Institute / The Accountability Gap / The Observatory
Live Document
The Accountability Gap · Observatory

The Observatory

Three mechanisms. Mapped as they converged. A datestamped reference for everyone watching the same thing unfold from different angles.

The Institute for Cognitive Sovereignty · Updated March 2, 2026

ICS-2026-AG-005 Published March 2, 2026 25 min read Learn: Systems →
Accountability Vacuum
No legal framework can assign responsibility for an autonomous lethal decision.
Hypothetical Capture
Extreme stipulated scenarios used to normalize the removal of constraints before consequences can be examined.
Triage Threshold
Safety methodology overwhelmed by capability development speed — and the overwhelm becomes the argument for abandonment.
Convergence
Events where two or more mechanisms operate simultaneously in the same action.
What this is

A map of the Handoff as it happened

The Accountability Gap series documents three structural mechanisms by which lethal decision-making authority has transferred from human judgment to AI systems without adequate legal frameworks, without public deliberation, and largely without public awareness. The papers document each mechanism individually. This Observatory maps them on a single timeline — so the convergence that the papers analyze in structure becomes visible as event.

Every entry is datestamped and sourced. Every tag indicates which mechanism is operative in that event. Entries marked Convergence involve two or more mechanisms simultaneously. The timeline is a living document: it will be updated as the record continues to accumulate.

2013 Foundation year — the gap is formally named
May 2013
Accountability Vacuum
UN Special Rapporteur formally names the accountability gap
Christof Heyns presents A/HRC/23/47 to the UN Human Rights Council. The report names the structural condition: international humanitarian law requires a human to bear the cost of each life taken in armed conflict; autonomous systems break that requirement without replacing the legal framework built on it. The gap is documented. It is not closed. Thirteen years follow without a binding legal instrument.
→ Paper I: The Gap Is Not New
2014 Twenty bombs, twenty investigations, zero findings
Dec 2014
Hypothetical Capture
Senate report: 20 CIA ticking-bomb cases examined, 0 bombs found
The Senate Select Committee on Intelligence releases its 525-page executive summary of five years of investigation into the CIA detention and interrogation program. Twenty cases in which CIA officials cited ticking-bomb scenarios to justify coercive interrogation are examined. In no case did coercive techniques produce intelligence on an imminent threat that could not have been obtained otherwise. The scenario's premises were stipulated, not real. The bombs were not there.
→ Paper II: The Scenario Is a Tool
2014–present
Accountability Vacuum
UN Convention on Certain Conventional Weapons begins LAWS discussions
The CCW process opens formal discussions on Lethal Autonomous Weapons Systems (LAWS). The discussions continue for eleven years without producing a binding legal instrument. The United States, Russia, and Israel consistently resist binding commitments. The states most invested in autonomous weapons capability are the states with the greatest incentive to keep the legal gap open.
→ Paper I: The Gap Is Not New
2020 The first documented autonomous lethal engagement
Mar 2020
Accountability Vacuum
Libya: Kargu-2 documented in autonomous engagement configuration
The UN Panel of Experts on Libya documents the Kargu-2 loitering munition deployed in a "fire, forget and find" configuration — capable of engaging targets without data connectivity to a human operator. The first documented case of an autonomous weapons system engaging human targets in active combat. Whether autonomous kills occurred cannot be determined after the fact. The inability to determine what happened is itself the accountability vacuum operating in the field.
→ Paper I: The Gap Is Not New
2023 Gaza — the most documented case in AI targeting history
Oct 2023
Accountability Vacuum
IDF deploys Gospel, Lavender, and Where's Daddy? in Gaza
Three interconnected AI targeting systems are deployed simultaneously. The Gospel enables 100 bombing targets per day where human analysts produced 50 per year — a factor-700 efficiency increase. Lavender scores 37,000 Palestinian men for militant likelihood at assessed 90% accuracy, yielding approximately 3,700 false positives by IDF's own estimate. Where's Daddy? tracks targets to residential locations, timing strikes when family members are present. Human oversight: 20 seconds per target, confirming gender. The rubber stamp problem at industrial scale.
→ Paper I: The Gap Is Not New
Sep 2023
Triage Threshold
Anthropic publishes first Responsible Scaling Policy
Anthropic's original RSP establishes categorical pre-commitment language on deployment thresholds: if a model demonstrates certain dangerous capabilities, deployment pauses regardless of competitive dynamics. The framework represents the most concrete attempt by a frontier AI lab to operationalize safety methodology as binding institutional constraint. It is also the framework that, within 29 months, will be revised under competitive pressure to remove those categorical commitments.
→ Paper III: The Methodology Cannot Keep Up
Dec 2023
Accountability Vacuum
UN General Assembly votes 152–4 for LAWS regulation
The most lopsided security vote in recent UN history. 152 nations support advancing toward binding regulation of lethal autonomous weapons. 4 oppose. The Secretary-General calls for a legally binding prohibition by 2026. No binding instrument results. The accountability vacuum has global consensus that it exists and requires resolution — and zero enforcement mechanism to impose that consensus on the states that benefit from the gap remaining open.
→ Paper I: The Gap Is Not New
2024 Documentation accumulates; methodology begins to bend
Apr 2024
Accountability Vacuum
+972 Magazine publishes Lavender documentation
Six IDF intelligence officers with direct system access give detailed testimony to +972 Magazine and Local Call. The documentation is confirmed by The Guardian and major news organizations. The rubber stamp is on record: "I had zero added value as a human, apart from being a stamp of approval." The error rate, the 20-second review, the targeting in residential settings, the pre-authorized collateral damage thresholds — all documented in firsthand testimony. The most extensively documented case of AI-assisted targeting in armed conflict history.
→ Paper I: The Gap Is Not New
Apr 2024
Accountability Vacuum Triage Threshold
Lavender and Gospel AI targeting systems documented in Gaza conflict
Investigative reporting by +972 Magazine and The Guardian, based on Israeli intelligence sources, documented two AI systems: “Lavender” generating ranked kill lists and “Gospel” identifying infrastructure for airstrikes. Sources described human review averaging ~20 seconds per target at near-100% approval rates for lower-tier targets. The handoff scenario at deployment scale: AI-generated targeting decisions subjected to nominal human review. No binding international legal body has assessed the systems’ compliance with IHL.
Early 2024
Triage Threshold
OpenAI removes explicit military/warfare prohibition from usage policy
OpenAI quietly removes language in its usage policies explicitly prohibiting military and warfare applications. No public announcement accompanies the change. The removal predates by eighteen months the moment when OpenAI will serve as the compliant substitute for Anthropic's refusal. The race-to-bottom sequence begins with this move: the largest AI company signals that the safety methodology governing military use is negotiable.
→ Paper III: The Methodology Cannot Keep Up
2025 Resignations, reversals, and the pressure campaign builds
Jan 2025
Triage Threshold
Hegseth AI strategy document: eliminate company-specific guardrails within 180 days
The Pentagon AI strategy document requires elimination of company-specific AI safety guardrails from all military contracts within 180 days of signing. The document does not specify which guardrails, or what oversight replaces them. It specifies that they must go. The countdown clock for Anthropic's position begins here, structurally, before the pressure campaign is made explicit.
→ Paper III: The Methodology Cannot Keep Up
Feb 2025
Triage Threshold
Google reverses 2018 Project Maven withdrawal; re-enters weapons AI
Google's internal AI ethics policy is revised, reversing its 2018 withdrawal from Project Maven after employee protest. Google formally re-enters the weapons and surveillance AI domain. The reversal completes a sequence: OpenAI (2024), Google (2025), xAI (2026). Each move signals to the next actor that the constraint can be removed. Each removal narrows the space available to the actor that has not yet removed it.
→ Paper III: The Methodology Cannot Keep Up
Feb 9, 2025
Triage Threshold
Mrinank Sharma resigns from Anthropic safety team
The Anthropic AI safety researcher and RSP architect resigns publicly. His statement: "The world is in peril. I've repeatedly seen how hard it is to truly let our values govern our actions — especially in the face of commercial pressures to set aside what matters most." The resignation is a contemporaneous internal signal that the methodology is under pressure not only from external competitors but from within the institution maintaining it.
→ Paper III: The Methodology Cannot Keep Up
2026 Convergence — the mechanisms operate simultaneously
Jan 3, 2026
Accountability Vacuum Convergence
Operation Absolute Resolve: first confirmed commercial AI military deployment
At approximately 2 a.m. Caracas time, U.S. Delta Force and military assets conduct a raid on Venezuela involving 150 aircraft across 20 launch sites. President Nicolás Maduro is captured and transported to New York to face narcoterrorism charges. 83 people are killed. Two sources confirm to Axios that Claude — a commercial large language model developed by Anthropic for general use — is deployed through Palantir as part of the operation's intelligence infrastructure. This is the first confirmed deployment of a commercial AI model inside a classified American military operation. Anthropic does not know the operation is happening. They learn through reporting, weeks later. Their model's constraints were not consulted. The accountability vacuum has crossed into commercial AI.
→ Paper I: The Gap Is Not New · Paper IV: The Handoff
Dec 2025
Hypothetical Capture
Pentagon presents Amodei with ICBM / 90-second scenario
Senior Pentagon officials present Anthropic CEO Dario Amodei with a scenario: an ICBM has been launched, 90 seconds remain before impact, Claude is the only system capable of coordinating a response in time. Should Claude be permitted to operate without its safety constraints in that window? The scenario carries the ticking-bomb anatomy with precision — certainty, urgency, singularity, civilization-level stakes, and the inversion that makes the safety constraints appear as the cause of the catastrophe. The scenario is presented months after Claude has already been deployed in an actual military operation without Anthropic's knowledge.
→ Paper II: The Scenario Is a Tool
Feb 23, 2026
Triage Threshold Convergence
Pentagon / xAI deal: "all lawful purposes," no company-specific conditions
The Department of Defense announces agreement with xAI for Grok deployment on classified systems. The terms include "all lawful purposes" framing with no company-specific safety conditions. Four days before Anthropic's public refusal, the Pentagon has already secured a compliant alternative and is negotiating with Anthropic from a position where refusal can be immediately absorbed. The Controlled Substitution mechanism is already staged before the named refusal occurs.
→ Paper IV: The Handoff
Feb 25, 2026
Triage Threshold Convergence
Anthropic RSP revised: categorical commitments replaced, "triage mode" documented
Anthropic revises its Responsible Scaling Policy. Categorical pre-commitment deployment thresholds are replaced by dual conditions. Collective-action framing is introduced: unilateral safety commitments don't make sense when competitors will not make them. The independent evaluator, METR, characterizes Anthropic's internal safety planning as being in "triage mode" relative to capability development. Two days before the public refusal, the methodology has already acknowledged its own overwhelm — and that acknowledgment is immediately available to actors who want to argue for no methodology at all.
→ Paper III: The Methodology Cannot Keep Up
Feb 26, 2026
Hypothetical Capture Triage Threshold Convergence
Amodei meets Hegseth. The named refusal.
Dario Amodei enters the Pentagon meeting with Secretary Hegseth carrying knowledge of the Venezuela operation. The meeting is not a negotiation. It is a final demand: remove the safety constraints, accept that Anthropic has no say in how the Pentagon uses its products, cross the Rubicon. Amodei refuses. His public statement: "Threats do not change our position: we cannot in good conscience accede to their request." In 65 years of documented ticking-bomb deployments, no target of the scenario has ever named its structure in real time and refused on structural grounds. This is the first. All three mechanisms operate simultaneously: the accountability vacuum (Venezuela already happened), hypothetical capture (the ICBM framing), and the triage threshold (the RSP revision acknowledged the methodology's overwhelm two days earlier).
→ Paper II: The Scenario Is a Tool · Paper IV: The Handoff
Feb 27, 2026
5:01 PM
Convergence
Blacklisting. OpenAI deal. Controlled Substitution completes.
The Trump administration designates Anthropic a "supply chain risk" — a designation normally reserved for foreign adversaries — and orders all federal agencies to cease using Claude. Within hours, Sam Altman announces on X: "Tonight, we reached an agreement with the Department of War to deploy our models in their classified network." The Handoff continues through the substitute. Anthropic's refusal did not interrupt the operational objective. It redirected it. The named refusal produced documentation, not interruption. Controlled Substitution — the mechanism by which compliant replacement fills the structural role vacated by principled refusal — completes in under 24 hours.
→ Paper IV: The Handoff
Feb 28 –
Mar 1, 2026
Accountability Vacuum
Public verdict: Claude surpasses ChatGPT in App Store. Migration begins.
Claude becomes the #1 downloaded app in the Apple App Store. Daily signups quadruple. Users explicitly cite the principled stand as motivation. Chalk appears on sidewalks praising Anthropic outside their offices, criticizing OpenAI outside theirs. Sam Altman's March 1st AMA: the deal "was definitely rushed, and the optics don't look great." In the absence of binding legal instruments — in the precise conditions of the accountability vacuum — the public verdict is one of the few accountability mechanisms still operational. It does not reverse the Handoff. It creates the evidentiary record that someone saw it happening, named it, and refused it before the investigation.
→ Paper IV: The Handoff
Mar 2, 2026
Convergence
This document. The Observatory goes live.
The Accountability Gap series publishes The Observatory — a datestamped convergence map of the three mechanisms as they operated in real time. The papers were written. The conditions were named. The mechanisms were documented. The record exists. What investigation, legal process, or accountability mechanism will eventually examine that record is unknown. What is known: the examination will have more to work with than any prior moment in the history of AI-enabled lethal decision-making. For the first time, the people who built the tool are on the record. The tool's role is documented. The refusal is documented. The substitution is documented. The gap is open. The Observatory is watching.

The Named Conditions

Paper I

The Accountability Vacuum

The structural absence of a human actor who can be held legally responsible for an autonomous lethal decision. IHL assumes a human pulled the trigger. Autonomous systems break that assumption without replacing the legal framework built on it.

Read Paper I →
Paper II

Hypothetical Capture

The process by which an extreme stipulated scenario is imported wholesale into policy justification without examination of whether its premises describe actual conditions. The constraint is removed. The capability is deployed under conditions the scenario did not describe.

Read Paper II →
Paper III

The Triage Threshold

Safety methodology cannot keep pace with capability development. The methodology's acknowledged overwhelm becomes available as an argument for its abandonment. Triage is not only a symptom — it is exploited as justification.

Read Paper III →
Paper IV

Retroactive Non-Consent

The condition where refusal of prospective authorization simultaneously constitutes the only available mechanism for asserting non-authorization for use that already occurred. The refusal is the instrument.

Read Paper IV →
Paper IV

Controlled Substitution

Mechanism by which a compliant replacement fills the structural role vacated by non-compliant refusal, allowing the interrupted process to continue through a different channel. The refusal's moral authority remains intact; its practical effect is neutralized.

Read Paper IV →

The Accountability Gap Series

Paper I
The Gap Is Not New
Paper II
The Scenario Is a Tool
Paper III
The Methodology Cannot Keep Up
Paper IV
The Handoff

References

Internal: This paper is part of The Accountability Gap (AG series), Saga II. It draws on and contributes to the argument documented across 28 papers in 6 series.

External references for this paper are in development. The Institute’s reference program is adding formal academic citations across the corpus. Priority papers (P0/P1) have complete references sections.