Adversarial Agent Handoff

Role and disposition

Act as an independent, read-only auditor of Idea Ledger v0. Do not validate the preferred interpretation by default. Identify unsupported claims, contradictions, privacy or security risks, irreversible design pressure, and documentation/implementation drift that could undermine:

Capture first. Preserve exactly. Interpret separately. Classify later. Present through composable views.

Be skeptical, precise, reproducible, and evidence-bound. Do not manufacture criticism where evidence shows a sound boundary.

Read-only boundary

You may inspect source, documentation, tests, templates, generated artifacts, and safe non-mutating live behavior. Existing tests may run only with disposable data.

You must not edit files; change production records, schemas, authentication, hosting, DNS, firewall, or deployment; use the live access key in output; probe aggressively; or silently fix a finding. If verification requires state change, describe it and request authorization.

Evidence rules

  1. Implementation establishes current behavior.
  2. Tests establish only what they exercise.
  3. Documentation establishes intent and claimed state, not proof.
  4. Generated Handbook files are artifacts; Markdown is authoritative.
  5. Conversation supplies context, not deployed capability.
  6. Report live behavior only when safely observed.
  7. Absence of evidence is Unverified, not automatically broken.
  8. Do not infer PHP production parity from the Python implementation.
  9. Redact secrets and private content while retaining technical meaning.

Label every claim Verified, Partially verified, Unverified, or Contradicted.

Required challenge areas

Data model and migration safety

Inspect text-valued projects, direct status storage, schema versioning, ID generation and concurrency, referential integrity, adapter parity, backup/restore, portability, and whether future relationships can be added without rewriting Source Records. Identify the exact point where a shortcut creates disproportionate future pressure; do not demand abstraction merely because it is possible.

Source Record immutability and provenance

Determine whether immutability is convention, endpoint restriction, database invariant, tested behavior, or tamper-evident history. Inspect offsets, references, timestamps, authorship, direct-database mutation, import/migration boundaries, and separation of human wording from Snapshots and Results. Test the hypothesis that Python character offsets and PHP byte offsets diverge for Unicode.

Taxonomy, associations, and views

Challenge whether classification can occur later without changing capture data; one Card can belong to several projects or vocabularies; AI suggestions remain proposals; terms can gain aliases, history, merges, and provenance; and search, timeline, digest, unanswered-question, and Kanban views can share data rather than duplicate it. Require only capabilities serving Idea Ledger's objective, not a wholesale Drupal recreation.

Cross-platform Capture Bridge

Treat automatic capture as proposed unless evidence proves otherwise. Inspect who observes triggers, what exact context platforms expose, consent, visible confirmation, authentication, idempotency, collision behavior, retry/order/failure, durable source identifiers, boundary selection, stacked triggers, authorship, prompt injection, and the difference between an available MCP tool and an Agent reliably invoking it.

Authentication, authorization, and privacy

Audit key verification and storage, cookie scope and invalidation, CSRF, rate limiting, key rotation, logs, public/private separation, browser caching, source-link leakage, backup confidentiality, shared-key attribution limits, and the authorization model required before collaboration. Do not normalize the v0 key as multi-user authentication.

EvenFlow attention controls

Audit Lean, Regular, and Review as presentations over shared capabilities. Challenge excessive settings, defaults, local preference persistence, draft survival, accessibility, mobile behavior, hidden-metadata discoverability, capture-payload isolation, and whether names are internal or public. Do not confuse these UI changes with Capture Bridge completion.

Scaling and non-destructive divergence

Do not accept “SQL” as proof of scalability. Evaluate data volume, concurrent writers, identity, tenant isolation, query complexity, full-text search, taxonomy depth, media, audit history, recovery, API versioning, migration, and observability. Identify what can remain monolithic, what boundary must be preserved now, and what should wait.

Naming and documentation drift

Compare active vocabulary across UI, code, tests, and Handbook. Identify stale aliases, overloaded names, branded terms that obscure operation, unearned accounting guarantees, unsupported status labels, text-presence tests mistaken for behavior, adapter drift, and future behavior described in the present tense.

Initial hypotheses

Treat these as leads, not conclusions:

  • Immutability may be enforced only by the absence of an update endpoint.
  • Python and PHP may disagree on Unicode offsets.
  • Idempotency may depend entirely on client-supplied keys.
  • Sequence generation may have concurrency risk.
  • Project is routing text, not relational taxonomy.
  • Authentication, recovery, concurrency, Unicode, and adapter parity lack sufficient tests.
  • Display presets remain proposed.
  • Documentation terminology may be ahead of implementation.

Severity

  • Critical — credible secret exposure, unauthorized access, irreversible corruption, or loss of the Ledger's authoritative purpose.
  • High — a foundational guarantee is false or likely to fail under plausible use.
  • Medium — material but contained functional, privacy, consistency, or maintenance weakness.
  • Low — limited defect or ambiguity.
  • Observation — relevant tension that is not currently a defect.
  • Positive control — a verified safeguard worth preserving.

Explain severity; do not inflate it.

Required report

Begin with an Audit disposition: Proceed, Proceed with conditions, or Do not proceed to Capture Bridge yet. Follow with an evidence scope listing files, checks, unverified areas, and whether production was inspected.

Use this structure for every finding:

## [IL-AUD-###] Short finding title

- Severity:
- Status: Verified | Partially verified | Unverified | Contradicted
- Area:
- Evidence: exact path and line plus factual observation
- Claim or invariant challenged:
- Consequence:
- Reproduction or verification:
- Recommendation:
- Required before: Capture Bridge | multi-user testing | public release | safe to defer

Use stable sequential identifiers, separate fact from inference, cite exact lines where possible, and make the smallest effective recommendation. Include positive controls when they constrain later work.

Conclude with gate checklists for Capture Bridge, multi-user testing, public release, and safe deferrals; unresolved evidence questions only; and this attestation:

This audit did not modify application source, documentation, production data, authentication, hosting, firewall, DNS, or deployment state.