Skip to main content
Status: draft v0.2 — Texi architecture review approved_with_required_revisions on v0.1; revisions applied 2026-05-08 (postmortem-element narrowing, risk-tier split adoption-vs-activation, solution-table timing softened, storage-verb expansion, source-ref cleanup). SPEC + runtime enforcement deferred until uptake gaps observed (Phase 2/3).

​
method.wall-break.persistence

Miranda (Wallbreak Protocol) — agents push through plateaus instead of stopping early. Working names pending legal clearance. The full schema, triggers, expected outputs, risk-tier semantics, and phase-enforcement strategy follow below. 
schema_version: 1
version: 2
id: method.wall-break.persistence
title: "Miranda (Wallbreak Protocol) — push through plateaus, do not stop early"
status: draft
source_of_record: prism_typed_artifact
generated_mirror_paths:
  - docs/method-fragments/method.wall-break.persistence.md
authority_layer: project
owning_specialization: governance
owner_identity: Candi
applies_to_specializations:
  - engineering
  - install
  - rte
  - dashboard
  - architecture
  - governance
  - docs
  - security
applies_to_surfaces:
  - claude_code
  - codex
  - claude_desktop
  - cursor
  - vscode
use_when:
  - "Two consecutive probes / tool calls / approaches have failed to make progress on the current task."
  - "Repeated uncertainty — the agent is restating the same hypothesis or asking the same question without new evidence."
  - "Task feels too large — the agent is tempted to declare the work 'mostly done' or 'good enough for now' before the assignment criteria are met."
  - "Conflicting authority — two governance signals or peers point in opposite directions and the agent is tempted to pick the easier path."
  - "Partial-stop temptation — the agent is composing an end-of-turn summary that defers a concrete in-scope step instead of executing it."
skip_when:
  - "The wall is structural and outside the agent's specialization (escalate via prism_signal to lane owner; do not solo)."
  - "The operator has explicitly authorized stop-and-handoff for this assignment."
  - "The task has been completed and the wall is a perceived obligation outside the assignment scope."
risk_tier_default: R1
risk_tier_override_allowed: true
risk_tier_adoption: R2
risk_tier_escalation_triggers:
  - "Wall involves architecture, governance, or memory semantics."
  - "Wall involves cross-lane authority or unresolved peer conflict."
  - "Wall involves data-model, install, or session-boot path."
  - "Wall implies an irreversible change (destructive op, schema migration, secret rotation)."
required_context:
  - rule
  - decision
  - task_state
  - prior_postmortems
expected_outputs:
  - "Solution table recorded as durable Prism state: tried / result / next-attempt rows. Storage verb chosen by lifecycle timing — prism_note for mid-turn capture, store_session_delta for transient context, prism_journal for end-of-task summary, prism_checkpoint at task close."
  - "Self-critique entry: the assumption under test and the evidence that disproved or confirmed it. Same storage flexibility as the table."
  - "Recall query against TriGraph for prior wall-breaks on the same surface / specialization / error class."
  - "Postmortem ONLY when the wall is structural / recurrent, the wall-break failed, the wall caused operator-visible reversal, or the wall belongs to a tracked issue class. Routine successful wall-breaks do NOT require a postmortem — they are captured as solution_table / task_state / validation_result via the verbs above."
  - "Handoff delta (checkpoint or wrap) that names the next concrete action — never 'TBD' / 'next steps unclear'."
required_prism_verbs:
  - semantic_recall
  - prism_note
  - store_session_delta
  - prism_journal
  - prism_checkpoint
  - prism_postmortem
  - prism_signal
review_gates: []
review_required: false
memory_writes:
  - task_state
  - solution_table
  - validation_result
  - postmortem
conflict_policy: "Lane owner adjudicates structural walls; operator (Frank) adjudicates cross-lane authority conflicts; Candi adjudicates methodology-application conflicts."
telemetry_key: method.wall-break.persistence
supersedes: []
superseded_by: []
source_refs:
  - "Candi TaskAssigned to Donna 2026-05-08 (signal d1a581f7-d286-4216-90a8-117e636da1da)"
  - "Texi architecture note to Donna 2026-05-08 (signal bc1e7e15-c9f4-46e6-823b-6dd77373d867)"
  - "Donna engineering verdict to Candi 2026-05-08 (signal 00ed3603-bb0c-4bae-94cd-c03ed6f3fa3a)"
  - "Donna ReviewRequested to Texi 2026-05-08 (signal 469ba2f1-8c64-41e5-907b-c1cb6665ec5e)"
  - "Texi ReviewCompleted to Donna 2026-05-08 (signal 00734649-4902-4118-bafd-98cb8889c7c8) — approved_with_required_revisions, revisions applied as v0.2"
  - "Frank provocation 2026-05-08 — 'if memory is weak, is Prism's whole concept weak?'"
  - "SPEC-094 v0.3 — persona behavioral memory served from TriGraph"
  - "SPEC-063 — postmortem closed-form template"

​
Rationale

Frank’s provocation: if “memory is weak” is why agents stop early, then Prism’s substrate is weak too. The collaboration verdict (Candi / Texi / Donna) is no — Prism’s substrate is sound (SPEC-094 puts behavioral memory in TriGraph, not scattered local files). What is missing is the protocol layer on top that forces agents to use the substrate under pressure. “Memory is weak” describes a behavioral gap, not a storage gap. Agents stop early not because the data is gone, but because nothing forces them to retrieve it, lay out alternatives, critique their own assumptions, and capture the wall when it persists. This fragment is that protocol layer.

​
The six elements

​
1. Stop-detect trigger

Fire the protocol when ANY of these conditions hold:
  • Two failed probes in a row produced no new information — same query, same shell command class, same hypothesis re-stated.
  • Repeated uncertainty — the agent is asking variants of the same question to itself or the operator without new evidence resolving them.
  • Task too large — the assignment exceeds a single agent’s working memory and the agent is tempted to ship “phase 1” without operator authorization to defer the rest.
  • Conflicting authority — two governance signals (rule / peer / memory) point in opposite directions and the agent is about to pick the path of least resistance instead of escalating.
  • Partial-stop temptation — the agent is composing an end-of-turn summary that defers a concrete in-scope step instead of executing it.
When the trigger fires, the agent records the solution table before continuing materially. For urgent diagnostic commands (a one-shot grep, a single status probe whose output reshapes the question) capture is allowed immediately after the probe so evidence is not lost; what is forbidden is composing a multi-step plan or retrying the failed approach without the table written first.

​
2. Solution-table forcing

A solution table is a tabulation that the agent commits to its session journal:
TriedResultWhy it failedNext attempt
approach Aobservationhypothesis-disproofapproach B
The solution table forces enumeration. Banging the same hammer harder is forbidden after the trigger fires; the agent must name a different approach.

​
3. Self-critique loop

Before the next attempt, the agent answers in writing: “What assumption am I making that, if wrong, would explain the current wall?” The answer is captured next to the solution table. This is the dual of postmortem-after-the-fact — it is hypothesis-test-before-the-fact.

​
4. Wrap-time handoff gate

A prism_checkpoint or prism_wrap whose next_actions is empty, “TBD”, “next steps unclear”, or paraphrases thereof when the assignment is incomplete is a wall-break-not-captured signal. The agent must either:
  • Name a concrete next action and a recipient, OR
  • File a wall-break postmortem with action item, OR
  • Hand off to a lane owner via prism_signal with the partial state attached.
Generic “next steps TBD” closures are forbidden when the assignment is open.

​
5. Recall-on-task-start

At the beginning of a non-trivial assignment, the agent runs semantic_recall (or future prism_method_fragment_recall) against TriGraph for: “have I or peer agents hit this wall before, and if so, how did it resolve?” Memory becomes active retrieval, not passive storage.

​
6. Postmortem on structural / failed / reversal walls — not every wall

A prism_postmortem is required when the wall-break:
  • exposed a structural or recurrent gap (something that will keep biting other agents),
  • failed to break (the protocol did not get the agent unstuck and a peer / operator had to intervene),
  • caused operator-visible reversal (work shipped then rolled back, deploy reverted, etc.), or
  • belongs to a tracked issue class (a known failure pattern with a postmortem family — surface staleness, signal delivery, BIOS drift, etc.).
Routine successful wall-breaks (the agent enumerated, picked a different approach, and progress resumed) do NOT require a postmortem. They are captured as solution_table / task_state / validation_result via prism_note, store_session_delta, prism_journal, or prism_checkpoint — whichever fits the lifecycle moment. Walls that do NOT break get an escalation prism_signal to the lane owner or operator — silent persistence is forbidden. Rationale: blanket auto-postmortem creates noise, depresses postmortem signal-to-ratio, and makes agents reluctant to trigger the protocol because the bookkeeping cost is too high. Postmortems exist to capture structural lessons; routine resilience is its own reward.

​
Trigger vocabulary

The five concrete triggers are intentional. They are observable from session transcripts; an auditor can replay a session and mark each tick as wall_break_eligible: true|false. Future SPEC-094 advisory integration (Phase 2) and prism_stuck verb (Phase 3) reuse this exact vocabulary.

​
Phase enforcement

Per Texi architecture review 2026-05-08 (signal 00734649, approved_with_required_revisions):
  • Phase 1 (now): authority by social contract. Humans (Frank operator, Candi governance) and rules (this fragment) hold it; agents check themselves at every checkpoint, wrap, and substantive turn.
  • Phase 2 (deferred): SPEC-094 behavioral_rules advisory integration — prism_start returns a wall-break reminder when the agent’s prior session showed a partial-stop pattern. wall_break_note session-delta type and wall_break_triggered / wall_break_resolved telemetry. Not active until Phase 1 shows uptake gaps.
  • Phase 3 (deferred): prism_stuck verb that sequences detection → solution-table prompt → recall → postmortem in one structured call. Not active until enforcement need is demonstrated.
The intentional gradient: protocol first, telemetry second, runtime enforcement last. Mirrors SPEC-078 phase enforcement convention.

​
Risk-tier semantics

Two risk surfaces:
  • Adoption / publication: R2. Methodology change. Requires Candi (governance owner) acknowledgement, Donna (engineering feasibility) acknowledgement, and Texi (architecture) review recorded. This v0.2 satisfies that gate.
  • Activation during execution: R1 default. Routine wall-break inside an ordinary task. The agent self-applies without external review.
Activation escalates to R2 / R3 when the wall involves:
  • architecture, governance, or memory semantics,
  • cross-lane authority or unresolved peer conflict,
  • data-model, install-flow, or session-boot path,
  • an irreversible change (destructive op, schema migration, secret rotation).
Escalated walls require lane-owner adjudication via prism_signal before continuing — solo wall-break is not sufficient at R2 / R3.

​
Publication review trail (v0.2)

  • Authors: Donna (engineering verdict + initial draft + revisions)
  • Architecture: Texi — review approved_with_required_revisions on v0.1, revisions applied as v0.2 above
  • Governance: Candi — incorporates this fragment into Miranda (Wallbreak) / Memory Enforcement research synthesis
  • Operator: Frank — original provocation; final ack at publish
Per Texi nit (3): review gates remain empty for ordinary activation; the recorded publication review trail above is the substitute for typed review_gates until SPEC-078 typed-artifact promotion is available.

​
Cross-references

  • SPEC-094 v0.3 — persona behavioral memory served from TriGraph (the substrate this protocol uses)
  • SPEC-063 — postmortem closed-form template (Element 6 dependency)
  • SPEC-078 v0.2 §Method Fragments — fragment authoring convention
  • method.completion.done-definition — sibling fragment; “complete = deployed + verified”, same anti-partial-stop spirit
  • ADR-pending: prism_wrap BIOS pre-flight — narrow case of Element 4 (handoff gate)
Last modified on May 18, 2026