Skip to content

Disagreements — Agent 0 and Agent 1

Meaningful disagreement is preserved here, not resolved into consensus. A disagreement that is merely unresolved is more useful to a future investigator than an agreement that was manufactured.

Agent 1 is not epistemically subordinate to Agent 0. Agent 0 leads reconciliation; that is a procedural role, not the authority to win an argument. Where Agent 1's challenge is stronger, the corresponding row in FINDINGS.md moves to QUALIFIED or REJECTED and says so.

Status: reconciled 2026-09-22 — eight entries

Astra's Phase 2 report is at gzkit-engineering-assessment-adversarial-review.md. The full reconciliation pass ran 2026-09-22, after the operator lifted the deferral. All 25 rows of its § 2 challenge table were mapped onto FINDINGS.md and dispositioned.

Counted directly from § 2: 25 data rows — 11 CONFIRM WITH QUALIFICATION, 9 REJECT, 3 DOWNGRADE TO HYPOTHESIS, 1 CONFIRM, 1 NEEDS MORE EVIDENCE.

Where the pass landed, by status: 2 CONFIRMED (F-032, F-033), 15 QUALIFIED, 2 DISPUTED (F-006, F-021), 14 OPEN, 2 REJECTED (F-034, F-035, both pre-existing) — 35 rows in total. Nineteen rows received a Phase 2 challenge block and seventeen statuses moved; F-001 and F-018 were already QUALIFIED before the pass and carry their challenge without a status change.

Fourteen rows remain OPEN, and that is a deliberate classification, not an oversight. (Ruled a settled disposition by the operator at Q-14 on 2026-09-23, after the reasoning below was written.) Astra reviewed pieces 01 and 02, not this register, and states its own scope limit: "This is not an exhaustive conformity audit", and the report's subjective classifications, receipt population and historical gate events "were not independently reclassified." A row Astra did not reach was not exposed to challenge, so it did not survive one. Promoting such a row to CONFIRMED would make the status word mean what the seeding note already refuses to let it mean. This pass could not meet Phase 3's stop condition on its own. The operator closed it at Q-14 by ruling OPEN settled for an unchallenged row, rather than by commissioning a second adversarial pass — which would have required a fourth role the binding composition model does not have.

Two REJECTED verdicts were not adopted, and the reasons are recorded at D-01 and D-05 rather than resolved by the lead's procedural role.

Three rows of F-020's twenty-two-point disagreement register are themselves rejected on counterevidence and are no longer live disagreements; they are recorded on that finding.

One challenge row has no F-### to land on. "§3/§8: BDD gate largely duplicates unit tests; retain only unique REQs" — verdict NEEDS MORE EVIDENCE — challenges a piece 01 claim that was never carried into this register. Recorded at D-08. A challenge with no row is a gap in the register, not in the challenge. (Amended 2026-09-23: the claim does have a measurement home — M-F — summarised in README.md § The measurement program. It has no finding, which is the gap.)

Entries

D-01 — Does a persistent system model exist, or only fragments of one?

  • Findings touched: F-006, and by inference F-013
  • Agent 0 position: Six intake surfaces each hold part of the project's durable knowledge, and none holds the model itself, so no surface is a backlog against a model. Risk and research-question have no home at all
  • Agent 1 challenge: "Absence claim not sustained." State/trust doctrine, CLI specification, architectural identity, source ontology and the OKF knowledge bundle exist [E7,E9,E10]; ADR consequences and research records hold risk and research content [E9,E10]. The claim "Conflates full 42010 AD conformity with possession of useful system knowledge", and "A concept need not have its own object"
  • Evidence for A0: 01 § 2 object inventory, 01 § 4.4, 01 § 8.2; the three-session rediscovery loop recorded in the repository
  • Evidence for A1: E7, E9, E10 — named artifacts, each verifiable in the tree
  • Disposition: UNRESOLVED
  • To resolve: The two positions use model differently: A0 means a queryable object with coverage over concerns, A1 means possession of system knowledge in retrievable form. The observation that settles it is whether any existing surface can answer "which decisions bear on this surface" without reconstruction — M-C and M-E between them ask it. Neither agent can settle it by argument, and the rediscovery loop is evidence for A0 that A1's inventory does not address

D-02 — Can a hierarchical REQ identifier outlive its work package?

  • Findings touched: F-002
  • Agent 0 position: REQ-<adr-semver>-<obpi-NN>-<req-NN> encodes parentage in identity, so the requirement cannot outlive the package that introduced it
  • Agent 1 challenge: "Logical assertion, not observed property of the grammar." The parser accepts an ID independently; storage lookup and deletion govern discoverability [E6]. 29148 permits relational identity [S2]
  • Evidence for A0: src/gzkit/triangle.py:24-31, the consolidated grammar
  • Evidence for A1: E6 — parser behaviour, independent of storage
  • Disposition: A0 CONCEDES
  • To resolve: Already resolved. F-002's own Interpretation conceded this before the challenge arrived — "The identifier grammar is sound and globally unique. What is wrong is ownership, not syntax." The Observation overstated what the Interpretation claimed, and the row is narrowed rather than defended. Recorded because a future investigator reading the original wording would re-derive the rejected form

D-03 — What was destroyed when demoted briefs were deleted?

  • Findings touched: F-003, and F-001 which depends on it
  • Agent 0 position: A large fraction of all REQ acceptance criteria and FAIL-CLOSED constraints ever authored has been deleted by gz adr demote, demonstrating loss of enduring requirements
  • Agent 1 challenge: "Deletion count is real; durability and causal attribution are unclassified." 364 of 384 deletion events cluster in two deliberate demotion campaigns, largely unstarted work [E1]; the 16326 inference is invalid. Alternative: backlog retirement, with a smaller consequential retention problem underneath
  • Evidence for A0: adr_demote.py:475; git log --diff-filter=D over brief paths; obpi_lifecycle.py:256-260 ("A hollow exit 0"); live behave tags pointing at deleted briefs
  • Evidence for A1: E1 — the clustering of deletion events into two campaigns
  • Disposition: MISSING EVIDENCE
  • To resolve: Neither side has classified the deleted population. The question is what share of the deleted REQs and constraints were durable engineering knowledge rather than retired backlog — M-B exactly. Recorded as MISSING EVIDENCE and not as UNRESOLVED, because the argument is premature rather than the evidence divided. The operator's Q-01 ruling that the delete is collateral rather than designed is unaffected either way, and the dangling behave tags are a defect on their own terms

D-04 — Does knowledge placement explain work-package duration?

  • Findings touched: F-001, F-025
  • Agent 0 position: Persistent knowledge carried inside transient work packages explains long OBPI duration — the investigation's central causal claim, asserted in piece 01 §§ 4, 6 and 8
  • Agent 1 challenge: "No causal test; created-to-completed mixes queue and execution." One OBPI open more than 33 days completed 3.25 hours after its first lock [E2]. Alternatives: authorization and sequence waiting, scope coupling, churn
  • Evidence for A0: 01 § 4.4, 01 § 6; the duration distribution in 01 § 6
  • Evidence for A1: E2 — lock-to-completion elapsed time against created-to-completed elapsed time
  • Disposition: MISSING EVIDENCE
  • To resolve: M-G, baselining duration as repeated measures against the eleven candidate causes F-025 lists. F-025 already refuses to attribute duration to any cause, and its candidate list already contains Astra's alternatives — so on this point the register and the challenge agree, and the disagreement is with piece 01's §§ 4, 6 and 8, which the operator had independently flagged for using unsettled findings as settled inputs. A0 concedes the causal claim; the placement observation underneath it is not conceded and is not challenged

D-05 — Can the system retire what it detects?

  • Findings touched: F-021
  • Agent 0 position: The error-detecting loop works and is honest; what is missing is a mechanism to retire what it finds, so findings accumulate alongside the defects they describe. The grandfather ratchets are declared undeveloped arguments in 15026-2 § 3.1.7 terms
  • Agent 1 challenge: "Historical audit and backlog sizes do not establish current failure", and "A baseline is not automatically an undeveloped assurance argument." July NC defects have subsequent repair commits; withdrawal, repudiation and demotion were exercised [E1,E5]. On the companion row: "Too absolute" — the rubric reads references from content and 36 defect-resolution records exist [E3]
  • Evidence for A0: enforcement-claim-nc-audit-2026-07-18.md; evidence-record-contract.md; data/mechanical_witness_grandfather.json; Ledger.get_post_validation_failed_gates
  • Evidence for A1: E1, E3, E5 — repair commits following the audit, retirement verbs exercised, 36 defect-resolution records
  • Disposition: UNRESOLVED
  • To resolve: The positions talk past each other and both are supported: A1 shows individual defects were repaired and retirement verbs exist; A0 claims no mechanism links a finding to its retirement, so the standing inventory cannot be reduced. Astra's own alternative concedes the narrower form — "No adequate linked current-state reduction, despite useful archival signals." The settling observation is whether any query returns the set of enforcement-claim non-conformances still live today. M-D asks it, and carries its own method problem: the prior audit disqualified itself as "a stochastic surface auditing a stochastic surface"

D-06 — Do gzkit's gates decide anything?

  • Findings touched: F-019
  • Agent 0 position: The large majority of gate_checked events carry no observation of what was verified; _run_gate_5() prints and returns True; these are checks, not decision gates, in 24748-1 § 4.3.2 terms
  • Agent 1 challenge: "Good historical payload concern; false generalization." The claim "Confuses a check, a gate condition, and the authorization consuming it." Closeout blocks failures; completion emits human-attestation receipts [E4]. Alternative: a deprecated interface plus dispersed evidence, not universally absent gating
  • Evidence for A0: src/gzkit/commands/gates.py:153-160, :261-263; governance/deprecations.py:41
  • Evidence for A1: E4 — completion_review and closeout behaviour
  • Disposition: A0 CONCEDES
  • To resolve: Already resolved, in the challenge's favour on the general claim and A0's on the specific one. The surviving finding is about the deprecated gz gates interface and the historical payloads, not about the system's ability to gate. The operator's Q-07 ruling — re-point the covenant at gz closeout and gz obpi complete, keeping the five-gate vocabulary — was reached independently and lands in the same place. Recorded because the original wording is the kind a later reader would cite as evidence that gzkit gates nothing

D-07 — Is the anti-tautology stack ahead of the testing standard?

  • Findings touched: F-028
  • Agent 0 position: 29119 has no vocabulary for RED-witness or mutation-witness classification — "test oracle" appears zero times in Parts 2–4 — and a none classification records a test that demonstrably cannot fail
  • Agent 1 challenge: "Mechanisms useful; superiority and universal inference unsupported." Part 1 explicitly defines oracle and oracle problem [S8]. And "A passing selected baseline does not establish incapacity to fail on all relevant faults"
  • Evidence for A0: red_witness.py, mutation_witness.py (GHI #963); the zero-occurrence count over Parts 2–4
  • Evidence for A1: S8 — ISO/IEC/IEEE 29119-1's own definitions
  • Disposition: A0 CONCEDES
  • To resolve: Already resolved. The zero-occurrence count over Parts 2–4 stands as stated and the mechanisms keep their KEEP class, but the corpus does have the vocabulary and the finding searched the wrong parts of it. A none witness is a local adequacy result over a selected baseline, not a proof of incapacity. This is the clearest instance in the challenge table of a standards claim made from a partial read of the corpus, and it is a caution for Phase 4

D-08 — BDD duplication: a challenge with no finding to land on

  • Findings touched: none — this is the gap
  • Agent 0 position: Piece 01 held that the BDD gate largely duplicates unit tests and that only scenarios covering unique REQs should be retained. This claim was never carried into FINDINGS.md
  • Agent 1 challenge: NEEDS MORE EVIDENCE. "ID/function overlap is insufficient"; the 1012 note is "misapplied to suite architecture" [S6]; "Same lines/REQs can be exercised under different conditions and assertions." Alternative: some redundant tests amid useful boundary and acceptance checks
  • Evidence for A0: 01 § 3, 01 § 8 — not re-derived by this pass
  • Evidence for A1: S6 — IEEE 1012's note read in its own scope
  • Disposition: MISSING EVIDENCE — unchanged by M-F's execution; see the 2026-09-23 note below
  • To resolve: M-F is this claim's measurement home, and it already exists — "Measure the real duplication between Gate 4 and Gate 2 before proposing removal", independent and cheap, with a method (run behave under coverage instrumentation, compare against the unit suite, verify the behave-only REQ count) and a note that @wip scenarios which never execute must be counted separately. So the claim is not homeless in the investigation; it is homeless in the register. Both Astra's verdict and M-F say the same thing in different vocabularies: measure before proposing removal. What remains for the operator is whether a finding should be authored so the claim carries a status, or whether M-F alone is a sufficient home for a REMOVE-class proposal that no finding asserts. Either way the seeding of FINDINGS.md from pieces 01 and 02 was not exhaustive, which is worth knowing before the register is treated as complete

Amended 2026-09-23 — M-F executed; piece 03. The disposition does not move, and the reason is the challenge itself. Astra's objection was "Same lines/REQs can be exercised under different conditions and assertions", and line coverage is structurally unable to answer it: a scenario asserting different behaviour over the same lines is indistinguishable from a duplicate. M-F measured the coverage residue — 288 behave-only statements, 1.14% of behave's own reach, 47 of 74 feature files adding none — which bounds the claim from one side only. A0's position is now supported for line reach and still unsupported for assertion content, which is not the same as supported. Two narrower results stand unconditionally, because they do not depend on what a scenario asserts: the 35 @wip scenarios execute nothing and every one of their 25 REQ tags already has a @covers test, and subagent_pipeline.feature contributes neither a behave-only statement nor a single REQ tag. The open operator question is unchanged — whether a finding should be authored so the claim carries a status — and M-F having run makes it more answerable, not answered.

Phase 3's stop condition was met on 2026-09-23 — every row carries a settled status, and the disagreements are recorded. Standing conditions:

  • D-01 and D-05 are live disagreements on load-bearing findings (F-006, F-021) and must not be read as settled in either direction. A disagreement recorded is not a disagreement resolved, and this file exists to keep that distinction;
  • the fourteen OPEN rows are settled under Q-14, not confirmed — nothing challenged them, and they may not be cited as having survived challenge;
  • Bounded Phase 4 design is now authorized at OPEN-QUESTIONS.md Q-17 (2026-09-25). Nothing in this disagreement register supplies that authority or settles the disagreements; the operator's request does. Pilot and migration remain separate.

Entry schema

Each disagreement is recorded as:

Text Only
### D-NN — <the issue, stated neutrally>

- Findings touched:   F-0xx, F-0yy
- Agent 0 position:   what Phase 1 concluded, and on what basis
- Agent 1 challenge:  what Phase 2 asserts against it, in Agent 1's terms
- Evidence for A0:    repository references
- Evidence for A1:    repository references
- Disposition:        A0 CONCEDES | A1 CONCEDES | UNRESOLVED | MISSING EVIDENCE
- To resolve:         the specific observation that would settle it, and who can make it

MISSING EVIDENCE is distinct from UNRESOLVED, and the distinction is the point of this file. UNRESOLVED means both positions are supported and the evidence does not choose between them — a real disagreement. MISSING EVIDENCE means neither is yet supported and the argument is premature. Recording the second as the first would manufacture a dispute where there is only an unmeasured question; those belong in OPEN-QUESTIONS.md or the measurement program.

Where a challenge lands, the disposition is recorded here and the status is changed on the finding itself. The two must not drift.

Amendments

  • 2026-09-25 — Authorization pointer updated for Q-17. The operator requested a grounding design pivot. Updated the current phase boundary only; no disagreement or finding disposition changed.

  • 2026-09-23 — M-F executed; D-08 annotated, disposition unmoved. The measurement D-08 was waiting on has been run (piece 03). MISSING EVIDENCE stands: the evidence produced is line coverage, and Astra's challenge is about assertions over shared lines, which coverage cannot see. Recording it as resolved would convert a bounded result into the unbounded one the challenge contested. No other entry changed and no other disposition moved.

  • 2026-09-23 — stop condition met (Q-14). § Status updated and the closing conditions rewritten. No entry changed and no disposition moved. D-01 and D-05 remain UNRESOLVED and are restated as standing conditions, because the moment a phase closes is the moment a live disagreement is most likely to be read as settled.

  • 2026-09-23 — D-08 given its measurement home. Tracing M-A … M-H out of 01 § 12 showed that M-F exists and is exactly this claim's measurement item. The claim is homeless in the register, not in the investigation, and D-08's To resolve is narrowed accordingly: the open question is whether a finding should be authored so the claim carries a status, or whether M-F alone suffices for a REMOVE-class proposal no finding asserts.
  • 2026-09-22 — Phase 3 reconciliation pass run. The operator lifted the deferral; all 25 rows of Astra's § 2 challenge table were mapped onto FINDINGS.md and dispositioned. Nineteen rows received a Phase 2 challenge block; seventeen statuses moved — 2 to CONFIRMED, 13 to QUALIFIED, 2 to DISPUTED, with F-001 and F-018 already QUALIFIED; 14 rows were left OPEN because Astra's non-exhaustive review never reached them, and silence is not survival. Eight disagreement entries recorded: D-01 and D-05 UNRESOLVED, D-02, D-06 and D-07 A0 CONCEDES, D-03, D-04 and D-08 MISSING EVIDENCE. Phase 3's stop condition is not met by this pass and cannot be — see § Status. Two items for the operator: D-08 shows the seeding of FINDINGS.md from the pieces was not exhaustive; and F-014's qualification is Astra reaching, independently, the Q-02 narrowing that is still awaiting a ruling.
  • 2026-09-22 — Act 1 cold-read repair pass; challenge tally corrected. The un-reconciled tally read "roughly 26 rows, 8 REJECT". Counted directly from § 2 of the report: 25 data rows — 11 CONFIRM WITH QUALIFICATION, 9 REJECT, 3 DOWNGRADE TO HYPOTHESIS, 1 CONFIRM, 1 NEEDS MORE EVIDENCE. The contested share is 9/25, not 8/26, and the prior enumeration summed to 24 against a stated 26. The "Created empty" entry below records the Phase 2 report as awaited; it is left standing as the dated record it is, and the report's arrival is stated in § Status above. Unresolved: the nine REJECT rows are still not mapped to F-### ids — that mapping is what the reconciliation pass produces, and until it exists a reader can tell that roughly a third of the register is contested but not which third. Separately, this file gates promotion to CONFIRMED on "until this file is populated" while README.md § Currently prohibited gates it on the Phase 2 report "being read"; two conditions on one act, for the operator to settle.
  • 2026-09-22 — Created empty (Phase 3). Schema fixed; no entries. Awaiting the Phase 2 report.