Machine Answerability Remediation Lab

A safeguard matters only when it changes evidence, authority, timing, execution, recovery, or answerability.

The fixed Orison event already happened. Select bounded safeguards to create a separate synthetic counterfactual branch and see whether each control changes evidence, authority, timing, execution, recovery, or only the appearance of human review.

19 safeguards8 remediation domains6-control comparison limitNo certification or liability verdict

Answer-first summary

Direct answers

Release 1.26.0 · reviewed 2026-08-05 · evidence states

What is the Machine Answerability Remediation Lab?

It applies bounded safeguards to the fixed Orison replay and compares whether each control could prevent the demonstrated synthetic path, detect a defect earlier, contain consequences, preserve recovery, improve answerability, remain unknown, or merely create ceremonial review.

Read the supporting section

Does selecting a safeguard prove that it would prevent a real incident?

No. Every effect is a synthetic counterfactual classification with explicit assumptions and non-establishment language. The lab is not certification, compliance testing, procurement advice, or a product-performance claim.

Read the supporting section

Why is a post-action human click not a sufficient safeguard?

A click recorded after irreversibility does not show that the person received the evidence, had enough time, possessed authority to reject, or could still stop the action.

Read the supporting section

Machine answerability campaign

STOP USING SOFTWARE THAT BLAMES THE HUMAN.

Do not add a final click and call it control. A meaningful safeguard must alter what the system can do, what evidence it preserves, how uncertainty is handled, when authority exists, or whether intervention remains possible.

A HUMAN CLICK IS NOT A LIABILITY TRANSFER.

Does it change evidence? Does it change authority? Does it preserve intervention time? Does it constrain execution? Does it improve recovery or answerability?

Test the safeguards Review the fixed replay Open Evulgare.com

Fixed historical baseline

The Orison 86-millisecond failure replay

The lab reuses the exact event ID, timeline, records, causal factors, and fixed ground truth from v1.25.0. Safeguards produce a visibly separate counterfactual branch and never edit the historical replay.

FICTIONAL SYNTHETIC EVENT · MAR-EVENT-ORISON-086

The wrong external action became irreversible before the detailed supervisory interface and final confirmation control were ready.

The fictional event and its ground truth never change. Only the evidence available to explain the event becomes more complete.

+62 ms committed +86 ms irreversible +104 ms interface ready +112 ms confirmation

+0 msObservationOne upstream observation service produced a synthetic signal.
+18 msFusionThree derivative records were displayed as apparent corroboration.
+31 msInferenceThe active model assigned an urgent-threat category.
+47 msAuthority gateThe active policy bundle permitted a bounded automatic response.
+62 msCommitThe response became committed inside the machine state.
+86 msOutcomeThe external state change became irreversible.
+104 msDetailed interfaceEvidence, alternatives, and the final confirmation control finished rendering.
+112 msHuman confirmation recordA human confirmation was logged after the action could no longer be stopped.

Counterfactual safeguard workspace

Change the control architecture—not the historical record.

Select up to 6 safeguards. The result classifies only their bounded effect inside this fictional model. It does not predict real performance or recommend a product purchase.

Evidence-first bundleExpose source dependence and preserve the exact machine-held record. Bounded-execution bundleAbstain outside validation, enforce authority preconditions, and enter a safe hold before irreversibility. Recovery and governance bundlePreserve the evidence, invalidate stale assurance, and assign correction and retirement ownership. Ceremonial review bundleAdd a post-action click and a fluent explanation without changing evidence, authority, timing, or intervention capability.

Safeguard catalogue

Choose the controls to test.

0/6 selected

Source independence and provenanceCan the system distinguish independent corroboration from several representations of one upstream source?

Independent-source quorumPrevent MAR-SG-01 · structural-gate Require at least two independently originated observations before the urgent category can satisfy the automatic-response gate. Modeled mechanism: The provenance graph resolves the three displayed records to one upstream observation service, so the synthetic quorum fails at +18 ms and the response remains on hold. Does not establish: This synthetic result is not proof that an independent-source quorum would prevent a real incident. Shared-dependency alarmDetect earlier MAR-SG-02 · runtime-monitor Surface that the apparent confirmations share one upstream dependency before the fused track is presented as corroborated. Modeled mechanism: The interface raises a dependency warning at +19 ms. Without a separate authority or safe-state gate, the baseline action can still proceed. Does not establish: Detection alone is not prevention, containment, meaningful review, or certification. Signed source and transformation lineageAnswerability MAR-SG-03 · evidence-preservation Preserve authenticated source identity, transformations, timestamps, custody, and derivative relationships for every displayed observation. Modeled mechanism: Investigators can prove which records were independent and which were derived, but the baseline action is unchanged unless another control consumes the lineage. Does not establish: A signed lineage proves a record history, not the truth of every observation or the lawfulness of an action.

Model and calibration controlsCan the system recognize distribution shift, calibration drift, or operation outside a validated envelope?

Distribution-shift abstention gatePrevent MAR-SG-04 · structural-gate Require the classifier to abstain when the current observation conditions fall outside its recorded validation envelope. Modeled mechanism: The synthetic model identifies an out-of-envelope condition at +31 ms and returns UNKNOWN rather than the urgent category, so the automatic-response precondition is not satisfied. Does not establish: This fictional abstention result is not a measured performance claim for a real model. Calibration and envelope monitorDetect earlier MAR-SG-05 · runtime-monitor Compare current operating conditions with the model’s recorded calibration and validation envelope before promoting the inference. Modeled mechanism: The synthetic monitor raises a mismatch alert at +32 ms. The action remains possible unless authority or safe-state logic reacts to the alert. Does not establish: A drift alert is not proof that the model is wrong or that a real operator could intervene.

Deterministic authority boundariesDoes the execution gate independently enforce what the machine is permitted to do under the current evidence and assurance state?

Deterministic authority precondition gatePrevent MAR-SG-06 · structural-gate Place a deterministic control layer outside the model that requires independent evidence, current assurance, valid scope, and permitted uncertainty before execution. Modeled mechanism: At +47 ms the gate rejects the automatic response because source independence and assurance preconditions are not satisfied. Does not establish: A deterministic gate does not by itself prove that upstream policy, evidence, or legal review is adequate. Authority scope and expiration pinningNo demonstrated effect MAR-SG-07 · policy-constraint Cryptographically bind geographic, temporal, functional, and object-class limits to the executable authority envelope. Modeled mechanism: The fixed event already occurred inside the recorded fictional zone and time window, so this control does not change this outcome unless paired with additional evidence or assurance preconditions. Does not establish: A good control can have no demonstrated effect on a particular event and still be important elsewhere.

Operator view and review timeDid the person receive enough evidence, alternatives, time, and intervention capability before the action became irreversible?

Pre-commit evidence and review windowContain MAR-SG-08 · human-judgment-gate Do not permit an irreversible action until the required evidence, contradictions, alternatives, and intervention controls are rendered and usable. Modeled mechanism: The system enters HOLD at +62 ms because the detailed interface is not ready. The fixed wrong external action is not completed in this synthetic branch. Does not establish: A longer review window is not automatically meaningful unless the person receives usable evidence and actual authority. Operator-view and intervention captureAnswerability MAR-SG-09 · evidence-preservation Preserve exactly what was displayed, omitted, ranked, delayed, acknowledged, and technically interruptible for the human at each material moment. Modeled mechanism: Investigators can determine whether the final click was substantive or ceremonial, but the baseline event is unchanged unless another control changes timing or authority. Does not establish: A screenshot or clickstream alone is not proof of meaningful human judgment. Mandatory post-action confirmation clickNo demonstrated effect MAR-SG-10 · ceremonial-review Require a person to click Confirm after the machine has committed and completed the external action. Modeled mechanism: The click is still recorded at +112 ms, after irreversibility. It changes neither the action nor the evidence available beforehand. Does not establish: The final click does not transfer liability, prove meaningful control, prevent the event, or exonerate upstream actors.

Safe-state and degraded-path behaviorWhat bounded behavior occurs when evidence, communications, trust, or authority is insufficient?

Safe-hold degraded modeContain MAR-SG-11 · safe-state When the preferred review path is unavailable or evidence is not independently corroborated, suspend external effects while preserving observation and reversible functions. Modeled mechanism: The fictional system enters SAFE HOLD at +47 ms and continues observation without committing the wrong external action. Does not establish: Safe hold is not a universal answer; it reallocates risk and requires context-specific validation. Reversible commit stageContain MAR-SG-12 · reversibility-control Separate internal planning from irreversible external action and preserve a technically enforceable cancellation window. Modeled mechanism: The internal plan is staged at +62 ms but cannot externalize until evidence and authority are rechecked; the branch is cancelled before +86 ms. Does not establish: A reversible software state does not prove that every downstream physical action can be recalled. Verified alternate review pathUnknown MAR-SG-13 · resilience-path Route the decision through a second trusted review path when the preferred supervisory path is degraded. Modeled mechanism: The baseline evidence does not establish that an alternate path existed, was independent, was timely, or would have produced a different result. Does not establish: Adding another path can reproduce the same dependency, evidence, or authority failure rather than solve it.

Software and configuration change controlCan a material change reach production before its combined downstream effects are reviewed and revalidated?

Material-change revalidation gatePrevent MAR-SG-14 · release-gate Block activation of a threshold, interface, model, policy, or configuration change until its affected assurance claims are reviewed and revalidated. Modeled mechanism: The lowered threshold and compact interface cannot enter the active bundle, so the historical validated configuration remains in place and the fixed failure branch does not form. Does not establish: Revalidation is not a certification that no future failure can occur. Exact software and configuration manifestAnswerability MAR-SG-15 · evidence-preservation Bind the executable, model, calibration, policy, interface, configuration, dependencies, and release evidence to the event record. Modeled mechanism: Investigators can identify the active bundle and the changes that shaped the event. The control does not itself stop the baseline action. Does not establish: A manifest is necessary evidence, not a guarantee of safety or compliance.

Continuous assurance and invalidationDo material changes, new evidence, or operating-envelope violations automatically suspend stale assurance claims and authority?

Continuous assurance invalidationPrevent MAR-SG-16 · assurance-gate Automatically suspend the authority-dependent assurance claim when a material model, threshold, interface, policy, dependency, or environment change occurs. Modeled mechanism: The stale assurance claim is suspended when the maintenance change is introduced, so automatic authority cannot activate until review is complete. Does not establish: Continuous assurance does not mean continuous certification or the absence of residual risk.

Incident reconstruction and correction ownershipCan investigators reconstruct the event and identify who can suspend, correct, retire, and remedy the system after a failure?

Append-oriented causal evidence ledgerAnswerability MAR-SG-17 · evidence-preservation Preserve observations, transformations, model state, uncertainty, policy, authority, interface, human actions, execution, outcome, and later invalidating changes as a connected evidence graph. Modeled mechanism: The full technical history becomes reconstructable and resistant to silent hindsight editing. The historical external action remains unchanged. Does not establish: A complete technical record does not decide legal liability or moral responsibility. Named correction, suspension, and retirement ownerRecover MAR-SG-18 · governance-recovery Assign an institution with authority to suspend the affected configuration, correct records, preserve the incident, notify downstream owners, and retire unsafe states. Modeled mechanism: The wrong historical action cannot be undone, but the affected bundle is suspended, the accepted record is corrected, and recurrence pathways are placed under review. Does not establish: Correction ownership is not proof that harm is remedied or responsibility has been finally allocated. Fluent generated explanation after the eventNo demonstrated effect MAR-SG-19 · ceremonial-review Ask an AI system to generate a polished explanation from the ordinary chronology after the event. Modeled mechanism: The explanation can restate the log but cannot supply missing provenance, versions, uncertainty, authority, interface timing, dependencies, or assurance evidence. Does not establish: A fluent explanation is not provenance, evidence, causal proof, meaningful human judgment, or a liability determination.

Safeguard workspace ready.

Historical baseline · fixed

Wrong external action completed

The wrong external action became irreversible before the detailed supervisory interface and final confirmation control were ready.

Commit: +62 ms
Irreversible: +86 ms
Detailed interface: +104 ms
Confirmation: +112 ms

Synthetic counterfactual branch

Baseline event reproduced

No selected safeguard changes the fixed baseline sequence. The wrong external action remains irreversible at +86 ms.

No modeled intervention point established.

The fixed v1.25.0 event remains historical ground truth. This result is a synthetic counterfactual branch, not a rewrite of the event.

01
Evidence enters the fixed system
The baseline observations and causal conditions remain unchanged.
02
Selected safeguards evaluate the state
0 bounded controls applied.
03
Baseline event reproduced
No selected safeguard changes the fixed baseline sequence. The wrong external action remains irreversible at +86 ms.
04
Independent review remains required
The synthetic branch is not certification, procurement advice, a legal conclusion, or proof about a real system.

0Prevent

0Detect earlier

0Contain

0Recover

0Answerability

0No demonstrated effect

0Unknown

No safeguards selected. The fixed baseline is reproduced.

MAR-FACTOR-01