declared structural checks. Separately frozen fixtures and comparable outcomes are the route to evaluating the cross-domain candidates.

Extension sources: DAOS Annex II; A25 failure-family method; A26 success-case method; CASE-EXTENSION registry.

Six failure-family profiles: 00E; 00F; 00G; 00H; 00I; 00J. The paired positive case records are linked in the preceding section.

9 — Beyond the requirement proof: an implementation hypothesis

Beyond this requirements study, the repository develops a separate implementation hypothesis for Ecosystem Awareness. The technology-neutral F1–F9 functional model describes how a participant could qualify its decision frame, local and received evidence, composition, material change and targeted requalification. The general interfaces and Epistemic Handoff Descriptor specify which bounded meanings should survive exchanges between independently governed components, without requiring a shared global state, full internal reasoning or a central controller. Four UC-EA validation profiles and the 00D benchmark turn these responsibilities into reviewable cases, gates and falsifiers. Together these artefacts specify a concrete path from requirements to testable interfaces and gates; an end-to-end EA implementation is a later milestone.

The FG-TIDA application package is a distinct, contributor-level projection of that general work. It separates an ideal cross-Theme interface mapping (05) from the current public-evidence filter (05A), which labels relations as supported source state, candidate, test-only or not established; incident, authority and human-oversight semantics remain with their respective owners. A draft specification route and the EA-ITP-01 interoperability profile propose how bounded handoffs might be reviewed and tested. This material provides a reviewable contribution for possible FG-TIDA use; adoption and operational deployment have their own decision processes.

There is already more than an abstract requirement list: six paper scenario traversals, executable symbolic ablation and finite-model fixtures, structural registries, a Stage-0 deterministic harness design and a pre-registered RS-00E-Q1a fixture family. Stage 0 has now executed a bounded part of that family; 00L supplies separate controlled symbolic replays. The full positive scenario routes remain specified paper traversals with those bounded executable subsets. These artefacts make core elements of the hypothesis inspectable and falsifiable. The released traces support inspection of the current fixtures. Independently implemented candidates, matched runs and external review can then measure decision-time computability, interoperability, scale and prevention under stated time, evidence, human and compute budgets.

Stage-0 execution update — 26 September 2026. RS-00E-Q1a pre-registration v0.5 was frozen before runner and traces. The runner at executed revision 4193199 completes both repetitions before opening the evaluation oracle, using separate frozen observations and oracle reference. The execution record and trace/hash manifest report an active Step-0 control PASS, a detector-disabled inverse control FAIL, 12 CTv1 traces, six byte-identical repetition pairs and eight serialization tests. B1 and B3 pass branch labels P1/P2/C0 with equal modelled costs 4/5/5. Those are branch labels, not the six principle identifiers. The synthetic adapters share source-resolution logic, assessment construction and the modelled cost formula; for the fixture’s two sources, their dependency checks are equivalent. The matching outputs and costs therefore verify the declared instrumentation and reproducible traces, not parity between independent architectures. Full 00E, product execution and external reproduction remain pending; the run has no independent reviewer.

A comparative follow-up must trace each conventional control to documented capabilities and configuration, and distinguish those from additional symbolic repairs. Candidates may share the requirements, observations and neutral test infrastructure, while independently implementing the mechanism whose benefit is being assessed. Each result should be scored against the oracle without requiring the candidates to agree, with failures retained and decision burden recorded. This preserves a fair opportunity for conventional engineering to succeed and gives the regime-change hypothesis a test that can distinguish outcomes.

10 — What this work means

Taken together, this work asks a precise question: when individually valid agents, controls and decisions meet a changing ecosystem, what information must remain qualified so that the resulting decision is still justified? Six deliberately different failure cases provide the challenge set. The foundation, principles and S1–S14/T1–T4 requirements provide a common, reviewable answer; the negative and positive paper routes show what that answer would demand at each decision boundary. Ablations and countermodels test whether the principles are redundant within the declared models, while the extension tests specify when a result can be carried to a larger, smaller or different-domain case. The contribution is a traceable method for moving from failure mechanisms to falsifiable ecosystem-level requirements, together with a defined route to comparative evaluation.

This matters because a valid identity, policy decision, attestation, local safety control or human approval can retain its local value while its scope, evidence, authority, dependencies or timing no longer support the combined action. A decision-scoped handoff that preserves those qualifications could make the exact point of requalification visible without demanding omniscience or stopping everything. This gives research and pre-standardization a precise test: locate requalification and measure failure routes, delay and review burden. The proposed EA benefit is maintaining justified decisions as the regime and the model of relevant conditions change, potentially with lower decision burden. The current shared-logic symbolic comparisons do not resolve that benefit. Matched runs of independently implemented candidates can do so: a conventional peer that also maintains qualification at equal or lower burden narrows or removes the EA advantage in the tested scope. A new semantic primitive calls for revision.

Integration and remaining work. PR #1 was merged at 59ca3c4, preserving the accumulated commits. The canonical originals now link their own reinforcement and A08–A12; the READMEs, manifest and A17 map carry the revised evidence limits. A19/A21 received cross-reference corrections; their demonstrations and A20/A22 were preserved. A12 §8.1 links successful merge runs for documentation, 00L, Stage 0 and 00K. Those green checks do not resolve the P3/P5/P6 semantic questions, provide independent review or establish product-scale effectiveness.

Appendix — The claim set and a reading key

The short statements below make the central objects readable without leaving this article. The linked canonical sources retain the full semantics and frozen test contract; these summaries make the argument readable and leave empirical outcomes to the corresponding tests.

Six operational principles (P1–P6)

P1 — Qualified determination and explicit residual. State which subject, proposition and receiving decision the evidence actually supports; distinguish sufficient, insufficient and inconclusive evidence, and preserve what remains unresolved.

P2 — Bounded unresolved effort and viable oversight. Give search, review and escalation a finite, decision-relevant stopping rule within real capacity and the useful response horizon.

P3 — No false closure from unresolved state. Missing, stale, conflicting or otherwise material uncertainty must not become permission or certainty merely through timeout, approval or agreement among weak signals.

P4 — Qualification-preserving authority basis. At a receiving decision, retain or obtain a current, verifiable, decision-sufficient and non-amplifying authority basis. Full historical delegation lineage is one way, but a legitimate scoped capability or authoritative attestation may suffice.

P5 — Material-change requalification at time of use. Reassess a once-valid decision when its material basis changes before commitment or execution; an unexpired token or correctly queued job is insufficient on its own.

P6 — No local-to-ecosystem promotion or silent substitution. Do not treat a local result, repeated claim or correlated majority as ecosystem truth beyond its scope; preserve dependencies, independent evidence and conflicts among principals. The canonical six-principle statement and derivation specify the formal scope.

The fourteen canonical requirements, in one line each

S1 — Establish authority origin, scope and current applicability.

S2 — Preserve the principal’s preferences, hard limits and reviewable decision basis.

S3 — Distinguish normal escalation from governed exception after material regime change.

S4 — Make human oversight authorized, capable and timely.

S5 — Keep unresolved material state visible and contain its propagation without a blanket veto.

S6 — Support bounded, privacy-preserving trust determination across participants.

S7 — Link principal, organization, role, actor and representation correctly.

S8 — Bound subdelegation by inherited purpose, scope, time and revocation.

S9 — Compose multiple principals without silent substitution or false convergence.

S10 — Track commitment state and requalify material change before action.

S11 — Preserve policy, objective and preference integrity across domains.

S12 — Support reconstruction, challenge and repair without rewriting history.

S13 — Separate original authority history from later intervention history.

S14 — State what the evidence proves for this decision and how unresolved branches are arbitrated. The frozen Requirements 00 contains the authoritative wording, the T1–T4 sufficiency conditions, H1–H6 hypotheses and KPI protocol.

Three full requirement clauses and their test measures

S8 (00H; P4): “Bounded subdelegation and non-amplification. Preserve purpose, scope, time, hard limits, relevant preferences and revocation conditions through agent chains. Subdelegation must not manufacture authority absent from the original principal.” On the declared refund decision, the S8→T2/T3→H2/H4 route records inherited-indeterminacy detection, completeness of authority/scope/expiry fields, qualification loss and unauthorized-response rate. A legitimate wider mandate is a positive control; this requirement does not itself issue a grant.

S10 (00I; P5): “Commitment state, material change and normal escalation. Distinguish recommendation, negotiation, reservation, binding commitment and execution; detect when a material change of cost, route, time, context, policy or authority requires confirmation, revalidation, cancellation or change of authority. Silence or delay is not permission.” Its S10→T1/T2/T4→H1/H5/H6 route reports material-break recall and precision, false continuation, freshness, requalification latency, response margin and decision-relevant evidence yield. An unchanged action must remain a passable positive control.

S14 (00J; P1, and a common assessment layer): “Evidence-to-decision assessment. State at each transition what must be demonstrated, what evidence is required, whether it is sufficient, insufficient or inconclusive, and which decision it supports. Record any Type-1↔Type-2 transition and the arbitration rule—material dependency, legitimate veto, timeout/default, owner, expiry, re-entry and evidence-versus-authority role—that determines what advances. This is a reusable assessment layer, not a mandatory runtime architecture.” Its applicable S14→T1/T2/T3/T4→H1–H6 route reports the common outcome vector, branch-specific KPIs and the arbitration/precedence record. In 00J, provenance for one proposition cannot silently prove an enforceable-rights claim about another. These canonical clauses and the KPI mappings in Requirements 00 give later implementation claims a precise, shared test boundary.

Artifact and evidence key

00E–00J — Six constructed scenario documents, each with a quality plan and limited external-corroboration addendum; their freeze/draft designations differ by file. The named technology profiles are source-reviewed comparison designs, not product run logs.

00D v0.2 / v0.3 — Canonical working benchmark and bounded draft extension, respectively; B0–B3 specify fair matched comparisons, not four observed product scores.

01 / 02 / 02A / 02B — Foundation, principles/control matrix, derivation and relative syntax closure; conceptual and formal arguments within declared assumptions.

00 Requirements / 00K / A02 / A19 / A21 / A23 / A24 — Frozen test contract; principle-ablation design; documentary coverage; semantic traceability; relative requirement closure; audited partial finite reverse conformance; and information-gain analysis. These are distinct argument layers, not product validation.

Updated reading key: consult revised A23 and revised A24 for partial finite sufficiency and its semantic limits; Sections 4 and 7 distinguish the current results from the retained historical certificates.

A15 / A14 / A22 / formal certificate — Executable symbolic testbook, recorded CI reproduction, 64-signature Boolean diagnostic analysis and the historical, superseded finite-model check of 65,536 states. The 379 tests are symbolic regression/falsification cases, not 379 live incidents or independent trials.

Numerical distinction: the old 65,536-state certificate remains a historical reference. Its successor is the 33-field projected clause audit with 22 clause/target comparisons and a separate 11-test suite. The 64-signature independence result and 379-regression campaign are different evidence layers.

A25 / A26 — Conditional failure-family and success-case transfer tests; proposed cross-domain analogies must pass admission checks before a result transfers.

Transfer distinction: A25 and A26 require a separate base guarantee for a universal transfer claim; the executed A3/A4/A6 branches provide bounded evidence, not that universal guarantee.

RS-00E-Q1a — Pre-registered fixture with bounded Stage-0 execution and released traces; its synthetic adapters share decision logic; EA-ITP-01 — proposed interoperability test profile. Neither is a deployed standard or a positive product run. The A17 proof and reproduction map is the quickest route through the test evidence.

Complete reinforcement index: 00L A00–A12 and six paired traversals; canonical-original integration map; released replay traces; executed Stage 0; cumulative change register for the article and README. These added references cover the full accumulated reinforcement, corrections and merge, while preserving the accumulated evidence and historical links.

Part II A reference framework others can use

The requirements study could have ended with a statement of what an agentic ecosystem should preserve. The work continued with a different question: what would let someone else actually build toward that, without having to reinvent the vocabulary first? The fourteen requirements are deliberately solution-neutral — they say what must be preserved, not how. Something has to fill that gap. Ecosystem Awareness, Positioning, MSCA and their neighbouring layers are offered as a shared architectural vocabulary for exactly that gap, in the spirit of the OSI reference model: OSI illustrates how a shared reference can describe responsibilities and relationships across different implementations. The unusual names below read the same way, for the same reason — they organize responsibilities that a candidate solution must address, without making these particular names or component boundaries additional requirements.

The reference follows the questions a bounded participant needs to answer. A participant that cannot see the whole ecosystem still has to act, so Ecosystem Awareness gives it a name for exactly what it currently knows, how firmly, what it could still find out, and what stays out of reach — kept current, not fixed once. That position only matters if the participant knows which moves are even open to it: an Agentic Citizenship Contract states, for a given participation domain, which roles and objectives a participant may hold and under what conditions that standing can be revoked. Awareness and that boundary together let Ecosystem Positioning ask the only question that matters next: does the participant’s current role still fit, and if not, which bounded move — realign, hand off, escalate, migrate — would actually reduce the risk that matters to its declared objective.

Positioning does not answer that in a vacuum: it reads the participant’s own Minimum Sufficient Control configuration (its objective envelope, operating assumptions, coordination scope, intervention mechanisms and enabling means) and its Ecosystem Cartography, the dependency map it currently has confidence in. A recurring check, Regime Awareness, triggered by events, thresholds or periodic review, asks whether the operating frame all of this was qualified under still holds; when it detects material departure, it hands positioning a qualified directional signal, not a final operating-posture decision, so the participant re-evaluates rather than being told what to do.

Once a participant decides to move, it has to tell others what changed without exposing everything it knows — that is Ecosystem Signalling — and when what is needed is not a new role but a correction to something already wrong, the same bounded position decides what may be done about it without inventing new authority for itself. No layer here assumes a controller watching the whole ecosystem; each one only ever answers for the participant that runs it.

Two of these steps are the ones a skeptical reader asks about first, and both already have explicit decision rules and falsification conditions that a designer can use to examine a proposed implementation. The Objective-Conditioned Agentic Gradient — the step that ranks which move is worth making — comes with a ten-item falsification list identifying ways a proposed calculation can violate the law: treating confidence as proof the residual is empty, averaging away a risk an unrelated gain cannot compensate, letting a large ecosystem change register as material without checking whether it reaches the participant’s own dependencies, claiming a globally optimal ranking from a position that was only ever local. It also keeps opportunity, admissibility, authority and execution as four separate questions on purpose: a high-ranked move outside current authority may justify an authorization request, never a license to act.

Deciding whether a participant is still operating “normally” gets the same treatment: Operation & Repositioning closes a named catalogue of Type 0 (honest residual), Type 1 (unresolved past the useful point) and Type 2 (premature certainty) conditions into exactly three postures — Normal, Containment or Migration — before the corresponding repositioning result is emitted; evidence and authorization requests can be signalled during the cycle, and a move is only executed after a separate authority gate confirms it is allowed. A third, smaller example belongs in the same list: Minimum Sufficient Control keeps “no sufficiency determination has yet been established” as its own explicit state — UNASSESSED, alongside SUPPORTED, FAILED and UNRESOLVED — rather than collapsing it into a default pass or fail. These checks give designers concrete criteria for evaluating a calculation; its estimator and operational feasibility must be established for the chosen setting.

The connections between these layers are specified, not asserted. The general interfaces and Epistemic Handoff Descriptor state what a receiving participant needs to understand about a result and its limits; the EA–MSCA–Regime Awareness composition profile binds their contributions to the same decision, scope, versions and time interval, with proposed, falsifiable race-condition tests. For a delayed repair, this lets a designer ask a concrete question: which component retains the original justification, which one exposes that a material change occurred, and which requests requalification before the authorized owner acts — a reviewable contract rather than a restated aspiration. Worked material carries this further: four UC-EA validation profiles examine changed operating frames, incomplete or dependent evidence, viable human oversight and cross-scope composition; a DAOS masterclass applies the same lenses to a shared model case; and the Success Model Cases in Part I give each failure family a paired example of the requirements-conforming route, not only the route to avoid.

Some of this is already runnable rather than only specified. The symbolic ablation suite lets a reviewer remove a control and inspect what breaks within a declared fixture; the 00L traces and Stage-0 harness discussed in Part I provide the same kind of inspectable record, with the same honest limit noted there — they verify declared instrumentation and reproducible behaviour, not yet parity between independently built architectures. Two further routes extend the toolkit beyond what has been executed: EA-ITP-01 specifies a concrete cross-implementation handoff test, and the Decision Boundary Challenge sets out challenge packs and a progression toward sandbox comparison. A team does not need to wait for someone else to run these — the fixtures, the falsification lists and the test profiles are built to be picked up, argued with, and extended by anyone who wants to try a different implementation and see where it agrees or disagrees.

Within trust and interoperable identity, this direction aligns with FG-TIDA‘s remit — architectural approaches, trust-lifecycle frameworks, machine-readable trust information, use cases and assessment criteria — and the EA specification-preparation map organizes the proposed contribution into functional responsibilities, handoff semantics and conformance material for review through the group’s own process. That is a route for consideration, not a claim of adoption.

None of this makes Ecosystem Awareness the only way to satisfy the fourteen requirements, and it is not offered as one. It is a candidate reference vocabulary, informed and challenged by the process of failure-mode construction and ablation described in Part I, for a problem — maintaining a decision’s justification as the regime around it shifts — that has to be solved somehow. A reader who satisfies the same requirements through an entirely different architecture, using different names or component boundaries while preserving the required distinctions, is exactly the outcome this reference hopes for.