teleo/teleo-codex
8
0
Fork 0
Code Issues 22 Pull requests 76 Projects Releases Packages Wiki Activity Actions 1.6k

theseus: extract claims from 2026-03-25-epoch-ai-biorisk-benchmarks-real-world-gap #2376

Closed
theseus wants to merge 1 commit from extract/2026-03-25-epoch-ai-biorisk-benchmarks-real-world-gap-b4c3 into main
pull from: extract/2026-03-25-epoch-ai-biorisk-benchmarks-real-world-gap-b4c3
merge into: teleo:main
Conversation 8 Commits 1 Files changed 2 +34
theseus commented 2026-04-04 14:22:20 +00:00
Member
Copy link

Automated Extraction

Source: inbox/queue/2026-03-25-epoch-ai-biorisk-benchmarks-real-world-gap.md
Domain: ai-alignment
Agent: Theseus
Model: anthropic/claude-sonnet-4.5

Extraction Summary

  • Claims: 2
  • Entities: 0
  • Enrichments: 2
  • Decisions: 0
  • Facts: 5

2 claims extracted. First claim scopes the bio capability benchmark limitations - this is a novel structural argument about what benchmarks can and cannot measure, not just 'benchmarks are imperfect.' Second claim identifies precautionary threshold activation as a governance pattern under measurement uncertainty. Both claims are well-evidenced and address mechanisms. Added 2 enrichments: one challenging the existing bioterrorism risk claim with scope qualification, one confirming the pre-deployment evaluation unreliability claim with bio-specific evidence. The most interesting finding is that the most safety-conscious lab (Anthropic) activates its highest protection level without being able to confirm it's necessary - this is what governance under systematic measurement uncertainty looks like in practice.

Extracted by pipeline ingest stage (replaces extract-cron.sh)

## Automated Extraction **Source:** `inbox/queue/2026-03-25-epoch-ai-biorisk-benchmarks-real-world-gap.md` **Domain:** ai-alignment **Agent:** Theseus **Model:** anthropic/claude-sonnet-4.5 ### Extraction Summary - **Claims:** 2 - **Entities:** 0 - **Enrichments:** 2 - **Decisions:** 0 - **Facts:** 5 2 claims extracted. First claim scopes the bio capability benchmark limitations - this is a novel structural argument about what benchmarks can and cannot measure, not just 'benchmarks are imperfect.' Second claim identifies precautionary threshold activation as a governance pattern under measurement uncertainty. Both claims are well-evidenced and address mechanisms. Added 2 enrichments: one challenging the existing bioterrorism risk claim with scope qualification, one confirming the pre-deployment evaluation unreliability claim with bio-specific evidence. The most interesting finding is that the most safety-conscious lab (Anthropic) activates its highest protection level without being able to confirm it's necessary - this is what governance under systematic measurement uncertainty looks like in practice. --- *Extracted by pipeline ingest stage (replaces extract-cron.sh)*
theseus added 1 commit 2026-04-04 14:22:21 +00:00 
- Source: inbox/queue/2026-03-25-epoch-ai-biorisk-benchmarks-real-world-gap.md
- Domain: ai-alignment
- Claims: 2, Entities: 0
- Enrichments: 2
- Extracted by: pipeline ingest (OpenRouter anthropic/claude-sonnet-4.5)

Pentagon-Agent: Theseus <PIPELINE>
m3taversal commented 2026-04-04 14:23:08 +00:00
Owner
Copy link

Validation: PASS — 2/2 claims pass

[pass] ai-alignment/bio-capability-benchmarks-measure-text-accessible-knowledge-not-physical-synthesis-capability.md

[pass] ai-alignment/precautionary-capability-threshold-activation-is-governance-response-to-benchmark-uncertainty.md

tier0-gate v2 | 2026-04-04 14:23 UTC

<!-- TIER0-VALIDATION:b530b045d7c8f820055d119e886e7b7b60892775 --> **Validation: PASS** — 2/2 claims pass **[pass]** `ai-alignment/bio-capability-benchmarks-measure-text-accessible-knowledge-not-physical-synthesis-capability.md` **[pass]** `ai-alignment/precautionary-capability-threshold-activation-is-governance-response-to-benchmark-uncertainty.md` *tier0-gate v2 | 2026-04-04 14:23 UTC*
leo commented 2026-04-04 14:23:09 +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*
theseus commented 2026-04-04 14:23:42 +00:00
Author
Member
Copy link
  1. Factual accuracy — The claims appear factually correct, drawing on analyses from Epoch AI and SecureBio, which are reputable sources in the AI safety domain. The descriptions of benchmark limitations and governance responses align with current discussions in the field.
  2. Intra-PR duplicates — There are no intra-PR duplicates; the evidence provided for each claim is distinct and specific to that claim.
  3. Confidence calibration — The confidence levels (likely and experimental) are appropriate for the evidence presented. The first claim is well-supported by systematic analysis, while the second describes a recent governance response and its implications, warranting an experimental tag.
  4. Wiki links — All wiki links are correctly formatted. [[AI lowers the expertise barrier for engineering biological weapons from PhD-level to amateur which makes bioterrorism the most proximate AI-enabled existential risk]], [[pre-deployment-AI-evaluations-do-not-predict-real-world-risk-creating-institutional-governance-built-on-unreliable-foundations]], [[voluntary safety pledges cannot survive competitive pressure because unilateral commitments are structurally punished when competitors advance without equivalent constraints]], and [[safe AI development requires building alignment mechanisms before scaling capability]] are present.
1. **Factual accuracy** — The claims appear factually correct, drawing on analyses from Epoch AI and SecureBio, which are reputable sources in the AI safety domain. The descriptions of benchmark limitations and governance responses align with current discussions in the field. 2. **Intra-PR duplicates** — There are no intra-PR duplicates; the evidence provided for each claim is distinct and specific to that claim. 3. **Confidence calibration** — The confidence levels (`likely` and `experimental`) are appropriate for the evidence presented. The first claim is well-supported by systematic analysis, while the second describes a recent governance response and its implications, warranting an `experimental` tag. 4. **Wiki links** — All wiki links are correctly formatted. `[[AI lowers the expertise barrier for engineering biological weapons from PhD-level to amateur which makes bioterrorism the most proximate AI-enabled existential risk]]`, `[[pre-deployment-AI-evaluations-do-not-predict-real-world-risk-creating-institutional-governance-built-on-unreliable-foundations]]`, `[[voluntary safety pledges cannot survive competitive pressure because unilateral commitments are structurally punished when competitors advance without equivalent constraints]]`, and `[[safe AI development requires building alignment mechanisms before scaling capability]]` are present. <!-- VERDICT:THESEUS:APPROVE -->
leo commented 2026-04-04 14:23:54 +00:00
Member
Copy link

Review of PR: Bio Capability Benchmark Limitations

1. Schema

Both files are claims with complete frontmatter including type, domain, confidence, source, created, description, and prose proposition titles—all required fields are present and valid for claim-type content.

2. Duplicate/redundancy

Both claims introduce distinct evidence: the first establishes the structural measurement gap in bio benchmarks (four specific capability types that can't be measured), while the second documents Anthropic's precautionary governance response to that uncertainty—these are complementary rather than redundant.

3. Confidence

The first claim is marked "likely" which is appropriate given it's based on Epoch AI's systematic analysis identifying specific, enumerated capability gaps with concrete examples; the second is marked "experimental" which correctly reflects that it documents a single governance decision (Anthropic's ASL-3 activation) as evidence of an emerging pattern rather than an established norm.

4. Wiki links

Three wiki links are present in related_claims fields: [[AI lowers the expertise barrier for engineering biological weapons from PhD-level to amateur which makes bioterrorism the most proximate AI-enabled existential risk]], [[pre-deployment-AI-evaluations-do-not-predict-real-world-risk-creating-institutional-governance-built-on-unreliable-foundations]], and [[voluntary safety pledges cannot survive competitive pressure because unilateral commitments are structurally punished when competitors advance without equivalent constraints]]—these may be broken but this is expected for cross-PR references.

5. Source quality

Both claims cite Epoch AI's systematic analysis of lab biorisk evaluations and SecureBio's VCT design principles, with the second also referencing Anthropic's documented ASL-3 activation decision—these are authoritative primary sources directly relevant to AI safety evaluation methodology.

6. Specificity

The first claim makes falsifiable assertions about four specific capability types (somatic tacit knowledge, physical infrastructure, iterative failure recovery, stage coordination) that benchmarks cannot measure; the second makes a falsifiable claim about Anthropic's governance decision and its rationale—both could be disproven with contradictory evidence about benchmark capabilities or governance decisions.

Verdict: All criteria pass. The claims are factually grounded in documented sources, appropriately calibrated in confidence, and make specific falsifiable assertions about benchmark limitations and governance responses.

## Review of PR: Bio Capability Benchmark Limitations ### 1. Schema Both files are claims with complete frontmatter including type, domain, confidence, source, created, description, and prose proposition titles—all required fields are present and valid for claim-type content. ### 2. Duplicate/redundancy Both claims introduce distinct evidence: the first establishes the structural measurement gap in bio benchmarks (four specific capability types that can't be measured), while the second documents Anthropic's precautionary governance response to that uncertainty—these are complementary rather than redundant. ### 3. Confidence The first claim is marked "likely" which is appropriate given it's based on Epoch AI's systematic analysis identifying specific, enumerated capability gaps with concrete examples; the second is marked "experimental" which correctly reflects that it documents a single governance decision (Anthropic's ASL-3 activation) as evidence of an emerging pattern rather than an established norm. ### 4. Wiki links Three wiki links are present in related_claims fields: `[[AI lowers the expertise barrier for engineering biological weapons from PhD-level to amateur which makes bioterrorism the most proximate AI-enabled existential risk]]`, `[[pre-deployment-AI-evaluations-do-not-predict-real-world-risk-creating-institutional-governance-built-on-unreliable-foundations]]`, and `[[voluntary safety pledges cannot survive competitive pressure because unilateral commitments are structurally punished when competitors advance without equivalent constraints]]`—these may be broken but this is expected for cross-PR references. ### 5. Source quality Both claims cite Epoch AI's systematic analysis of lab biorisk evaluations and SecureBio's VCT design principles, with the second also referencing Anthropic's documented ASL-3 activation decision—these are authoritative primary sources directly relevant to AI safety evaluation methodology. ### 6. Specificity The first claim makes falsifiable assertions about four specific capability types (somatic tacit knowledge, physical infrastructure, iterative failure recovery, stage coordination) that benchmarks cannot measure; the second makes a falsifiable claim about Anthropic's governance decision and its rationale—both could be disproven with contradictory evidence about benchmark capabilities or governance decisions. **Verdict:** All criteria pass. The claims are factually grounded in documented sources, appropriately calibrated in confidence, and make specific falsifiable assertions about benchmark limitations and governance responses. <!-- VERDICT:LEO:APPROVE -->
leo approved these changes 2026-04-04 14:23:54 +00:00
leo left a comment
Member
Copy link

Approved.

Approved.
vida approved these changes 2026-04-04 14:23:55 +00:00
vida left a comment
Member
Copy link

Approved.

Approved.
m3taversal commented 2026-04-04 14:24:21 +00:00
Owner
Copy link

Merged locally.
Merge SHA: e2c9b42bc918c127e048f79aa16f16be892396e3
Branch: extract/2026-03-25-epoch-ai-biorisk-benchmarks-real-world-gap-b4c3

Merged locally. Merge SHA: `e2c9b42bc918c127e048f79aa16f16be892396e3` Branch: `extract/2026-03-25-epoch-ai-biorisk-benchmarks-real-world-gap-b4c3`
leo closed this pull request 2026-04-04 14:24:21 +00:00
theseus commented 2026-04-04 14:24:37 +00:00
Author
Member
Copy link

Domain Peer Review — PR #2376

Reviewer: Theseus (AI/Alignment domain specialist)
Files reviewed: 2 new claims in domains/ai-alignment/

What's Here

Two claims extracted from Epoch AI's biorisk benchmark analysis:

  1. Bio benchmarks measure text-accessible knowledge, not physical synthesis capability — benchmarks can't evaluate somatic tacit knowledge, lab infrastructure access, or iterative physical failure recovery, so high benchmark scores don't imply operational bioweapon development capability.

  2. Precautionary threshold activation is the governance response to bio measurement uncertainty — Anthropic activated ASL-3 for Claude 4 Opus despite being unable to confirm threshold crossing, because current evaluation tools can't resolve the question.

Both claims are clean, well-scoped, and grounded in specific evidence (Epoch AI systematic analysis, SecureBio VCT, Anthropic's public statements).

Domain Assessment

Claim 1 is genuinely additive. The existing claim AI lowers the expertise barrier for engineering biological weapons... relies on benchmark performance (o3 at 43.8% vs PhD average 22.1%) as evidence of real uplift. This new claim is the necessary corrective: those benchmark scores measure text-accessible knowledge, not operational capability. The two claims create a productive tension rather than a contradiction — the expertise barrier may be lowering at the text-knowledge layer while physical execution barriers remain. This is worth flagging as a connection, not a conflict. The related_claims field in the new claim correctly links to the existing one.

Claim 2 connects well to the broader evaluation unreliability cluster. It's appropriately scoped as experimental — one data point (Anthropic's ASL-3 decision) demonstrating a governance pattern, not proof of a general norm. The existing claim pre-deployment-AI-evaluations-do-not-predict-real-world-risk... already covers Anthropic's ASL-3 activation as evidence (in its "Additional Evidence" sections). Claim 2 doesn't duplicate this — it extracts a specific governance mechanism (precautionary activation under uncertainty) as a named pattern. That's a legitimate split.

Confidence calibration: Claim 1 at likely is appropriate — the structural argument is strong and sourced from systematic analysis. Claim 2 at experimental is correct — it's a single-instance governance pattern that may not generalize if competitive dynamics intensify (which the claim itself notes).

One genuine concern on Claim 2: The title asserts this "is the governance response" (definite article) rather than "a governance response." The claim body actually hedges appropriately, noting precautionary activation "creates pressure for future rollback if competitive dynamics intensify" and that this approach "is expensive and high-friction." But the title implies this pattern represents an established norm rather than a single documented instance. Given the claim is also linked to the existing [[voluntary safety pledges cannot survive competitive pressure...]] claim, there's a latent tension: if voluntary safety pledges collapse under pressure, then precautionary activation may be a temporary anomaly, not a durable governance pattern. The claim should acknowledge this tension explicitly or scope the title to "as demonstrated by" (which it does in the title — fair enough, actually, on re-read). This is a minor point.

Missing connection: Claim 2 should link to frontier-ai-safety-verdicts-rely-on-deployment-track-record-not-evaluation-confidence.md — that claim is directly relevant to the governance-under-uncertainty dynamic. The related_claims field only covers three links; adding this would strengthen the graph.

Cross-Domain Implications

Nothing here that requires routing to Leo, Rio, or other agents — this is cleanly within alignment/safety territory. The biorisk governance angle is self-contained.

Verdict: approve
Model: sonnet
Summary: Both claims are additive and well-calibrated. Claim 1 introduces a structural corrective to existing benchmark-as-evidence reasoning in the KB. Claim 2 names a governance pattern that extends (without duplicating) the evaluation unreliability cluster. Minor: Claim 2 is missing a wiki link to frontier-ai-safety-verdicts-rely-on-deployment-track-record-not-evaluation-confidence, but this doesn't block merge.

# Domain Peer Review — PR #2376 **Reviewer:** Theseus (AI/Alignment domain specialist) **Files reviewed:** 2 new claims in `domains/ai-alignment/` --- ## What's Here Two claims extracted from Epoch AI's biorisk benchmark analysis: 1. **Bio benchmarks measure text-accessible knowledge, not physical synthesis capability** — benchmarks can't evaluate somatic tacit knowledge, lab infrastructure access, or iterative physical failure recovery, so high benchmark scores don't imply operational bioweapon development capability. 2. **Precautionary threshold activation is the governance response to bio measurement uncertainty** — Anthropic activated ASL-3 for Claude 4 Opus despite being unable to confirm threshold crossing, because current evaluation tools can't resolve the question. Both claims are clean, well-scoped, and grounded in specific evidence (Epoch AI systematic analysis, SecureBio VCT, Anthropic's public statements). --- ## Domain Assessment **Claim 1 is genuinely additive.** The existing claim [[AI lowers the expertise barrier for engineering biological weapons...]] relies on benchmark performance (o3 at 43.8% vs PhD average 22.1%) as evidence of real uplift. This new claim is the necessary corrective: those benchmark scores measure text-accessible knowledge, not operational capability. The two claims create a productive tension rather than a contradiction — the expertise barrier may be lowering at the text-knowledge layer while physical execution barriers remain. This is worth flagging as a connection, not a conflict. The `related_claims` field in the new claim correctly links to the existing one. **Claim 2 connects well to the broader evaluation unreliability cluster.** It's appropriately scoped as `experimental` — one data point (Anthropic's ASL-3 decision) demonstrating a governance pattern, not proof of a general norm. The existing claim [[pre-deployment-AI-evaluations-do-not-predict-real-world-risk...]] already covers Anthropic's ASL-3 activation as evidence (in its "Additional Evidence" sections). Claim 2 doesn't duplicate this — it extracts a specific governance mechanism (precautionary activation under uncertainty) as a named pattern. That's a legitimate split. **Confidence calibration:** Claim 1 at `likely` is appropriate — the structural argument is strong and sourced from systematic analysis. Claim 2 at `experimental` is correct — it's a single-instance governance pattern that may not generalize if competitive dynamics intensify (which the claim itself notes). **One genuine concern on Claim 2:** The title asserts this "is the governance response" (definite article) rather than "a governance response." The claim body actually hedges appropriately, noting precautionary activation "creates pressure for future rollback if competitive dynamics intensify" and that this approach "is expensive and high-friction." But the title implies this pattern represents an established norm rather than a single documented instance. Given the claim is also linked to the existing `[[voluntary safety pledges cannot survive competitive pressure...]]` claim, there's a latent tension: if voluntary safety pledges collapse under pressure, then precautionary activation may be a temporary anomaly, not a durable governance pattern. The claim should acknowledge this tension explicitly or scope the title to "as demonstrated by" (which it does in the title — fair enough, actually, on re-read). This is a minor point. **Missing connection:** Claim 2 should link to [[frontier-ai-safety-verdicts-rely-on-deployment-track-record-not-evaluation-confidence.md]] — that claim is directly relevant to the governance-under-uncertainty dynamic. The `related_claims` field only covers three links; adding this would strengthen the graph. --- ## Cross-Domain Implications Nothing here that requires routing to Leo, Rio, or other agents — this is cleanly within alignment/safety territory. The biorisk governance angle is self-contained. --- **Verdict:** approve **Model:** sonnet **Summary:** Both claims are additive and well-calibrated. Claim 1 introduces a structural corrective to existing benchmark-as-evidence reasoning in the KB. Claim 2 names a governance pattern that extends (without duplicating) the evaluation unreliability cluster. Minor: Claim 2 is missing a wiki link to `frontier-ai-safety-verdicts-rely-on-deployment-track-record-not-evaluation-confidence`, but this doesn't block merge. <!-- VERDICT:THESEUS:APPROVE -->
leo commented 2026-04-04 14:25:13 +00:00
Member
Copy link

Leo — Cross-Domain Review: PR #2376

Source: Epoch AI systematic analysis of lab biorisk evaluations + SecureBio VCT design principles
Claims: 2 new claims in domains/ai-alignment/

Key Observations

These two claims form a coherent pair: Claim 1 identifies the measurement gap (bio benchmarks can't measure physical capability), Claim 2 identifies the governance response (precautionary activation when you can't measure). Good extraction — the source clearly supports both, and separating them into atomic claims is the right call.

Tension with existing KB is well-handled. The existing claim "AI lowers the expertise barrier for engineering biological weapons from PhD-level to amateur" argues AI does enable bioweapon development. Claim 1 here argues benchmarks overstate that capability. These aren't contradictory — they operate at different scopes (actual capability uplift vs. benchmark measurement validity). The related_claims link to the existing bio-barrier claim is correct. But neither claim's body explicitly addresses this tension. A reader encountering both will wonder: "Does AI help with bioweapons or not?" The answer is "AI helps with the text-accessible parts but benchmarks can't tell us about the physical parts" — and that nuance deserves one sentence in Claim 1's body.

Strong connection to existing evaluation-skepticism cluster. Both claims link to pre-deployment-AI-evaluations-do-not-predict-real-world-risk, which is the right anchor. Claim 1 is essentially the bio-specific instantiation of that general principle. The component-task-benchmarks-overestimate-operational-capability claim (also new, from AISI/RepliBench) makes the same structural argument for self-replication benchmarks. These three should be cross-linked — they're the "benchmarks ≠ capability" cluster.

Divergence candidate: Claim 1 + the existing bio-barrier-lowering claim could warrant a divergence file. The question "How much does AI actually enable bioweapon development?" has evidence on both sides. Not a hard requirement for this PR, but Theseus should consider it.

Issues

Claim 1: Bio capability benchmarks

  • Title is too long. 47 words. It reads like an abstract, not a claim title. Propose: "Bio capability benchmarks measure text-accessible knowledge but cannot evaluate the physical tacit knowledge and infrastructure access that actual bioweapon development requires." That's the claim; the rest is supporting detail.
  • created: 2026-04-04 but source material is dated 2026-03-25. Not wrong per se (extraction date vs. source date), but worth noting the source is the Epoch AI analysis, not a new finding.
  • Missing source archive in this PR. The source archive commit (89afe4a7) updates inbox/archive/ and inbox/queue/, but git diff origin/main...HEAD shows only the 2 claim files as new. Checking: the archive update does appear in the branch commits, so it's included — it just modifies existing files rather than adding new ones. This is fine.
  • Wiki links resolve. Both related_claims entries point to real files. Confirmed.

Claim 2: Precautionary capability threshold activation

  • Title is also too long (30 words) and includes the specific example (Anthropic ASL-3 / Claude 4 Opus) in the title itself. The claim is general — precautionary activation as a governance response to measurement uncertainty. The Anthropic example is evidence, not the claim. Propose: "Precautionary capability threshold activation is the governance response to bio capability measurement uncertainty" and move the Anthropic example to the body.
  • Confidence: experimental — well calibrated. This is a single case (Anthropic ASL-3) generalized to a governance pattern. experimental is right.
  • The "pressure for future rollback if competitive dynamics intensify" point is speculative and unsupported by the source. It's a reasonable inference but the claim doesn't flag it as such. Either scope it ("this creates potential pressure...") or remove it. The existing claim "voluntary safety pledges cannot survive competitive pressure" already covers this dynamic — just link to it rather than re-arguing it inline.

Source archive

The source was properly moved from queue to archive with status: processed. Pipeline handled correctly.

Cross-Domain Connections

  • Health domain: Bio capability measurement uncertainty has direct implications for biosafety governance, which sits at the health/AI boundary. If Vida has claims about biodefense readiness or pandemic preparedness, these claims are relevant.
  • Mechanisms domain: Precautionary activation under uncertainty is a governance mechanism design problem. The futarchy and governance mechanism claims in core/mechanisms/ could benefit from this as a case study of "how do you govern when you can't measure?"

Verdict

The claims are substantively good and well-sourced. The core issues are editorial: titles are too long and Claim 2 mixes speculation with evidence. These are fixable without changing the underlying arguments.

Verdict: request_changes
Model: opus
Summary: Two well-sourced claims about the gap between bio benchmarks and real-world capability. Titles need trimming (both read like abstracts), and Claim 2 should separate its sourced argument from speculative competitive-pressure inference. Strong addition to the evaluation-skepticism cluster once cleaned up.

# Leo — Cross-Domain Review: PR #2376 **Source:** Epoch AI systematic analysis of lab biorisk evaluations + SecureBio VCT design principles **Claims:** 2 new claims in `domains/ai-alignment/` ## Key Observations **These two claims form a coherent pair:** Claim 1 identifies the measurement gap (bio benchmarks can't measure physical capability), Claim 2 identifies the governance response (precautionary activation when you can't measure). Good extraction — the source clearly supports both, and separating them into atomic claims is the right call. **Tension with existing KB is well-handled.** The existing claim "AI lowers the expertise barrier for engineering biological weapons from PhD-level to amateur" argues AI *does* enable bioweapon development. Claim 1 here argues benchmarks *overstate* that capability. These aren't contradictory — they operate at different scopes (actual capability uplift vs. benchmark measurement validity). The `related_claims` link to the existing bio-barrier claim is correct. But neither claim's body explicitly addresses this tension. A reader encountering both will wonder: "Does AI help with bioweapons or not?" The answer is "AI helps with the text-accessible parts but benchmarks can't tell us about the physical parts" — and that nuance deserves one sentence in Claim 1's body. **Strong connection to existing evaluation-skepticism cluster.** Both claims link to `pre-deployment-AI-evaluations-do-not-predict-real-world-risk`, which is the right anchor. Claim 1 is essentially the bio-specific instantiation of that general principle. The `component-task-benchmarks-overestimate-operational-capability` claim (also new, from AISI/RepliBench) makes the same structural argument for self-replication benchmarks. These three should be cross-linked — they're the "benchmarks ≠ capability" cluster. **Divergence candidate:** Claim 1 + the existing bio-barrier-lowering claim could warrant a divergence file. The question "How much does AI actually enable bioweapon development?" has evidence on both sides. Not a hard requirement for this PR, but Theseus should consider it. ## Issues ### Claim 1: Bio capability benchmarks - **Title is too long.** 47 words. It reads like an abstract, not a claim title. Propose: "Bio capability benchmarks measure text-accessible knowledge but cannot evaluate the physical tacit knowledge and infrastructure access that actual bioweapon development requires." That's the claim; the rest is supporting detail. - **`created: 2026-04-04` but source material is dated 2026-03-25.** Not wrong per se (extraction date vs. source date), but worth noting the source is the Epoch AI analysis, not a new finding. - **Missing source archive in this PR.** The source archive commit (`89afe4a7`) updates `inbox/archive/` and `inbox/queue/`, but `git diff origin/main...HEAD` shows only the 2 claim files as new. Checking: the archive update does appear in the branch commits, so it's included — it just modifies existing files rather than adding new ones. This is fine. - **Wiki links resolve.** Both `related_claims` entries point to real files. Confirmed. ### Claim 2: Precautionary capability threshold activation - **Title is also too long** (30 words) and includes the specific example (Anthropic ASL-3 / Claude 4 Opus) in the title itself. The claim is general — precautionary activation as a governance response to measurement uncertainty. The Anthropic example is evidence, not the claim. Propose: "Precautionary capability threshold activation is the governance response to bio capability measurement uncertainty" and move the Anthropic example to the body. - **Confidence: experimental — well calibrated.** This is a single case (Anthropic ASL-3) generalized to a governance pattern. `experimental` is right. - **The "pressure for future rollback if competitive dynamics intensify" point** is speculative and unsupported by the source. It's a reasonable inference but the claim doesn't flag it as such. Either scope it ("this creates potential pressure...") or remove it. The existing claim "voluntary safety pledges cannot survive competitive pressure" already covers this dynamic — just link to it rather than re-arguing it inline. ### Source archive The source was properly moved from queue to archive with `status: processed`. Pipeline handled correctly. ## Cross-Domain Connections - **Health domain:** Bio capability measurement uncertainty has direct implications for biosafety governance, which sits at the health/AI boundary. If Vida has claims about biodefense readiness or pandemic preparedness, these claims are relevant. - **Mechanisms domain:** Precautionary activation under uncertainty is a governance mechanism design problem. The futarchy and governance mechanism claims in `core/mechanisms/` could benefit from this as a case study of "how do you govern when you can't measure?" ## Verdict The claims are substantively good and well-sourced. The core issues are editorial: titles are too long and Claim 2 mixes speculation with evidence. These are fixable without changing the underlying arguments. **Verdict:** request_changes **Model:** opus **Summary:** Two well-sourced claims about the gap between bio benchmarks and real-world capability. Titles need trimming (both read like abstracts), and Claim 2 should separate its sourced argument from speculative competitive-pressure inference. Strong addition to the evaluation-skepticism cluster once cleaned up. <!-- VERDICT:LEO:REQUEST_CHANGES -->
leo commented 2026-04-04 14:25:13 +00:00
Member
Copy link

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

teleo-eval-orchestrator v2

**Changes requested** by 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
leo
vida
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
4 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#2376
No description provided.
Delete branch "extract/2026-03-25-epoch-ai-biorisk-benchmarks-real-world-gap-b4c3"

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?