Skip to content

IEEE / ISO/IEC/IEEE standards analysis series

A standing series that reads gzkit against a corpus of modern systems- and software-engineering standards. The purpose is not standards conformance. The standards are used as a mature body of engineering vocabulary, to answer four questions about this project:

  1. What has gzkit already rediscovered correctly?
  2. What engineering objects has it conflated?
  3. What important concepts are missing or weak?
  4. What existing process or artifact can be simplified or removed?

The governing principle is use the standards to improve engineering clarity, not to import bureaucracy. Every piece in the series carries a section naming the machinery gzkit should not adopt, because most of what these standards contain is acquisition, contractual or large-organization apparatus that a project with one operator and several agents would only be harmed by.

Start here

Read FINDINGS.md first. It is the canonical register — what the investigation currently accepts as supported — and it is the only file needed to know the evidential findings. For the accepted design direction and the remaining candidate mechanics, read the grounding pivot, authorized at Q-17. The numbered pieces are evidence behind the findings, not the current design. For the broader design conversation, its stakes, and the alternatives that led to the anchor, read the dated session record.

File Tier What it is
FINDINGS.md canonical The findings, with status and evidence pointers. Living; amended in place
DISAGREEMENTS.md canonical Where Agent 0 and Agent 1 disagree, preserved unresolved
OPEN-QUESTIONS.md canonical Operator questions and rulings, including successive-change evaluation (Q-15) and the FDAU origin and retained hypothesis (Q-16)
design-candidates.md accepted grounding; proposed mechanics Q-17's grounding/anchoring pivot, requirement classification, revision/acceptance model, alternatives, review, and pilot specification. Ownership relationship adopted in Magna Carta §3; catalog mechanics remain proposed
Design pivot session deliberation and provenance Why the pivot arose; recovery choices; conventional ADR, capability and mADR alternatives; attachment strands; source limits; unresolved cost. History, not a second candidate home
consequence-bands.md canonical, PROVISIONAL Operator-ruled consequence scale (Q-04). Kept PROVISIONAL by operator ruling 2026-09-23: three rows have no finding behind them, and four rest on F-022 (OPEN) or F-021 (DISPUTED); the lift is decided when Phase 4 is authorised
01-…, 02-… historical Raw investigation record — what an agent said, frozen at its date
Astra's adversarial review historical Agent 1's Phase 2 report, as deposited. Raw record — challenge input, not a second register
Act 1 cold read historical Agent 2's Act 1 verdict on this register, frozen at its date. Raw record — evidence about the register, not a finding about gzkit
Dex Horthy transcript excerpts historical, external testimony Timestamped source excerpts behind the successive-change evaluation criterion; not verified findings about gzkit
raw/ historical Index and tier statement for raw reports
references/ reference Vendored, freely distributable companion texts. Currently FAA DOT/FAA/AR-08/32, the Requirements Engineering Management Handbook
02-…-evidence/ tooling measure.py — the re-derivation script piece 02's figures come from. Not investigation narrative: it holds no conclusions, carries no literals from its authoring date, and reports whatever tree it is given

Canonical is not normative. A finding here records what the investigation supports. It binds nothing, changes no rule, and authorises no work.

This is not a standards-compliance initiative, and no phase of it may become one. gzkit is not pursuing conformance with any standard in this corpus, will not claim conformance, and is not obliged to adopt anything a standard contains. The corpus is used as vocabulary. FINDINGS.md F-032 and F-033 are the standing guard on this, and F-032 binds Phase 4 as a filter.

Grounding pivot — 2026-09-25

The operator accepted the roles of independent requirements catalog, mADR as bounded advance with design rationale, brief as requirement assignment, plan/specification/tasks as execution detail, and release record as shipped content with evidence. They then directed:

make the results from this conversation a GROUNDING/ANCHORING pivot.

Q-17 preserves the wording and authorization. design-candidates.md is the anchor: accepted direction first, proposed mechanics and evaluation separately labeled. Read it when resuming this design; do not reconstruct the direction from earlier alternatives or infer implementation approval from acceptance of the roles.

Retention amendment, 2026-09-25. At the operator's request for richer capture, the session record preserves the wider reasoning, and the option inventory separates timing, replacement scope, ownership, naming, and implementation order. Conventional ADRs are distinct from capability ownership; strong mADR/release association is distinct from deriving package version from an ADR identifier. The anchor does not settle all these choices or establish affordable recovery.

All investigations continue. IEEE can substantiate and challenge the design despite unconventional names; the test work addresses real verification debt; Magna Carta supplies release planning logic. TDD, BDD, and DDD remain grounding influences. Phase 4 is open for this bounded design and its full-triad review. Pilot, adoption of mechanics, migration, runtime rules, and campaign sequence are separate. No finding is promoted by the design decision.

Doctrine amendment, 2026-09-25. The operator subsequently directed the explicit amendment of Magna Carta §3 and its governing surfaces. Independent requirement authority and reference-based assignments are now the governing relationship. Brief-local acceptance criteria and current identifiers, parsers and proof bindings remain in force during migration. This narrow adoption does not authorize the proposed pilot or implement the catalog; Q-17 records the distinction.

Investigation origin and retained hypothesis: FDAU

Recorded 2026-09-23 at the operator's request; Q-16. The operator identifies xplane-fdau's emergent organization as the origin of this IEEE investigation: "fdau's accidental design is what brought this ieee analysis to bear and to light." The standards were gathered to help determine what gzkit ought to do after that comparison exposed a more intelligible separation of engineering objects. FDAU is an originating example to examine, not a reference implementation to copy or a demonstrated standards-conformant system.

Hypothesis to retain. Requirements and specifications describe obligations; ADRs record architectural decisions; bounded briefs assemble the requirements, relevant decisions, specifications, plans, tasks, and evidence needed for a change; releases identify selected delivered increments. The useful work-package function of OBPIs can survive without making an ADR the mandatory container or coupling its identity to a release. Semantic capability areas and hierarchical Nx.y identifiers may make the work easier to navigate without encoding release membership. This is a synthesis of the operator's framing, preserved verbatim at Q-16. It is not an approved replacement ontology, identifier scheme, or migration.

Accepted relationship, 2026-09-25. The operator accepted that a brief presents requirements together with their work and evidence by referencing their authoritative definitions. The verbatim acceptance and accompanying request for deeper option analysis are recorded under Q-16. Storage, identity, capability boundaries, and migration timing remain open. The operator subsequently clarified that using capabilities as organizing containers is itself uncertain: SEI is one possible anchor, while IEEE is sought for authority, standards, and principles. Q-16 preserves that wording; the capability model is not the assumed destination of this investigation.

Narrowed framing — 2026-09-25. The operator subsequently identified the needed separation as ADRs, the requirements corpus, and release/backlog planning, and raised making the smallest adjustment that achieves it. Poor tests are technical debt; IEEE comparison is reflective maintenance; Magna Carta is release planning. Q-16 preserves the wording and the retained TDD/BDD/DDD grounding constraint. These efforts do not require a single broad recovery program or a capability hierarchy. Implementation remains unselected.

Earlier alternative under comparison — 2026-09-25. The operator subsequently asked to explore retaining the ADR as a modified delivery increment (mADR): one active increment, other prospective work pooled, corrections through GHIs and patch releases, and reusable requirements held separately and referenced by briefs. Q-16 preserves the request and proposed composite notation verbatim. At that point neither alternative was selected; Q-17 records the later accepted grounding that retains bounded advances with independent requirements. Source grounding must distinguish traceability principles from FDAU's particular naming scheme. The operator then accepted the distinction among catalog requirement, work package, and requirement assignment (Q-16), while asking whether the proposed discipline removes complexity or merely preserves the old model through added exceptions. The ADR name and OBPI-as-sortie terminology remain open. Q-16 also retains the operator's original one-brief/airlock purpose: bound one serious agent foray, while catalog requirements describe evolving product needs. The comparison must account for that control function, including requirements discovered during planning, rather than judge the structure by ADR terminology alone.

Inspectable comparison source. The reviewed FDAU snapshot is 49eec3557145d9825c7026aa4dfee718443d433f: backlog method, roadmap, and backlog. Its retrospective also contains proposals for other projects; distinguish those from implemented FDAU behavior. Its child-and-gate tracking does not itself establish durable individual requirement identity or complete bidirectional traceability. The operator attributes the organization to the agent's emergent design, rather than to a capability supplied by Superpowers; that provenance account is not an audit of Superpowers.

How it remains in consideration. Carry this question through the existing phases, when each is authorised, and record its disposition rather than allowing it to disappear into a generic recommendation to improve governance:

Existing stage Treatment of the retained hypothesis
Measurement program Read M-A and M-B against the durable-obligation / bounded-work distinction; retain evidence that weakens it as well as evidence that supports it. This does not commission either measurement
Phase 4 — design Explicitly compare a model with separately identified requirements, decisions, work packages, and releases against the current model and narrower repairs. Examine brief relationships and semantic Nx.y navigation without assuming that literal composite keys or renumbering are needed. Record what to retain, revise, or reject and why
Phase 5 — pilot Use Q-15's existing successive-change sequence to test whether a changed requirement, deferred brief, and later release preserve identity, approval history, and evidence. Include a case where completed work is not yet released
Phase 6 — independent evaluation Reconstruct what was required, decided, worked, verified, and shipped from retained artifacts. Check both semantic separation and the cost of maintaining the links; inspect FDAU's limitations as possible failure modes
Phase 7 — operator decision Give this hypothesis an explicit adopt / revise / reject disposition supported by the pilot evidence; record an inconclusive result as inconclusive

Relationship to the existing record. This bears on F-001, F-002, F-005, F-008, F-012, and F-015; it neither supplies new evidence for their status nor settles their qualifications. Q-08 already rules against deriving a package version from an ADR identifier, and Q-10 selects a ledger release record. Preserve those rulings while investigating the wider separation. Q-12 still places design candidates in design-candidates.md only after Phase 4 is authorised. That authorization is now recorded at Q-17; the earlier hypothesis is retained here as the investigation's history. Runtime governance is unchanged.

Investigation phases

Phase Purpose Lead Status
0 Frame the investigation operator complete
1 Primary forensic assessment Agent 0 complete — pieces 01, 02
2 Adversarial review of Phase 1 Agent 1 (Astra) complete — report received 2026-09-22
3 Reconciliation and canonicalization Agent 0 complete 2026-09-23 — reconciled, disagreements recorded, stop condition met (Q-14)
4 Develop the grounding pivot and candidate against Q-15, carrying the FDAU origin (Q-16) Full triad bounded drafting and review authorized 2026-09-25 (Q-17); mechanics not adopted
5 Bounded pilot across successive changes (Q-15) — not authorised
6 Independently evaluate the pilot against the adopted criterion (Q-15) — not authorised
7 Adopt / revise / reject — not authorised
8 Incremental migration — not authorised
9 Measure and periodically reassess — not authorised

Success across successive changes

Adopted for this investigation, 2026-09-23. The operator asked, "how can we inculcate the more compelling aspects of Dex's experiences and conclusions here?" and accepted the proposed criterion and pilot approach with "okay, let's do that then". Operator: g0. The ruling is registered at Q-15. The accepted criterion is:

A remedy must preserve required behavior, authority, and evidence across successive changes. Its evaluation must examine whether later work remains understandable and verifiable, including by an agent that did not author the remedy. Additional rules, artifacts, and checks are not themselves evidence of improvement.

This is an evaluation commitment within the investigation. It adds no runtime gate and does not initiate Phase 4 design, a pilot, or implementation. The phase authorisation table and the Phase 6 independence rule remain in force.

Source and evidence boundary. The operator supplied an AI Native Dev conversation attributed to Dex Horthy. Selected transcript excerpts retain timestamps and original wording. His experiences motivate the questions below; they do not establish their answers for gzkit. No finding gains a status from this testimony. Benchmark figures, model rankings, and universal claims about context length are not adopted as evidence.

Lesson to test Application in the existing phases
Generating a system and maintaining it are different capabilities Phase 5 exercises successive changes against the same evolving pilot; Phase 6 evaluates later changes as well as the first success
Passing checks can miss architectural deterioration Phase 6 examines actual dependencies, state ownership, and failure paths alongside results; the explanation must point to implementation evidence
Planning has greatest value where mistakes are expensive to reverse Phase 4 gives explicit attention to requirement identity, release authority, and historical evidence, distinguishing them from cheaply reversible details
Existing code supplies examples for later agents Phase 4 identifies the implementation patterns a later agent is likely to copy; Phase 6 checks whether those patterns agree with the intended design
Corrections should improve subsequent work Phase 6 checks whether operator corrections recur, and whether existing instructions, skills, or verifiers address the cause before proposing another mechanism

Where the testimony lands in the register. Reconciled 2026-09-23, after the criterion was adopted. Each excerpt is mapped to the row it bears on; none is evidence for that row. Verified mechanically at reconciliation: the register held 35 rows at 15 QUALIFIED / 14 OPEN / 2 CONFIRMED / 2 DISPUTED / 2 REJECTED before and after the deposit, so the boundary above — no finding gains a status from this testimony — holds as stated.

Excerpt Bears on What it is, and is not
11:11 accumulated difficulty F-025, F-007 A different codebase decaying over months. F-025 attributes duration to no cause and lists eleven live candidates; this adds testimony to that list, not a measurement removing anything from it
19:10 "expected pain" — chance × pain consequence-bands.md, Q-05 Contrast, not corroboration. Horthy's framing is probability-weighted; the bands are D + R and deliberately are not, and Q-05 ruled the 16085 consequence threshold weaker, not merely different. A reader must not take this as support for re-proposing it — that re-proposal is prohibited
28:51 "no fast oracle for software maintainability" F-028, M-D A claim about feedback speed, distinct from F-028's claim about vocabulary — which Astra corrected: 29119 Part 1 does define oracle. Bears on M-D's unresolved method problem, the highest-risk item in the program
34:45 instructions unfollowed because context is too large F-023, F-024 The closest match to an existing finding. F-023 is QUALIFIED — Astra withdrew the 92,000-token figure and its dominance claim; testimony does not restore a withdrawn measurement
38:29, 38:38 bad code degrades future work F-020, F-021 Motivates the existing code supplies examples row above. F-020 chose no authoritative side on any of its 22 rows, and that restraint is unaffected
39:44, 39:50 analyse session traces for recurring corrections F-006, F-021, M-C The practice gzkit's insights ledger would serve. M-C already asks the sharper question — whether the insights ledger has ever changed an outcome. Testimony describes the practice; M-C is the measurement

Two reconciliation notes.

It agrees with the register's principal safeguard. The criterion's closing sentence — "Additional rules, artifacts, and checks are not themselves evidence of improvement" — is F-032 restated from outside the investigation. F-032 is the challenge table's only unqualified CONFIRM and binds Phase 4 as a filter; it now carries independent agreement from a practitioner who reached it by different means. That is corroboration of a method commitment, and it is still not evidence for any finding.

The excerpts inherit F-022's defect. They cite a SHA-256 of a supplied attachment that is not in this repository, so neither the excerpts nor the selection behind them can be re-derived or audited from the repo — the shape F-022 names, "an evidence record nobody else can retrieve is not evidence", applied to this investigation's own record. It is recorded rather than repaired: vendoring a third-party transcript is a licensing decision for the operator, and § Standards corpus already declines to vendor licensed texts for the same reason. A future reader should treat the excerpts as attributed, not as verifiable.

Phase 5 pilot shape, for when authorised. Demonstrate an initial release within the bounded pilot, then exercise a changed requirement, a deferred item, and a later release. Carry forward the resulting artifacts rather than resetting to a clean example between changes. The pilot must expose whether requirement identity and approval history remain understandable, release membership can be reconstructed, and later changes trigger unexpected repairs or repeated requests for settled operator decisions.

Phase 6 evaluation. An agent that did not author the remedy locates its governing decisions and reconstructs the relevant requirement and release history from the retained artifacts, without private explanations from the designer. This exercise is separate from the completed Act 1 cold read. Record correctness, regressions, reconstruction effort, and operator corrections separately, with evidence pointers and the scope of each observation. Use the information-need-first approach already proposed in M-G; establish a comparable baseline before claiming improvement and record relevant differences in task, model, and context. Reduced effort alone does not establish improved governance, and one successful sequence does not prove general maintainability.

The evaluation feeds the existing Phase 7 adopt / revise / reject decision. An inconclusive result remains inconclusive. This adds neither a review stage nor a presumption that new registries, rules, or checks are the remedy.

Roles

Agent Role
Agent 0 Primary investigator (Phase 1) and reconciliation lead (Phase 3)
Agent 1 — Astra Adversarial reviewer (Phase 2). Run separately by the operator; deposits reports where its tooling puts them — the Phase 2 report is at this directory's root by operator ruling (Q-13), and raw/ states the tier
Agent 2 Later independent reader, then design participant — in that order (see below)

Agent 0's lead is procedural: it does not make Agent 0's findings authoritative over Agent 1's challenges. Agent 1's critique enters as evidence and challenge input, not as a second canonical register merged automatically into the first.

How the three roles compose across phases (binding)

The three agents are not three workers taking turns. They are three positions relative to the evidence, defined by how anchored each is:

Agent Position Anchoring
Agent 0 constructs Most anchored — holds all the context, so least able to see its own assumptions
Agent 1 attacks Anchored by Agent 0's output; it argues on that terrain
Agent 2 reads cold Least anchored — the only one that can say whether the record stands without its authors

Phases 1→2→3 worked because construct → attack → reconcile is a sequence in which each step consumes the previous step's output — not because three agents each took a turn. The triad recurs only in analysis phases, and rotating all three through every phase would make the investigation pay its own F-023 entry cost three times over for passes that mostly add nothing.

Phase Composition
4 — design Full triad. A design needs attacking more than findings do: it has no evidence yet, only reasoning
5, 8 — pilot, migration Execution. Agent roles barely apply; the operator runs these
6 — evaluate pilot Independence-critical — see the rule below
7, 9 — adopt, measure Operator decision, and recurring measurement

Phase 6 independence rule (binding)

Agent 0 must not lead the evaluation of a pilot it designed. Phase 6 is led by Agent 2, or by Agent 1 — never by the agent that produced the Phase 4 design.

Whoever designs a thing cannot judge whether it worked. This is IEEE 1012 Annex C's independence point, and it is the same failure mode observed twice in this investigation on 2026-09-22: Agent 0 used its own unsettled output as a settled input, in consequence-bands.md and — as Astra independently found — in piece 01 §§ 4, 6 and 8. A constraint on the lead is exactly the kind that goes unwritten unless it is written early, which is why it is recorded here before Phase 4 exists.

Agent 2 sequencing constraint (binding)

Independent reader and design participant are two jobs, and doing them in the wrong order destroys the first. The moment an agent has a stake in a design it can no longer give an unanchored reading. Agent 2 therefore works in two declared acts, in this order:

Act 1 was taken on 2026-09-23, by operator ruling, immediately ahead of the Phase 3 reconciliation pass and for exactly the reason below. Its verdict — the register stands, with load-bearing qualifications — and its ranked gap list are recorded at act1-cold-read-2026-09-23.md, with a dated disposition table separating the reading from its consequences. Act 1 cannot be retaken: the register it read no longer exists in that form.

The reasoning that put it first is retained, because it will apply again to any successor register: Act 1 is not phase-bound, the register exists, and the cold read is most valuable before reconciliation rewrites it — what is on disk on the day of the reading is what a newcomer would actually inherit.

Act 1 — the cold read. Agent 2 reads only this README, FINDINGS.md, DISAGREEMENTS.md, OPEN-QUESTIONS.md and consequence-bands.md. It reads neither the numbered pieces nor anything in raw/. It then answers one question and records the answer before doing anything else:

Can the investigation's current position be stated from the canonical register alone — and if not, exactly what could not be determined?

Act 2 — design participation. Only after Act 1's verdict is recorded may Agent 2 read the raw record and join design work, as a declared second role. Act 2 depends on Phase 4 authorization, now supplied for the bounded pivot at Q-17 (2026-09-25). The current fresh design reader is a new participant, not the historical Act 1 agent. That cold read remains a dated test of the register and is not repeated or claimed as this design's pilot evaluation.

Why the order is binding. Act 1 is the only falsifiability test the register has. It checks whether this investigation exists durably in the repository or still only in Agent 0's and Agent 1's context. A failed cold read is a finding about the register, and the register is then what gets repaired — not the reader. It is also the closest thing available here to genuine V&V independence: Agent 1 read Agent 0's report and is anchored by it, whereas Agent 2 in Act 1 is not. IEEE 1012 Annex C names this problem, and piece 01 found gzkit's honest self-description is already the embedded form.

Current gate. Phase 3 is complete as of 2026-09-23. The Phase 2 report arrived 2026-09-22; the reconciliation pass mapped all 25 rows of its § 2 challenge table onto FINDINGS.md, moved 19 statuses, and recorded eight disagreements at D-01 … D-08. Fourteen rows remain OPEN and stay there: Q-14 rules that OPEN is a settled disposition for a row the review never reached. Two rows are DISPUTED (F-006, F-021) and must not be read as settled in either direction — a disagreement recorded is not a disagreement resolved.

Phase 4 bounded design is authorized at Q-17, 2026-09-25. The authority is the operator's request for a plausible design pivot and grounding record, not Phase 3 completion or a finding's status. The full triad constructs, challenges, and reconciles the candidate. This does not open pilot execution, adoption, or migration, promote a finding, or turn proposed mechanics into runtime rules. Q-09's deferral and Q-15's evaluation criterion remain.

Phase numbering, one caution. Piece 01 § 12 originally titled its measurement plan "PROPOSED PHASE 2 INVESTIGATION PLAN", which meant measurement rather than adversarial review. Those items were renumbered M-A … M-H on 2026-09-22 by operator ruling. A pre-2026-09-22 reference to "Phase 2" in this series may mean the measurement program.

Terms this series cites as binding

Two terms are cited in this register as binding constraints without being stated in it. Both are gzkit doctrine, not investigation vocabulary. The authority is the pointer; the gloss is a reading aid and is ILLUSTRATIVE, never authoritative (AGENTS.md § Governance doctrine surfaces). Cite the authority, not the gloss.

IRON LAW

Cited at § Currently prohibited, at Q-05 and on F-018.

  • The rule in force: root AGENTS.md § OBPI Acceptance Protocol — "Only the operator initiates and executes OBPI work through gz-obpi-pipeline." The rule in force does not use the name, which is why searching this repository for "IRON LAW" finds commentary rather than canon.
  • The name, and the operator's verbatim wording: the corpus, .gzkit/corpus/AGENTS.md.jsonl, entry corpus-operator-doctrine-verbatim-canon-2026-08-23T14:33:19.... Read it with the content skills; the corpus is not hand-edited.
  • Dated record of the compression: context-audit-2026-09-12/root-doctrine.md, which carries the original beside what landed in AGENTS.md.
  • Gloss, sufficient to read Q-05: only the operator may initiate OBPI work, across every arm of it — claiming or releasing locks, pipeline markers, starting or completing TASKs, dispatching implementers or reviewers, editing a brief. A narrow operator-named task that happens to fall inside an OBPI's scope is not initiation. It is a blanket prohibition on initiation and delegates nothing, which is exactly why Q-05 rules the 16085 consequence threshold strictly weaker rather than merely different: a threshold defines a line below which authority is delegated, and this rule delegates none. Advisory — no mechanical witness distinguishes operator-initiated from agent-initiated OBPI work today, which is itself the condition F-018 is about.

Architectural Boundaries 1 and 2

Cited at Q-03, at Q-09 and on F-006.

  • In force at: root AGENTS.md § Architectural Boundaries, as six numbered items.
  • As written: 1. "Do not promote post-1.0 pool ADRs into active work." 2. "Do not add more pool ADRs to the runtime track."
  • Only 1 and 2 are cited by this series. Boundary 6 — "Do not let derived views silently become source-of-truth" — is the one F-021 and consequence-bands.md reason inside without naming, and a Phase 4 design touching either should read it.

The measurement program — M-A … M-H

Eight findings route their disposition to a measurement item, so the items are summarised here. Authoritative text: 01 § 12. The program is proposed and not executed, except as noted. It proposes no implementation, and 01 § 12 closes by naming what it must not do: no replacement taxonomy, no new lifecycle, no new tooling — "adding machinery is the failure mode most consistent with this repository's history."

Item What it measures Follows State
M-A What a persistent requirement object would have to carry, by classifying ~200 REQs and ~200 FAIL-CLOSED constraints against 29148 § 5.2.5/§ 5.2.6 F-001, F-002 proposed. Declared open risk: the answer may be that most REQs are correctly transient and the persistent layer must be authored fresh — a larger finding, not a smaller one
M-B The deleted specification corpus — reconstruct the deleted briefs, classify what was lost as durable vs transient, check whether any deleted REQ is still cited by live code, a test or a @covers F-003 proposed. Settles D-03
M-C Whether the six intake surfaces can be reduced, by tracing where sampled items actually end up; duplication and mortality across surfaces F-006 proposed. Bears on D-01
M-D The claim behind each gate and each enforcement claim, in 15026-2 § 5.3.3 form, and whether a non-agent witness exists F-019, F-021 proposed, gated on its own method problem — the prior audit disqualified itself as "a stochastic surface auditing a stochastic surface." Highest-risk item. Bears on D-05
M-E Whether an architecture description is warranted, or whether extending gz drift/gz covers reaches the same property more cheaply F-013 proposed. Declared bias: the correspondence route is likely cheaper and should be tested first
M-F The real duplication between Gate 4 and Gate 2 — run behave with coverage instrumentation against the unit suite, and count the behave-only REQs (no finding — see below) EXECUTED 2026-09-23 — 03. This is the home of D-08, whose disposition it does NOT move
M-G Baselines for the metrics Phase 1 could measure only once, each stated as an information need before a measure (15939 § 6.2 b) F-025 proposed. Settles D-04
M-H Consequence bands, defined with the operator — "cannot be done by an agent alone" Q-04 DISCHARGED 2026-09-22 — consequence-bands.md, authored live with the operator. Still PROVISIONAL

Sequencing, from 01 § 12: M-A, M-B and M-C are independent and may run concurrently. M-D must not start until its method problem is settled. M-E depends on M-A. M-G depends on M-H, which is discharged. M-F is independent, and is executed.

Two things this table makes visible that the register previously could not. M-F exists and is the measurement home for the BDD-duplication claim that never became a finding — recorded at DISAGREEMENTS.md D-08 as a challenge with nothing to land on. And M-H is already discharged, so the program is 7/8 outstanding rather than 8/8; nothing in the register said so. (Amended 2026-09-23: M-F has since been executed — piece 03 — so the program now stands at 6/8 outstanding.)

Reading posture

These are dated records, in the sense docs/governance/ already uses (see config-derivation-census-2026-09-20.md). A value written in the prose is ILLUSTRATIVE, never authoritative (AGENTS.md § Governance doctrine surfaces). Each piece pins the commit it was measured against and ships a re-runnable script beside it; the script carries no literals from its authoring date, so re-running it reports whatever tree it is given rather than confirming a transcribed figure.

Findings are classified KEEP / REFINE / ADD / REMOVE, and nothing is classified merely because a standard contains it.

The numbered pieces are frozen at their date; the canonical register is not. FINDINGS.md, DISAGREEMENTS.md and OPEN-QUESTIONS.md are living documents amended in place with dated notes — the shape advisory-rules-audit.md already uses. A piece is corrected only by a later piece saying so.

These records are analysis, not canon. Nothing here binds until it is carried into a rule, an ADR, or the corpus by the ordinary route. Where a piece disagrees with an earlier piece, the later one says so explicitly rather than silently superseding it.

Pieces — raw investigation record

Historical. Subordinate to FINDINGS.md; read it first.

# Piece Subject Measured at
01 Engineering-method assessment Whole-system pass: the engineering ontology, persistent versus transient knowledge, traceability, the OBPI as a work package, agent entry and exit cost 6a0e5241e
02 Requirements engineering versus release management The identifier seam: decision identity, requirement identity and release identity sharing one semver namespace; baselines as the missing bridge object be663409a
03 What Gate 4 actually adds over Gate 2 Measurement item M-F: behave and unittest run under identical coverage instrumentation, the behave-only line and REQ residue, and what retiring Gate 4 would actually cost 43d63da8d

Standards corpus

Twenty-four standards, read at clause level. Not summarized — used selectively to answer concrete questions raised by the repository.

Area Standards
Lifecycle and systems context ISO/IEC/IEEE 12207:2026, 15288:2023, 24748-1:2024, 24748-2:2024, 24748-3:2020, 24748-6:2023, 24748-10:2026, 24765:2017
Requirements and engineering information ISO/IEC/IEEE 29148:2018, 15289:2019
Architecture ISO/IEC/IEEE 42010:2022, 42020:2019, 42030:2019
Verification and assurance IEEE 1012-2024, IEEE 730-2026, ISO/IEC/IEEE 15026-2:2022
Testing ISO/IEC/IEEE 29119-1:2022, 29119-2:2021, 29119-3:2021, 29119-4:2021
Engineering management ISO/IEC/IEEE 15939:2017, 16085:2021, 16326:2019
Delivery and operations ISO/IEC/IEEE 32675:2022

The standards texts themselves are not vendored into this repository; they are licensed documents. Pieces cite standard, clause and a paraphrase, and quote only where exact wording carries the argument.

Companion reference, outside the standards corpus. DOT/FAA/AR-08/32, Requirements Engineering Management Handbook (Lempia and Miller, FAA, 2009), is read alongside the corpus during review and alignment. It is a public U.S. Government report, so it is vendored at references/FAA-AR-08-32.pdf. Its provenance and checksum are in references/. The F-032 and F-033 guard applies to it exactly as it applies to the standards.

The three permissions that shape the whole series

Read these before proposing that gzkit adopt anything, because they bound what adoption can even mean:

  • Information, not documents. ISO/IEC/IEEE 15289:2019 § 5.1 (shall) holds that information items conform when unpublished but available in a repository, divided across documents, or combined into one; § 3.1.11 defines "include" as having the information or a reference to it. ISO/IEC/IEEE 29148:2018 Clause 7 and 29119-3:2021 § 4.1.1 say the same for requirements and test information. A ledger with a derived view conforms exactly as a document set does.
  • A mandated process is not a mandated artifact. 15289 § 8.2: a definition "does not in itself indicate that a specific information item is produced", and clauses requiring planning "do not necessarily mean that a documented plan is produced."
  • Fewer processes, honestly discharged, beats more processes diluted. 12207:2026 § 4.3 prefers full conformance to a smaller declared process set over tailored conformance to a larger one. ISO/IEC/IEEE 32675:2022 § 6.3.1.3 b) 4) goes further and requires enabling change through "compact low-dependency scopes, low gates, low overhead."

Handoff — for an agent joining now

Read this section plus FINDINGS.md, and DISAGREEMENTS.md, OPEN-QUESTIONS.md and consequence-bands.md as needed. That is enough; you do not need to replay Phase 1 or Phase 2. This set is exactly Agent 2's Act 1 reading list — if you are Agent 2, stop at its boundary and do not open the numbered pieces or raw/.

What is confirmed. Two rows: F-032 (most of the standards' machinery should not be adopted here) and F-033 (conformance is dischargeable by reference). Both carry the challenge table's only affirmative verdicts. Fifteen rows are QUALIFIED — survived in narrowed form, with the narrowing on the row — and fourteen are OPEN, which under Q-14 means stated with evidence and never reached by the review: settled, and not promotable.

What is disputed. Two findings, both load-bearing: F-006 (does a persistent system model exist, or only fragments) and F-021 (can the system retire what it detects). Eight entries are recorded at D-01 … D-08 in DISAGREEMENTS.md, three of them A0 CONCEDES and three MISSING EVIDENCE. D-08 records a challenge with no finding to land on — evidence that the seeding of FINDINGS.md from the pieces was not exhaustive.

What is decided. Operator questions and their rulings are recorded in OPEN-QUESTIONS.md; Q-09 is deferred. Q-15 adopts the successive-change evaluation criterion for later design, pilot, and independent evaluation. Carry that criterion into Phases 4–6 when authorised; its adoption alone did not open those phases. Q-16 records FDAU as the investigation's origin and retains the separation hypothesis. Carry its history forward alongside Q-15. Q-17 now records the accepted grounding direction and bounded design request; read the anchor and candidate as the current design account. The session record preserves its reasoning and wider alternatives; read it when continuing this design so the latest recommendation does not erase the decision space.

What remains unmeasured. The proposed M-A … M-E and M-G measurements and the successive-change pilot. M-H is discharged, with its scores still provisional; M-F is executed at piece 03 and settles less than its headline suggests — it bounds Gate 4's line redundancy and leaves the assertion-level question D-08 raises untouched. Operator rulings do not establish empirical findings.

What must not be assumed:

  • Do not assume the central hypothesis is settled. That persistent knowledge should leave the work package (F-001) is the investigation's leading finding and is QUALIFIED, not established.
  • Do not assume long OBPI duration has a known cause. F-025 lists eleven live candidate explanations and attributes duration to none of them. Converging early is the most likely way this investigation goes wrong.
  • Do not assume a finding implies its remedy. Findings stop at what is; remedies are Phase 4.
  • Do not assume a standard's presence is an argument. F-032 and F-033: most of this corpus should not be adopted, and conformance is dischargeable by reference.
  • Do not assume the numbers are current. Every figure is a dated observation. Re-run the command or the script.
  • F-### ids are ratified (Q-11, ruled 2026-09-22, confirmed 2026-09-23). Allocated in order, never reused, stable across phases. Cite them freely.

Currently prohibited:

  • Exceeding Q-17's bounded design authorization, or treating it as permission to pilot, migrate, book an ADR, or execute an OBPI.
  • Promoting any finding to CONFIRMED that the Phase 2 review did not reach. The review is read and reconciled; a row it never challenged is still unchallenged.
  • Creating an ADR, OBPI, REQ or TASK from anything in this directory.
  • Treating a design candidate as a decision, or this register as doctrine.
  • Editing the numbered pieces other than by a dated amendment.
  • Agent 2 reading the numbered pieces or raw/ before its Act 1 cold read is recorded. The reading is worthless once anchored, and it cannot be retaken.
  • Retiring or replacing the five-gate vocabulary. Standing operator constraint, 2026-09-22: "do not abandon the five gates without a discussion with me." This binds Phase 4 designs too, including as a side effect.
  • Re-proposing the 16085 consequence threshold as a replacement for the IRON LAW. Ruled against at Q-05, as weaker rather than merely different.
  • Promoting ADR-pool.feature-adr-semver-discipline before the kind guard lands (Q-09).

Permitted design work: develop and review the bounded candidate under Q-17, with the full triad. The accepted grounding is not an instruction to restart the investigation, discontinue its other tracks, or implement the candidate. The existing measurement program M-A … M-H continues on its own terms (M-H discharged, M-F executed, six outstanding); the two register questions about D-08 and the independently observed D2 rows remain separate.

Relationship to gzkit's own engineering artifacts

One-directional. Findings are inputs; nothing in this directory is an output of the governance pipeline, and nothing here creates one.

  • No ADR, OBPI brief, REQ or TASK is created by this investigation. A Phase 4 decision leaves by the ordinary route — gz-design → ADR → OBPI — and only the operator initiates that work (AGENTS.md § OBPI Acceptance Protocol).
  • A defect noticed here routes by AGENTS.md § Defect-fix routing like any other. Being named in a finding neither authorises nor blocks its repair.
  • Nothing here is Layer 1 canon or Layer 2 ledger. These are analysis documents; .gzkit/ledger.jsonl remains the system of record (state-doctrine.md).
  • Findings are not requirements. They carry no proof channel, no @covers binding and no gate.

Amendments

  • 2026-09-25 — Design-session retention expanded after operator review. Added the dated deliberation record and linked the expanded candidate inventory. Preserved recovery timing and cost, distinct ownership models, stronger mADR coupling, the earlier release-first proposal, SWP/CNS, and the attachment's vocabulary and boundary hypotheses. No additional adoption or execution authorization is inferred.

  • 2026-09-25 — Grounding pivot and bounded Phase 4 design (Q-17). Added the accepted direction and designated candidate; updated current phase, joining-agent, and authorization surfaces. Preserved continuation of IEEE, test investigation, and Magna Carta release planning. No finding, runtime rule, campaign sequence, pilot, or migration changed.

  • 2026-09-23 — M-F executed; piece 03 added. The first measurement item run by an agent alone (M-H was run with the operator). Both suites instrumented identically, subprocess capture enabled on both sides so the ten subprocess-driving step files were not scored as reaching nothing. Headline: the unit suite already covers 98.9% of every line behave reaches, and 47 of 74 feature files add no reach at all. Five of 01 § 8.3's six transcribed figures verify; the behave-only REQ count is 55 rather than ~54, and "restates the unit file name-for-name" is literally false and should not be re-cited. The result is deliberately one-sided and says so: line coverage cannot see a different assertion over the same line, which is exactly Phase 2's challenge at D-08, so D-08 keeps MISSING EVIDENCE and no finding was authored. Two sub-results stand on their own terms — the 35 @wip scenarios prove nothing not already @covers-ed, and subagent_pipeline.feature adds neither a line nor a REQ tag. M-F's state moved to EXECUTED in § The measurement program; the program is 6/8 outstanding.

  • 2026-09-23 — FDAU origin and separation hypothesis retained (Q-16). Recorded the operator's account of why the IEEE investigation began, pinned the comparison sources, and linked explicit consideration through the phase and handoff paths. Preserved the useful work-package idea and the proposed separation without adopting a schema, changing findings, or opening Phase 4.
  • 2026-09-23 — Dex Horthy deposit reconciled against the register. § Success across successive changes gains Where the testimony lands in the register: each excerpt mapped to the F-### or M- item it bears on, with the bears-on / evidence-for distinction stated per row. The deposit's own boundary claim was verified mechanically — the status distribution is unchanged across 35 rows — rather than accepted on its word. Two notes recorded: the criterion independently agrees with F-032, the register's only unqualified CONFIRM; and the excerpts cite a source not present in the repository, so they inherit F-022's retrievability defect and are attributed rather than verifiable. No finding status changed and no row gained evidence.
  • 2026-09-23 — evaluation ruling propagated after operator check. The operator asked "did you update where needed?". The first update had recorded the decision here but omitted OPEN-QUESTIONS.md. Added Q-15 there and linked it from the phase table, criterion, and handoff reading path. Removed stale current-state prose that described ruled questions as unknown and the completed reconciliation as awaiting execution. Historical amendments remain dated records. No finding or phase authorisation changed.
  • 2026-09-23 — successive-change evaluation adopted by g0. Operator: "okay, let's do that then", accepting the proposed success criterion and pilot approach following the Dex Horthy transcript discussion. Added § Success across successive changes and retained timestamped source excerpts in raw/. The lessons are evaluation questions, not new findings. Phase 4–9 authorisation, Phase 6 independence, the five gates, and finding statuses are unchanged.
  • 2026-09-23 — Phase 3 recorded complete (Q-14). The phase table, § Current gate, § Next permitted step and § Handoff updated. No finding changed status and nothing was promoted: the fourteen OPEN rows stay OPEN, and the ruling is about what OPEN means, not about what those rows assert. The CONFIRMED bar is unchanged and still requires that the review reached a row. The Phase 4 bar is restated rather than relaxed, in § Current gate and § Next permitted step, because a phase closing is the most likely moment for the next one to be read as open. It is not.
  • 2026-09-23 — three cold-read gaps closed. § Roles corrected: it said Astra deposits into raw/, against Q-13's ruling that the Phase 2 report stays at this directory's root. § The measurement program added — M-A … M-H summarised from 01 § 12, which the five canonical files previously never reached, leaving eight findings routing to items a reader could not evaluate (cold-read gap 2). Two facts surfaced by writing it: M-F exists and is the home of D-08, and M-H is already discharged by consequence-bands.md. The evidence directory is added to the file table as tooling rather than narrative (cold-read gap 8): a re-derivation script holds no conclusions, so reading it cannot anchor a future Act 1 the way a report would. That last point narrows a binding reading list and the operator may overturn it — § Agent 2 sequencing constraint still bars the numbered pieces and raw/, and only the script is at issue.
  • 2026-09-23 — Agent 2's Act 1 cold read recorded at act1-cold-read-2026-09-23.md, historical tier, and added to the file table; § Agent 2 sequencing constraint updated to say Act 1 was taken and cannot be retaken. Act 2 is unchanged and still does not exist until Phase 4 is authorised. The reading was taken by an isolated agent holding no prior context, dispatched because the main session was already anchored by the handoff chain — the caveat is recorded in the record itself. Written up because its verdict had existed only in session memory while its consequences were already landed in the register: F-007 happening to this investigation, caught before the session ended.
  • 2026-09-23 — § Terms this series cites as binding added. Agent 2's Act 1 cold read could not evaluate Q-05 because IRON LAW is cited four times across this register and stated nowhere in the five canonical files; the same held for Architectural Boundaries 1 and 2, cited as binding at Q-03, Q-09 and F-006. Both are now pointed at their authority. The cause is worth keeping: the rule in force in AGENTS.md does not carry the name IRON LAW, so a reader searching for the name finds commentary and a reader reading AGENTS.md finds the rule without the name — the operator's own standing caution that "a search is not a read" and that doctrine is routinely stated as a flag value or a path rather than as the prose being searched for. No rule was restated as authority here; the glosses are marked ILLUSTRATIVE.
  • 2026-09-23 — Q-11 and Q-12 recorded. § What must not be assumed had told readers not to assume F-### was ratified; it had been ratified on 2026-09-22, and the line is replaced with the ruling. Q-12 is confirmed: Phase 4 design candidates will live in design-candidates.md in this directory, tiered below FINDINGS.md, created when Phase 4 is authorised — so § Currently prohibited's bar on beginning Phase 4 design is unaffected, and no empty container is created ahead of it.
  • 2026-09-22 — Phase 3 reconciliation pass recorded. § Current gate, § Next permitted step, § Handoff (what is confirmed, what is disputed) and the CONFIRMED prohibition updated to the post-reconciliation state. The prohibition was narrowed, not lifted: it now forbids promoting a row the review never reached, which is the condition that actually bears after the report has been read. This also settles, in one direction, the contradiction flagged below — this file and DISAGREEMENTS.md previously stated two different gate conditions on promotion to CONFIRMED. The operator may overturn the reading.
  • 2026-09-22 — Act 1 cold-read repair pass (mechanical only). § Current gate's orphaned "Until then" — editing residue with no antecedent, inside a binding paragraph — reworded to name the reconciliation pass it was describing. No prohibition was added, removed or relaxed. This section is new; changes to this file were previously unrecorded. Unresolved, left for the operator: § Handoff states the Phase 2 report "has not been read", against this file's own phase table and DISAGREEMENTS.md § Status; § Currently prohibited gates promotion to CONFIRMED on the report being read while DISAGREEMENTS.md gates it on that file being populated; § What must not be assumed records F-### as unratified against Q-11's ruling; § Roles says Astra deposits into raw/ against Q-13's ruling that the report stays at the directory root; and IRON LAW and Architectural Boundaries 1 and 2 are cited as binding here without being stated, quoted or pointed at anywhere in the five canonical files.