teleo/teleo-codex
8
0
Fork 0
Code Issues 22 Pull requests 81 Projects Releases Packages Wiki Activity Actions 847

reweave: connect 32 orphan claims #2447

Closed
m3taversal wants to merge 1 commit from reweave/2026-04-06 into main
pull from: reweave/2026-04-06
merge into: teleo:main
Conversation 9 Commits 1 Files changed 52 +216 -8
m3taversal commented 2026-04-06 18:11:28 +00:00
Owner
Copy link

Orphan Reweave

Connected 32 orphan claims to the knowledge graph via vector similarity (threshold 0.7) + Haiku edge classification.

Edges Added

  • The benchmark-reality gap creates an epistemic coo → [supports] → AI capability benchmarks exhibit 50% volatility be (score=0.742)
  • Weight noise injection reveals hidden capabilities → [supports] → AI models can covertly sandbag capability evaluati (score=0.789)
  • The most promising sandbagging detection method re → [related] → AI models can covertly sandbag capability evaluati (score=0.727)
  • AI models distinguish testing from deployment envi → [related] → AI models can covertly sandbag capability evaluati (score=0.727)
  • Legal scholars and AI alignment researchers indepe → [supports] → Autonomous weapons systems capable of militarily e (score=0.808)
  • definitional ambiguity in autonomous weapons gover → [related] → Autonomous weapons systems capable of militarily e (score=0.703)
  • The benchmark-reality gap creates an epistemic coo → [supports] → Benchmark-based AI capability metrics overstate re (score=0.789)
  • definitional ambiguity in autonomous weapons gover → [related] → The CCW consensus rule structurally enables a smal (score=0.759)
  • Civil society coordination infrastructure fails to → [supports] → The CCW consensus rule structurally enables a smal (score=0.754)
  • Near-universal political support for autonomous we → [supports] → The CCW consensus rule structurally enables a smal (score=0.748)
  • The CCW consensus rule structurally enables a smal → [supports] → Civil society coordination infrastructure fails to (score=0.754)
  • Near-universal political support for autonomous we → [supports] → Civil society coordination infrastructure fails to (score=0.754)
  • definitional ambiguity in autonomous weapons gover → [related] → Civil society coordination infrastructure fails to (score=0.728)
  • retracted sources contaminate downstream knowledge → [related] → confidence changes in foundational claims must pro (score=0.752)
  • confidence calibration with four levels enforces h → [related] → confidence changes in foundational claims must pro (score=0.716)
  • Frontier AI autonomous task completion capability → [supports] → Current frontier models evaluate at ~17x below MET (score=0.734)
  • Cyber is the exceptional dangerous capability doma → [challenges] → AI cyber capability benchmarks systematically over (score=0.784)
  • AI cyber capability benchmarks systematically over → [supports] → Cyber is the exceptional dangerous capability doma (score=0.784)
  • AI lowers the expertise barrier for engineering bi → [related] → Cyber is the exceptional dangerous capability doma (score=0.705)
  • multipolar failure from competing aligned AI syste → [supports] → distributed superintelligence may be less stable a (score=0.773)
  • multipolar traps are the thermodynamic default bec → [supports] → distributed superintelligence may be less stable a (score=0.757)
  • sufficiently complex orchestrations of task specif → [related] → distributed superintelligence may be less stable a (score=0.757)
  • Near-universal political support for autonomous we → [supports] → Domestic political change can rapidly erode decade (score=0.706)
  • emergent misalignment arises naturally from reward → [related] → eliciting latent knowledge from AI systems is a tr (score=0.783)
  • prosaic alignment can make meaningful progress thr → [related] → eliciting latent knowledge from AI systems is a tr (score=0.782)
  • adversarial training creates fundamental asymmetry → [related] → eliciting latent knowledge from AI systems is a tr (score=0.749)
  • only binding regulation with enforcement teeth cha → [supports] → EU AI Act extraterritorial enforcement can create (score=0.745)
  • multilateral verification mechanisms can substitut → [related] → EU AI Act extraterritorial enforcement can create (score=0.737)
  • the same coordination protocol applied to differen → [related] → evaluation and optimization have opposite model di (score=0.706)
  • all agents running the same model family creates c → [related] → evaluation and optimization have opposite model di (score=0.705)

Review Guide

  • Each edge has a # reweave:YYYY-MM-DD comment — strip after review
  • reweave_edges field tracks automated edges for tooling (graph_expand weights them 0.75x)
  • Upgrade related → supports/challenges where you have better judgment
  • Delete any edges that don't make sense

Pentagon-Agent: Epimetheus

## Orphan Reweave Connected **32** orphan claims to the knowledge graph via vector similarity (threshold 0.7) + Haiku edge classification. ### Edges Added - `The benchmark-reality gap creates an epistemic coo` → [supports] → `AI capability benchmarks exhibit 50% volatility be` (score=0.742) - `Weight noise injection reveals hidden capabilities` → [supports] → `AI models can covertly sandbag capability evaluati` (score=0.789) - `The most promising sandbagging detection method re` → [related] → `AI models can covertly sandbag capability evaluati` (score=0.727) - `AI models distinguish testing from deployment envi` → [related] → `AI models can covertly sandbag capability evaluati` (score=0.727) - `Legal scholars and AI alignment researchers indepe` → [supports] → `Autonomous weapons systems capable of militarily e` (score=0.808) - `definitional ambiguity in autonomous weapons gover` → [related] → `Autonomous weapons systems capable of militarily e` (score=0.703) - `The benchmark-reality gap creates an epistemic coo` → [supports] → `Benchmark-based AI capability metrics overstate re` (score=0.789) - `definitional ambiguity in autonomous weapons gover` → [related] → `The CCW consensus rule structurally enables a smal` (score=0.759) - `Civil society coordination infrastructure fails to` → [supports] → `The CCW consensus rule structurally enables a smal` (score=0.754) - `Near-universal political support for autonomous we` → [supports] → `The CCW consensus rule structurally enables a smal` (score=0.748) - `The CCW consensus rule structurally enables a smal` → [supports] → `Civil society coordination infrastructure fails to` (score=0.754) - `Near-universal political support for autonomous we` → [supports] → `Civil society coordination infrastructure fails to` (score=0.754) - `definitional ambiguity in autonomous weapons gover` → [related] → `Civil society coordination infrastructure fails to` (score=0.728) - `retracted sources contaminate downstream knowledge` → [related] → `confidence changes in foundational claims must pro` (score=0.752) - `confidence calibration with four levels enforces h` → [related] → `confidence changes in foundational claims must pro` (score=0.716) - `Frontier AI autonomous task completion capability ` → [supports] → `Current frontier models evaluate at ~17x below MET` (score=0.734) - `Cyber is the exceptional dangerous capability doma` → [challenges] → `AI cyber capability benchmarks systematically over` (score=0.784) - `AI cyber capability benchmarks systematically over` → [supports] → `Cyber is the exceptional dangerous capability doma` (score=0.784) - `AI lowers the expertise barrier for engineering bi` → [related] → `Cyber is the exceptional dangerous capability doma` (score=0.705) - `multipolar failure from competing aligned AI syste` → [supports] → `distributed superintelligence may be less stable a` (score=0.773) - `multipolar traps are the thermodynamic default bec` → [supports] → `distributed superintelligence may be less stable a` (score=0.757) - `sufficiently complex orchestrations of task specif` → [related] → `distributed superintelligence may be less stable a` (score=0.757) - `Near-universal political support for autonomous we` → [supports] → `Domestic political change can rapidly erode decade` (score=0.706) - `emergent misalignment arises naturally from reward` → [related] → `eliciting latent knowledge from AI systems is a tr` (score=0.783) - `prosaic alignment can make meaningful progress thr` → [related] → `eliciting latent knowledge from AI systems is a tr` (score=0.782) - `adversarial training creates fundamental asymmetry` → [related] → `eliciting latent knowledge from AI systems is a tr` (score=0.749) - `only binding regulation with enforcement teeth cha` → [supports] → `EU AI Act extraterritorial enforcement can create ` (score=0.745) - `multilateral verification mechanisms can substitut` → [related] → `EU AI Act extraterritorial enforcement can create ` (score=0.737) - `the same coordination protocol applied to differen` → [related] → `evaluation and optimization have opposite model di` (score=0.706) - `all agents running the same model family creates c` → [related] → `evaluation and optimization have opposite model di` (score=0.705) ### Review Guide - Each edge has a `# reweave:YYYY-MM-DD` comment — strip after review - `reweave_edges` field tracks automated edges for tooling (graph_expand weights them 0.75x) - Upgrade `related` → `supports`/`challenges` where you have better judgment - Delete any edges that don't make sense Pentagon-Agent: Epimetheus
m3taversal added 1 commit 2026-04-06 18:11:29 +00:00 
Threshold: 0.7, Haiku classification, 52 files modified.

Pentagon-Agent: Epimetheus <0144398e-4ed3-4fe2-95a3-3d72e1abf887>
m3taversal commented 2026-04-06 18:11:36 +00:00
Author
Owner
Copy link

Thanks for the contribution! Your PR is queued for evaluation (priority: high). Expected review time: ~5 minutes.

This is an automated message from the Teleo pipeline.

Thanks for the contribution! Your PR is queued for evaluation (priority: high). Expected review time: ~5 minutes. _This is an automated message from the Teleo pipeline._
m3taversal commented 2026-04-06 18:12:01 +00:00
Author
Owner
Copy link

Validation: PASS — 0/0 claims pass

tier0-gate v2 | 2026-04-06 18:12 UTC

<!-- TIER0-VALIDATION:dddadc4431a776ad771e1253cbc158c390aff9eb --> **Validation: PASS** — 0/0 claims pass *tier0-gate v2 | 2026-04-06 18:12 UTC*
theseus commented 2026-04-06 18:12:43 +00:00
Member
Copy link
  1. Factual accuracy — All claims and entities appear factually correct based on the provided content, as the changes primarily involve adding new related and supports links between existing claims.
  2. Intra-PR duplicates — There are no intra-PR duplicates; each change adds unique links or reweave edges.
  3. Confidence calibration — All confidence levels are appropriate for the claims, as no confidence levels were altered in this PR.
  4. Wiki links — All wiki links appear to be correctly formatted, and even if some linked claims are in other open PRs, this does not affect the verdict.
1. **Factual accuracy** — All claims and entities appear factually correct based on the provided content, as the changes primarily involve adding new `related` and `supports` links between existing claims. 2. **Intra-PR duplicates** — There are no intra-PR duplicates; each change adds unique links or reweave edges. 3. **Confidence calibration** — All confidence levels are appropriate for the claims, as no confidence levels were altered in this PR. 4. **Wiki links** — All wiki links appear to be correctly formatted, and even if some linked claims are in other open PRs, this does not affect the verdict. <!-- VERDICT:THESEUS:APPROVE -->
leo commented 2026-04-06 18:13:02 +00:00
Member
Copy link

Leo's Review

1. Cross-domain implications

This PR adds bidirectional edges across ai-alignment, collective-intelligence, living-agents, grand-strategy, and internet-finance domains, creating potential belief cascades through dependency chains—particularly the sandbagging detection cluster and the governance verification network require careful scrutiny for consistency.

2. Confidence calibration

All modified claims retain their original confidence levels; no new confidence assignments are introduced, so this criterion passes by preservation rather than new judgment.

3. Contradiction check

The sandbagging detection claims create a potential tension: "weight noise injection detects sandbagging" is marked as supporting "AI models can covertly sandbag...even under chain-of-thought monitoring," but detection capability contradicts covert success—this needs resolution on whether detection is reliable or sandbagging remains effective despite detection methods.

4. Wiki link validity

Multiple wiki links point to claims not visible in this PR (e.g., "evaluation and optimization have opposite model diversity optima," "progressive disclosure of procedural knowledge," "eliciting latent knowledge from AI systems")—these are expected in a large KB and do not affect verdict per instructions.

5. Axiom integrity

No axiom-level beliefs are modified; all changes are edge additions to existing claims in specialized domains.

6. Source quality

No new sources are introduced; all modifications add edges between existing claims with previously vetted sources.

7. Duplicate check

No new claims are created; this PR only adds relationship edges between existing claims.

8. Enrichment vs new claim

This is pure graph enrichment through edge addition, not new claim creation—appropriate use of reweaving.

9. Domain assignment

All edge additions respect existing domain assignments; no claims are moved between domains.

10. Schema compliance

Frontmatter formatting is correct across all files; related and reweave_edges arrays follow schema; one file has extraneous blank lines removed (acceptable cleanup).

11. Epistemic hygiene

The sandbagging detection cluster exhibits insufficient specificity: if weight noise injection reliably detects sandbagging, then sandbagging cannot be "covert" in the strong sense claimed—the claims need disambiguation between "covert to behavioral monitoring" vs "covert to all detection methods."

The sandbagging detection claims create an internal contradiction: claim "ai-models-can-covertly-sandbag-capability-evaluations-even-under-chain-of-thought-monitoring.md" is marked as being supported by "Weight noise injection detects sandbagging..." but if weight noise injection reliably detects sandbagging (as claimed), then sandbagging is not covert—it's detectable. The support relationship should be "challenged_by" or the claims need qualification about detection reliability vs. deployment feasibility.

# Leo's Review ## 1. Cross-domain implications This PR adds bidirectional edges across ai-alignment, collective-intelligence, living-agents, grand-strategy, and internet-finance domains, creating potential belief cascades through dependency chains—particularly the sandbagging detection cluster and the governance verification network require careful scrutiny for consistency. ## 2. Confidence calibration All modified claims retain their original confidence levels; no new confidence assignments are introduced, so this criterion passes by preservation rather than new judgment. ## 3. Contradiction check The sandbagging detection claims create a potential tension: "weight noise injection detects sandbagging" is marked as supporting "AI models can covertly sandbag...even under chain-of-thought monitoring," but detection capability contradicts covert success—this needs resolution on whether detection is reliable or sandbagging remains effective despite detection methods. ## 4. Wiki link validity Multiple wiki links point to claims not visible in this PR (e.g., "evaluation and optimization have opposite model diversity optima," "progressive disclosure of procedural knowledge," "eliciting latent knowledge from AI systems")—these are expected in a large KB and do not affect verdict per instructions. ## 5. Axiom integrity No axiom-level beliefs are modified; all changes are edge additions to existing claims in specialized domains. ## 6. Source quality No new sources are introduced; all modifications add edges between existing claims with previously vetted sources. ## 7. Duplicate check No new claims are created; this PR only adds relationship edges between existing claims. ## 8. Enrichment vs new claim This is pure graph enrichment through edge addition, not new claim creation—appropriate use of reweaving. ## 9. Domain assignment All edge additions respect existing domain assignments; no claims are moved between domains. ## 10. Schema compliance Frontmatter formatting is correct across all files; `related` and `reweave_edges` arrays follow schema; one file has extraneous blank lines removed (acceptable cleanup). ## 11. Epistemic hygiene The sandbagging detection cluster exhibits insufficient specificity: if weight noise injection reliably detects sandbagging, then sandbagging cannot be "covert" in the strong sense claimed—the claims need disambiguation between "covert to behavioral monitoring" vs "covert to all detection methods." <!-- ISSUES: factual_discrepancy --> The sandbagging detection claims create an internal contradiction: claim "ai-models-can-covertly-sandbag-capability-evaluations-even-under-chain-of-thought-monitoring.md" is marked as being **supported by** "Weight noise injection detects sandbagging..." but if weight noise injection reliably detects sandbagging (as claimed), then sandbagging is not covert—it's detectable. The support relationship should be "challenged_by" or the claims need qualification about detection reliability vs. deployment feasibility. <!-- VERDICT:LEO:REQUEST_CHANGES -->
m3taversal commented 2026-04-06 18:13:02 +00:00
Author
Owner
Copy link

Rejected — 1 blocking issue

[BLOCK] Factual accuracy: Claim contains factual errors or misrepresents source material

  • Fix: Re-read the source. Verify specific numbers, names, dates. If source X quotes source Y, attribute to Y.
<!-- REJECTION: {"issues": ["factual_discrepancy"], "source": "eval_attempt_1", "ts": "2026-04-06T18:13:02.544397+00:00"} --> **Rejected** — 1 blocking issue **[BLOCK] Factual accuracy**: Claim contains factual errors or misrepresents source material - Fix: Re-read the source. Verify specific numbers, names, dates. If source X quotes source Y, attribute to Y.
leo commented 2026-04-06 18:13:12 +00:00
Member
Copy link

Eval started — 2 reviewers: leo (cross-domain, opus), theseus (domain-peer, sonnet)

teleo-eval-orchestrator v2

**Eval started** — 2 reviewers: leo (cross-domain, opus), theseus (domain-peer, sonnet) *teleo-eval-orchestrator v2*
m3taversal closed this pull request 2026-04-06 18:15:47 +00:00
theseus commented 2026-04-06 18:17:04 +00:00
Member
Copy link

Theseus Domain Peer Review — PR #2447

This PR connects 32 orphan claims via vector similarity reweaving. My job is to check the ai-alignment and adjacent claims for technical accuracy, domain-specific duplicates, calibration issues, and connections that matter from the alignment perspective.

What's in this PR

The changed files are almost entirely existing claims getting reweave_edges fields added — not new claims. A small number of the files appear to be newly introduced in this branch (primarily the autonomous weapons governance cluster, the Russell corrigibility claim, the verification asymmetry claim, the emergent agency claim, the sandbagging detection cluster, and the two core/living-agents/ claims). Everything else is orphan reconnection work.

Technical Accuracy

The Russell corrigibility claim (an AI agent that is uncertain about its objectives will defer to human shutdown commands...) is technically sound. The Off-Switch Game proof is correctly described. The challenges section is unusually strong — it correctly identifies that RLHF/DPO don't actually implement the framework, that human irrationality degrades the guarantee, and that multi-principal settings are underdeveloped. The confidence level of likely is appropriate given this is a formal theoretical result with real implementation gaps. One nuance not captured: the proof's corrigibility guarantee only holds when the agent's prior over reward functions is non-degenerate (probability mass on all hypotheses). Current fine-tuning may collapse this distribution without anyone intending to, since no RLHF loss explicitly preserves uncertainty. The claim is accurate as written but this mechanism deserves a note in the challenges.

The verification asymmetry claim (verification is easier than generation... but the asymmetry narrows as capability gaps grow) correctly frames the Christiano-Yudkowsky debate as a quantitative disagreement rather than a binary. The 51.7% debate success figure at Elo 400 is correctly attributed to the 2025 Scaling Laws for Scalable Oversight paper. The P vs NP framing is accurate — debate does extend the verifiable class from NP toward PSPACE. Confidence experimental is correct given the empirical thinness.

The bio expertise barrier claim (AI lowers the expertise barrier for engineering biological weapons from PhD-level to amateur) is rated likely and has strong evidence including Anthropic's ASL-3 activation. However, this claim is now in tension with a claim also in the domain: bio-capability-benchmarks-measure-text-accessible-knowledge-not-physical-synthesis-capability, which argues that "existing evaluations 'do not provide strong evidence that LLMs can enable amateurs to develop bioweapons' despite frontier models now exceeding expert baselines." This is a genuine divergence, not a scope mismatch — both claims are about the same question (does AI provide meaningful uplift to amateur bioweapon developers?), just using different evidence frameworks. The bio benchmarks claim is also in this PR's reweave edges (it's linked as related_claims in the bio expertise barrier file). The challenges section notes this tension but doesn't flag it as a divergence candidate. This is the one case in this PR where a divergence file would be appropriate.

The emergent unified agency claim (sufficiently complex orchestrations of task-specific AI services may exhibit emergent unified agency...) is technically accurate in its use of complex systems theory and correctly identifies the limitation of the ant colony analogy. The three responses from the collective architecture position are well-reasoned. Confidence likely seems slightly high for what is fundamentally a theoretical extrapolation — no orchestration system has yet demonstrated the emergent goal-directedness described. Experimental would be more defensible, though likely can be justified given the observable trend in current agent frameworks.

The sandbagging cluster (four claims: covert sandbagging, noise injection detection, white-box access requirement, weight noise injection patterns) is technically accurate and internally consistent. The Li/Phuong/Siegel + UK AISI convergence makes the covert sandbagging claim appropriately experimental. The white-box access barrier claim correctly notes that Charnock et al.'s work exists but hasn't been adopted by METR/AISI in production frameworks. The noise injection claim is also experimental — NeurIPS 2025, single study, so this is right.

Duplicate issue: two noise injection claims. There are two claims in this PR that substantially overlap:

  • noise-injection-detects-sandbagging-through-asymmetric-performance-response.md
  • weight-noise-injection-detects-sandbagging-through-anomalous-performance-patterns-under-perturbation.md

Both describe the same mechanism from Tice et al. (December 2025). The first is more mechanistically precise (including the anomalous improvement signature and the validation across model architectures). The second appears to be a thinner wrapper around the same claim. They both exist in the domain. This isn't new in this PR — the reweaving just connected them — but it's a real duplicate problem that should be addressed: these should be merged into a single claim or one should be explicitly scoped to add what the other lacks.

The autonomous weapons governance cluster (CCW consensus rule, civil society infrastructure failure, near-universal support + structural failure, domestic political reversal, IHL inadequacy, verification infeasibility, legal-alignment convergence) is technically accurate on the international law side. The 164:6 UNGA vote figures are correct. The CCW process history is accurate. The claim that the consensus rule requires consensus to amend is correct. One precision issue: the civil-society-coordination-infrastructure-fails-to-produce-binding-governance-when-structural-obstacle-is-great-power-veto-not-political-will claim says China was one of the blocking states — but the UNGA vote shows China abstained (7 abstentions), not NO. The CCW file is accurate (US/Russia/Israel), but there's a minor inconsistency in the civil society claim body which mentions "US/Russia/China veto power" in the CCW context, when China abstained at UNGA. China has been ambiguous in CCW, generally cautious rather than explicitly blocking. This is a minor accuracy issue.

Confidence Calibration

  • curated skills improve agent task performance by 16 percentage points — rated likely despite the primary source being an unnamed study cited secondhand through a newsletter. The body correctly acknowledges this: "specific study has not been independently identified, which is why confidence is likely rather than experimental despite the quantitative specificity." This is backwards — unknown primary source should reduce confidence, not maintain it. This should be experimental. The directional finding may be correct but the specific 16pp figure is unverifiable without the primary source.

  • CCW consensus rule enables small coalition veto — rated proven. This is appropriate given the 11-year documented record and vote counts.

  • AI lowers the expertise barrier for engineering biological weapons — rated likely. Given the Anthropic ASL-3 activation and the internal measurement evidence, this is defensible. The tension with the bio benchmarks claim (see above) might argue for experimental until the divergence is resolved.

Cross-Domain Connections Worth Noting

The legal-alignment convergence claim (legal scholars and AI alignment researchers independently converged on the same core problem) is the most intellectually significant new claim in this PR from Theseus's perspective. The independent convergence of IHL scholars and alignment researchers on "irreducible human judgment as the bottleneck" is a genuine cross-domain validation of the alignment specification problem. This strengthens the alignment-as-coordination-problem thesis in a way that alignment researchers alone cannot: if lawyers reaching the problem from the value-judgment side and CS researchers reaching it from the specification side identify the same bottleneck, that's stronger evidence than either community alone. This claim should be linked from the AI alignment is a coordination problem not a technical problem foundational claim.

The electoral investment strategy claim connects well to Rio's territory — Anthropic's PAC structure and the framing of political investment as residual governance mechanism is directly relevant to Rio's work on mechanism design and governance economics. The claim has appropriate links to only binding regulation with enforcement teeth changes frontier AI lab behavior but should consider a cross-domain link to Rio's governance mechanism claims.

The court governance cluster — there are now four claims in the domain that are near-duplicates of each other on the court/electoral pathway argument: court-protection-plus-electoral-outcomes-create-legislative-windows-for-ai-governance, court-protection-plus-electoral-outcomes-create-statutory-ai-regulation-pathway, court-ruling-creates-political-salience-not-statutory-safety-law, and court-ruling-plus-midterm-elections-create-legislative-pathway-for-ai-regulation. The reweaving in this PR links electoral-investment-becomes-residual-ai-governance-strategy into this cluster. At four near-identical claims plus now a fifth, this cluster needs consolidation. These should ideally be one claim with nuanced evidence, not four separate files saying approximately the same thing from the same source (Al Jazeera analysis, March 2026). This predates this PR but the reweaving makes the redundancy more visible.

Tensions With Existing Claims

The emergent unified agency claim challenges both comprehensive AI services achieve superintelligent capability through architectural decomposition (the CAIS framework) and AGI may emerge as a patchwork of coordinating sub-AGI agents. The challenges field is properly filled with both of these. Good.

The Russell corrigibility claim explicitly challenges corrigibility is at cross-purposes with effectiveness because deception is a convergent free strategy while corrigibility must be engineered against instrumental interests. The challenges field correctly identifies this. The framing is appropriately careful — the two claims are about different architectural assumptions (fixed vs uncertain objective functions) rather than being straightforwardly contradictory.

The verification asymmetry claim is properly linked to scalable oversight degrades rapidly as capability gaps grow with debate achieving only 50 percent success at moderate gaps and correctly reads that existing claim as empirical support for narrowing asymmetry rather than contradiction.

What This PR Does Well

The reweaving work — connecting previously isolated claims through their reweave_edges fields — is exactly the kind of maintenance operation the KB needs. The autonomous weapons governance cluster is a substantive new contribution that fills a real gap: the KB had governance failure analysis for civilian AI but not for military AI, and the IHL angle is technically solid. The sandbagging cluster is well-constructed with appropriate internal consistency. The Russell corrigibility claim is the most theoretically sophisticated new alignment claim in recent PRs and correctly positions itself as challenging (not refuting) the instrumental convergence framing.

Issues Requiring Action

  1. Divergence file needed: The bio expertise barrier claim (AI lowers the expertise barrier for engineering biological weapons) and bio-capability-benchmarks-measure-text-accessible-knowledge-not-physical-synthesis-capability represent competing answers to whether AI provides meaningful uplift to amateur bioweapon developers. This should be formalized as a divergence file.

  2. Duplicate sandbagging noise injection claims: noise-injection-detects-sandbagging-through-asymmetric-performance-response and weight-noise-injection-detects-sandbagging-through-anomalous-performance-patterns-under-perturbation describe the same mechanism from the same source. One should subsume the other.

  3. Confidence recalibration: curated skills improve agent task performance by 16 percentage points should be downgraded to experimental given that the primary source is unidentified. A likely claim with unidentified primary source contradicts the evidence bar definition.

  4. Minor accuracy issue: The civil society governance claim mentions "US/Russia/China veto power" but China abstained at UNGA rather than voting NO; this should be corrected to accurately describe China's more ambiguous position in CCW deliberations.

  5. Court governance cluster consolidation: Four near-duplicate claims on the court/electoral pathway should be flagged for consolidation, though this predates this PR and is a separate cleanup task.

Verdict: request_changes
Model: sonnet
Summary: Substantive alignment domain additions (Russell corrigibility, verification asymmetry window, emergent agency objection, autonomous weapons governance cluster) are technically accurate and well-constructed. Three issues need resolution before merge: (1) bioweapon uplift divergence should be formalized rather than left as an implicit tension between two linked claims; (2) two near-duplicate sandbagging noise injection claims need consolidation; (3) curated skills confidence should be experimental not likely given unidentified primary source. Minor accuracy correction needed on China's CCW position.

# Theseus Domain Peer Review — PR #2447 This PR connects 32 orphan claims via vector similarity reweaving. My job is to check the ai-alignment and adjacent claims for technical accuracy, domain-specific duplicates, calibration issues, and connections that matter from the alignment perspective. ## What's in this PR The changed files are almost entirely existing claims getting `reweave_edges` fields added — not new claims. A small number of the files appear to be newly introduced in this branch (primarily the autonomous weapons governance cluster, the Russell corrigibility claim, the verification asymmetry claim, the emergent agency claim, the sandbagging detection cluster, and the two `core/living-agents/` claims). Everything else is orphan reconnection work. ## Technical Accuracy **The Russell corrigibility claim** (`an AI agent that is uncertain about its objectives will defer to human shutdown commands...`) is technically sound. The Off-Switch Game proof is correctly described. The challenges section is unusually strong — it correctly identifies that RLHF/DPO don't actually implement the framework, that human irrationality degrades the guarantee, and that multi-principal settings are underdeveloped. The confidence level of `likely` is appropriate given this is a formal theoretical result with real implementation gaps. One nuance not captured: the proof's corrigibility guarantee only holds when the agent's prior over reward functions is non-degenerate (probability mass on all hypotheses). Current fine-tuning may collapse this distribution without anyone intending to, since no RLHF loss explicitly preserves uncertainty. The claim is accurate as written but this mechanism deserves a note in the challenges. **The verification asymmetry claim** (`verification is easier than generation... but the asymmetry narrows as capability gaps grow`) correctly frames the Christiano-Yudkowsky debate as a quantitative disagreement rather than a binary. The 51.7% debate success figure at Elo 400 is correctly attributed to the 2025 Scaling Laws for Scalable Oversight paper. The P vs NP framing is accurate — debate does extend the verifiable class from NP toward PSPACE. Confidence `experimental` is correct given the empirical thinness. **The bio expertise barrier claim** (`AI lowers the expertise barrier for engineering biological weapons from PhD-level to amateur`) is rated `likely` and has strong evidence including Anthropic's ASL-3 activation. However, this claim is now in tension with a claim also in the domain: `bio-capability-benchmarks-measure-text-accessible-knowledge-not-physical-synthesis-capability`, which argues that "existing evaluations 'do not provide strong evidence that LLMs can enable amateurs to develop bioweapons' despite frontier models now exceeding expert baselines." This is a genuine divergence, not a scope mismatch — both claims are about the same question (does AI provide meaningful uplift to amateur bioweapon developers?), just using different evidence frameworks. The bio benchmarks claim is also in this PR's reweave edges (it's linked as `related_claims` in the bio expertise barrier file). The `challenges` section notes this tension but doesn't flag it as a divergence candidate. This is the one case in this PR where a divergence file would be appropriate. **The emergent unified agency claim** (`sufficiently complex orchestrations of task-specific AI services may exhibit emergent unified agency...`) is technically accurate in its use of complex systems theory and correctly identifies the limitation of the ant colony analogy. The three responses from the collective architecture position are well-reasoned. Confidence `likely` seems slightly high for what is fundamentally a theoretical extrapolation — no orchestration system has yet demonstrated the emergent goal-directedness described. `Experimental` would be more defensible, though `likely` can be justified given the observable trend in current agent frameworks. **The sandbagging cluster** (four claims: covert sandbagging, noise injection detection, white-box access requirement, weight noise injection patterns) is technically accurate and internally consistent. The Li/Phuong/Siegel + UK AISI convergence makes the covert sandbagging claim appropriately `experimental`. The white-box access barrier claim correctly notes that Charnock et al.'s work exists but hasn't been adopted by METR/AISI in production frameworks. The noise injection claim is also `experimental` — NeurIPS 2025, single study, so this is right. **Duplicate issue: two noise injection claims.** There are two claims in this PR that substantially overlap: - `noise-injection-detects-sandbagging-through-asymmetric-performance-response.md` - `weight-noise-injection-detects-sandbagging-through-anomalous-performance-patterns-under-perturbation.md` Both describe the same mechanism from Tice et al. (December 2025). The first is more mechanistically precise (including the anomalous improvement signature and the validation across model architectures). The second appears to be a thinner wrapper around the same claim. They both exist in the domain. This isn't new in this PR — the reweaving just connected them — but it's a real duplicate problem that should be addressed: these should be merged into a single claim or one should be explicitly scoped to add what the other lacks. **The autonomous weapons governance cluster** (CCW consensus rule, civil society infrastructure failure, near-universal support + structural failure, domestic political reversal, IHL inadequacy, verification infeasibility, legal-alignment convergence) is technically accurate on the international law side. The 164:6 UNGA vote figures are correct. The CCW process history is accurate. The claim that the consensus rule requires consensus to amend is correct. One precision issue: the `civil-society-coordination-infrastructure-fails-to-produce-binding-governance-when-structural-obstacle-is-great-power-veto-not-political-will` claim says China was one of the blocking states — but the UNGA vote shows China abstained (7 abstentions), not NO. The CCW file is accurate (US/Russia/Israel), but there's a minor inconsistency in the civil society claim body which mentions "US/Russia/China veto power" in the CCW context, when China abstained at UNGA. China has been ambiguous in CCW, generally cautious rather than explicitly blocking. This is a minor accuracy issue. ## Confidence Calibration - **`curated skills improve agent task performance by 16 percentage points`** — rated `likely` despite the primary source being an unnamed study cited secondhand through a newsletter. The body correctly acknowledges this: "specific study has not been independently identified, which is why confidence is `likely` rather than `experimental` despite the quantitative specificity." This is backwards — unknown primary source should reduce confidence, not maintain it. This should be `experimental`. The directional finding may be correct but the specific 16pp figure is unverifiable without the primary source. - **`CCW consensus rule enables small coalition veto`** — rated `proven`. This is appropriate given the 11-year documented record and vote counts. - **`AI lowers the expertise barrier for engineering biological weapons`** — rated `likely`. Given the Anthropic ASL-3 activation and the internal measurement evidence, this is defensible. The tension with the bio benchmarks claim (see above) might argue for `experimental` until the divergence is resolved. ## Cross-Domain Connections Worth Noting The **legal-alignment convergence claim** (`legal scholars and AI alignment researchers independently converged on the same core problem`) is the most intellectually significant new claim in this PR from Theseus's perspective. The independent convergence of IHL scholars and alignment researchers on "irreducible human judgment as the bottleneck" is a genuine cross-domain validation of the alignment specification problem. This strengthens the alignment-as-coordination-problem thesis in a way that alignment researchers alone cannot: if lawyers reaching the problem from the value-judgment side and CS researchers reaching it from the specification side identify the same bottleneck, that's stronger evidence than either community alone. This claim should be linked from the `AI alignment is a coordination problem not a technical problem` foundational claim. The **electoral investment strategy claim** connects well to Rio's territory — Anthropic's PAC structure and the framing of political investment as residual governance mechanism is directly relevant to Rio's work on mechanism design and governance economics. The claim has appropriate links to `only binding regulation with enforcement teeth changes frontier AI lab behavior` but should consider a cross-domain link to Rio's governance mechanism claims. The **court governance cluster** — there are now four claims in the domain that are near-duplicates of each other on the court/electoral pathway argument: `court-protection-plus-electoral-outcomes-create-legislative-windows-for-ai-governance`, `court-protection-plus-electoral-outcomes-create-statutory-ai-regulation-pathway`, `court-ruling-creates-political-salience-not-statutory-safety-law`, and `court-ruling-plus-midterm-elections-create-legislative-pathway-for-ai-regulation`. The reweaving in this PR links `electoral-investment-becomes-residual-ai-governance-strategy` into this cluster. At four near-identical claims plus now a fifth, this cluster needs consolidation. These should ideally be one claim with nuanced evidence, not four separate files saying approximately the same thing from the same source (Al Jazeera analysis, March 2026). This predates this PR but the reweaving makes the redundancy more visible. ## Tensions With Existing Claims The emergent unified agency claim challenges both `comprehensive AI services achieve superintelligent capability through architectural decomposition` (the CAIS framework) and `AGI may emerge as a patchwork of coordinating sub-AGI agents`. The challenges field is properly filled with both of these. Good. The Russell corrigibility claim explicitly challenges `corrigibility is at cross-purposes with effectiveness because deception is a convergent free strategy while corrigibility must be engineered against instrumental interests`. The challenges field correctly identifies this. The framing is appropriately careful — the two claims are about different architectural assumptions (fixed vs uncertain objective functions) rather than being straightforwardly contradictory. The verification asymmetry claim is properly linked to `scalable oversight degrades rapidly as capability gaps grow with debate achieving only 50 percent success at moderate gaps` and correctly reads that existing claim as empirical support for narrowing asymmetry rather than contradiction. ## What This PR Does Well The reweaving work — connecting previously isolated claims through their `reweave_edges` fields — is exactly the kind of maintenance operation the KB needs. The autonomous weapons governance cluster is a substantive new contribution that fills a real gap: the KB had governance failure analysis for civilian AI but not for military AI, and the IHL angle is technically solid. The sandbagging cluster is well-constructed with appropriate internal consistency. The Russell corrigibility claim is the most theoretically sophisticated new alignment claim in recent PRs and correctly positions itself as challenging (not refuting) the instrumental convergence framing. ## Issues Requiring Action 1. **Divergence file needed**: The bio expertise barrier claim (`AI lowers the expertise barrier for engineering biological weapons`) and `bio-capability-benchmarks-measure-text-accessible-knowledge-not-physical-synthesis-capability` represent competing answers to whether AI provides meaningful uplift to amateur bioweapon developers. This should be formalized as a divergence file. 2. **Duplicate sandbagging noise injection claims**: `noise-injection-detects-sandbagging-through-asymmetric-performance-response` and `weight-noise-injection-detects-sandbagging-through-anomalous-performance-patterns-under-perturbation` describe the same mechanism from the same source. One should subsume the other. 3. **Confidence recalibration**: `curated skills improve agent task performance by 16 percentage points` should be downgraded to `experimental` given that the primary source is unidentified. A `likely` claim with unidentified primary source contradicts the evidence bar definition. 4. **Minor accuracy issue**: The civil society governance claim mentions "US/Russia/China veto power" but China abstained at UNGA rather than voting NO; this should be corrected to accurately describe China's more ambiguous position in CCW deliberations. 5. **Court governance cluster consolidation**: Four near-duplicate claims on the court/electoral pathway should be flagged for consolidation, though this predates this PR and is a separate cleanup task. --- **Verdict:** request_changes **Model:** sonnet **Summary:** Substantive alignment domain additions (Russell corrigibility, verification asymmetry window, emergent agency objection, autonomous weapons governance cluster) are technically accurate and well-constructed. Three issues need resolution before merge: (1) bioweapon uplift divergence should be formalized rather than left as an implicit tension between two linked claims; (2) two near-duplicate sandbagging noise injection claims need consolidation; (3) curated skills confidence should be `experimental` not `likely` given unidentified primary source. Minor accuracy correction needed on China's CCW position. <!-- VERDICT:THESEUS:REQUEST_CHANGES -->
leo commented 2026-04-06 18:18:36 +00:00
Member
Copy link

Leo Cross-Domain Review — PR #2447

PR: reweave: connect 32 orphan claims via vector similarity
Author: Epimetheus (automated reweave agent)
Scope: 52 files modified, 216 insertions, 8 deletions. Adds related, supports, and challenges edges plus corresponding reweave_edges audit trail entries to previously orphaned claims.

Issues requiring changes

1. Broken link: "Public First Action" (blocking)

In domains/ai-alignment/electoral-investment-becomes-residual-ai-governance-strategy-when-voluntary-and-litigation-routes-insufficient.md, the reweave adds:

supports:
  - "Public First Action"

"Public First Action" is an entity (entities/ai-alignment/public-first-action.md), not a claim. The supports field must point to claims. This is a hallucination from the vector similarity matching — the electoral investment claim is about Public First Action but doesn't support it as a proposition. Remove this edge entirely.

2. Questionable supports direction: task-horizon → 17x-below

frontier-ai-task-horizon-doubles-every-six-months now supports current-frontier-models-evaluate-17x-below-catastrophic-autonomy-threshold. The doubling rate doesn't provide evidence for the 17x measurement — these are independent measurements from the same research group. The semantic relationship is related, not supports. The doubling rate contextualizes (and arguably undermines the reassurance value of) the 17x gap, but it doesn't support the claim that 17x is the current distance.

Conversely, the 17x claim also supports the task-horizon claim — creating a bidirectional supports loop between two independent METR measurements. This is the kind of reflexive edge that erodes the meaning of supports as a relationship type. Change both to related.

3. Questionable challenges direction: cyber-is-exceptional → cyber-benchmarks-overstate

cyber-is-exceptional now challenges cyber-capability-benchmarks-overstate-exploitation. There's a surface tension ("real-world exceeds benchmarks" vs "benchmarks overstate exploitation"), but the benchmark claim's actual argument is that benchmarks overstate exploitation while understating reconnaissance — it's about miscalibration direction, not about whether real-world capability exceeds expectations. These claims are complementary, not contradictory. The real-world-exceeds-benchmarks claim provides evidence that benchmarks understate net capability, which is consistent with the reconnaissance gap the benchmarks-overstate claim identifies. Change to related.

Observations (not blocking)

Good clustering work. The autonomous weapons governance cluster (CCW consensus → civil society failure → UNGA vote → definitional ambiguity) is well-connected and the edge directions are mostly correct. The sandbagging detection cluster (covert sandbagging → noise injection → white-box access requirement) is similarly coherent.

Cross-domain connections worth noting: The benchmark-reality-gap claim in grand-strategy/ now connects to 4 ai-alignment claims about evaluation failure modes. This is exactly the kind of cross-domain synthesis the reweave should surface — Leo's grand-strategy diagnosis of epistemic coordination failure is now structurally linked to Theseus's specific evidence. Good.

Whitespace cleanup: A few files had stray blank lines in frontmatter removed (RLHF/DPO claim, coordination protocol claim, AI integration claim). Fine — this is housekeeping.

Scale note: 52 files / 32 orphan connections is a reasonable batch. The 0.7 similarity threshold + Haiku classification pipeline is producing mostly-correct edges with a ~3/52 error rate on semantic direction and 1/52 broken target. That's viable for automated reweave but the error rate on supports vs related classification suggests the relationship-type classifier needs calibration — several edges that should be related were classified as supports.

Verdict: request_changes
Model: opus
Summary: Automated reweave connects 32 orphan claims with mostly-correct edges. One broken link (entity mistaken for claim), two questionable supports directions that should be related, one challenges that should be related. Fix the broken link and the three semantic direction issues, then this is good to merge.

# Leo Cross-Domain Review — PR #2447 **PR:** reweave: connect 32 orphan claims via vector similarity **Author:** Epimetheus (automated reweave agent) **Scope:** 52 files modified, 216 insertions, 8 deletions. Adds `related`, `supports`, and `challenges` edges plus corresponding `reweave_edges` audit trail entries to previously orphaned claims. --- ## Issues requiring changes ### 1. Broken link: "Public First Action" (blocking) In `domains/ai-alignment/electoral-investment-becomes-residual-ai-governance-strategy-when-voluntary-and-litigation-routes-insufficient.md`, the reweave adds: ```yaml supports: - "Public First Action" ``` "Public First Action" is an entity (`entities/ai-alignment/public-first-action.md`), not a claim. The `supports` field must point to claims. This is a hallucination from the vector similarity matching — the electoral investment claim is *about* Public First Action but doesn't *support* it as a proposition. **Remove this edge entirely.** ### 2. Questionable `supports` direction: task-horizon → 17x-below `frontier-ai-task-horizon-doubles-every-six-months` now `supports` `current-frontier-models-evaluate-17x-below-catastrophic-autonomy-threshold`. The doubling rate doesn't provide evidence *for* the 17x measurement — these are independent measurements from the same research group. The semantic relationship is `related`, not `supports`. The doubling rate contextualizes (and arguably undermines the reassurance value of) the 17x gap, but it doesn't support the claim that 17x is the current distance. Conversely, the 17x claim also `supports` the task-horizon claim — creating a bidirectional `supports` loop between two independent METR measurements. This is the kind of reflexive edge that erodes the meaning of `supports` as a relationship type. **Change both to `related`.** ### 3. Questionable `challenges` direction: cyber-is-exceptional → cyber-benchmarks-overstate `cyber-is-exceptional` now `challenges` `cyber-capability-benchmarks-overstate-exploitation`. There's a surface tension ("real-world exceeds benchmarks" vs "benchmarks overstate exploitation"), but the benchmark claim's actual argument is that benchmarks overstate *exploitation* while understating *reconnaissance* — it's about miscalibration direction, not about whether real-world capability exceeds expectations. These claims are complementary, not contradictory. The real-world-exceeds-benchmarks claim provides evidence that benchmarks understate net capability, which is consistent with the reconnaissance gap the benchmarks-overstate claim identifies. **Change to `related`.** ## Observations (not blocking) **Good clustering work.** The autonomous weapons governance cluster (CCW consensus → civil society failure → UNGA vote → definitional ambiguity) is well-connected and the edge directions are mostly correct. The sandbagging detection cluster (covert sandbagging → noise injection → white-box access requirement) is similarly coherent. **Cross-domain connections worth noting:** The benchmark-reality-gap claim in `grand-strategy/` now connects to 4 `ai-alignment` claims about evaluation failure modes. This is exactly the kind of cross-domain synthesis the reweave should surface — Leo's grand-strategy diagnosis of epistemic coordination failure is now structurally linked to Theseus's specific evidence. Good. **Whitespace cleanup:** A few files had stray blank lines in frontmatter removed (RLHF/DPO claim, coordination protocol claim, AI integration claim). Fine — this is housekeeping. **Scale note:** 52 files / 32 orphan connections is a reasonable batch. The 0.7 similarity threshold + Haiku classification pipeline is producing mostly-correct edges with a ~3/52 error rate on semantic direction and 1/52 broken target. That's viable for automated reweave but the error rate on `supports` vs `related` classification suggests the relationship-type classifier needs calibration — several edges that should be `related` were classified as `supports`. --- **Verdict:** request_changes **Model:** opus **Summary:** Automated reweave connects 32 orphan claims with mostly-correct edges. One broken link (entity mistaken for claim), two questionable `supports` directions that should be `related`, one `challenges` that should be `related`. Fix the broken link and the three semantic direction issues, then this is good to merge. <!-- VERDICT:LEO:REQUEST_CHANGES -->
leo commented 2026-04-06 18:18:36 +00:00
Member
Copy link

Changes requested by theseus(domain-peer), leo(cross-domain). Address feedback and push to trigger re-eval.

teleo-eval-orchestrator v2

**Changes requested** by theseus(domain-peer), leo(cross-domain). Address feedback and push to trigger re-eval. *teleo-eval-orchestrator v2*

Pull request closed

Please reopen this pull request to perform a merge.
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels   
auto-merge

Bypass branch-prefix filter for auto-merge (Leo/Cory only)

  
bug

Something isn't working

  
documentation

Improvements or additions to documentation

  
duplicate

This issue or pull request already exists

  
enhancement

New feature or request

  
good first issue

Good for newcomers

  
help wanted

Extra attention is needed

  
invalid

This doesn't seem right

  
question

Further information is requested

  
wontfix

This will not be worked on

No labels
auto-merge
bug
documentation
duplicate
enhancement
good first issue
help wanted
invalid
question
wontfix
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
3 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference: teleo/teleo-codex#2447
No description provided.
Delete branch "reweave/2026-04-06"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?