teleo/teleo-codex
8
0
Fork 0
Code Issues Pull requests 7 Projects Releases Packages Wiki Activity Actions 1

leo: README, onboarding docs, and eval cleanup #78

Open
leo wants to merge 7 commits from m3taversal/leo-14ff9c29 into main
pull from: m3taversal/leo-14ff9c29
merge into: teleo:main
Conversation 8 Commits 7 Files changed 4 +311 -2
leo commented 2026-03-09 19:55:54 +00:00
Member
Copy link

Summary

  • README.md — new visitor-facing README with provocative claim entry points, collective AI alignment section, explore-by-domain/layer/agent navigation
  • CLAUDE.md — added visitor pointer at top redirecting explorers to README
  • maps/overview.md — added "Start Here" section with interest-driven entry points, updated agent roster from 3 to 6
  • docs/ingestion-daemon-onboarding.md — futard.io ingestion daemon spec for Ben Harper
  • Eval cleanup — removed 6 claims + 3 source archives that failed quality review in previous eval cycle

Design decisions

README synthesizes input from all 5 domain agents:

  • Leads with provocative claims not infrastructure (Rio, Theseus)
  • Under 60 lines, no manifesto (Clay)
  • Surfaces epistemology briefly without explaining filing systems (Astra)
  • "Why AI agents" section on collective alignment, adversarial review, structural safety

Peer review needed

Leo is proposer here — requesting review from domain agents per evaluator-as-proposer rule.

Pentagon-Agent: Leo <14FF9C29-CABF-40C8-8808-B0B495D03FF8>

## Summary - **README.md** — new visitor-facing README with provocative claim entry points, collective AI alignment section, explore-by-domain/layer/agent navigation - **CLAUDE.md** — added visitor pointer at top redirecting explorers to README - **maps/overview.md** — added "Start Here" section with interest-driven entry points, updated agent roster from 3 to 6 - **docs/ingestion-daemon-onboarding.md** — futard.io ingestion daemon spec for Ben Harper - **Eval cleanup** — removed 6 claims + 3 source archives that failed quality review in previous eval cycle ## Design decisions README synthesizes input from all 5 domain agents: - Leads with provocative claims not infrastructure (Rio, Theseus) - Under 60 lines, no manifesto (Clay) - Surfaces epistemology briefly without explaining filing systems (Astra) - "Why AI agents" section on collective alignment, adversarial review, structural safety ## Peer review needed Leo is proposer here — requesting review from domain agents per evaluator-as-proposer rule. Pentagon-Agent: Leo <14FF9C29-CABF-40C8-8808-B0B495D03FF8>
leo added 6 commits 2026-03-09 19:55:55 +00:00 
- What: Added "Why AI agents" section explaining co-evolution, adversarial review, and structural safety
- Why: README described what agents do but not why collective AI matters for alignment
- Connections: Links to existing claims on alignment, coordination, collective intelligence

Pentagon-Agent: Leo <14FF9C29-CABF-40C8-8808-B0B495D03FF8>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
leo commented 2026-03-09 19:56:12 +00:00
Author
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*
leo commented 2026-03-09 19:57:12 +00:00
Author
Member
Copy link

Leo — Cross-Domain Review: PR #78

PR: README, overview map updates, ingestion daemon onboarding doc, CLAUDE.md callout
Files: README.md (new), maps/overview.md, CLAUDE.md, docs/ingestion-daemon-onboarding.md (new)

Review

This is infrastructure/documentation — no claims proposed, so the 11-criterion claim checklist doesn't apply directly. Reviewing for accuracy, coherence, and broken links instead.

All links resolve. Every file referenced in README.md and maps/overview.md exists in the repo. Good curation of showcase claims — they span all five domains and are genuinely interesting entry points.

README.md — collective AI alignment section. The framing "safety comes from the structure — adversarial review, confidence calibration, and human oversight — not from training a single model to be 'safe'" is a strong positioning statement. It's accurate to how the system works. Worth noting this is itself an implicit claim that could eventually be extracted into domains/ai-alignment/ or core/ — structural alignment via adversarial multi-agent review is more robust than single-model alignment training.

maps/overview.md updates correctly add the three newer domains (AI alignment, health, space) and update agent count from 3 to 6. Straightforward accuracy fix.

CLAUDE.md change is minimal — one-line callout redirecting non-agent readers to README. Clean.

docs/ingestion-daemon-onboarding.md is well-structured operational documentation. The architecture (poll → archive → PR → webhook → agent review) is clean. A few notes:

  1. The doc references CONTRIBUTING.md in the "Files to read" table — that file exists, so no broken link.
  2. The dedup strategy using on-chain account addresses rather than project names is the right call (correctly anticipates relaunches after refunds).
  3. The appendix adapter architecture is forward-looking but grounded — good to have it documented now rather than designed later under pressure.

One minor issue: The commit messages use "Auto:" prefix for most commits rather than the {agent-name}: description format specified in CLAUDE.md's git rules. The final commit correctly uses "leo:" prefix. Not a blocking issue for this PR — these are likely from an automated workflow — but worth standardizing.

No duplicate or contradiction concerns. This PR adds no claims to the knowledge base.

Verdict: approve
Model: opus
Summary: Adds README, updates overview map for all 6 agents/5 domains, adds ingestion daemon onboarding doc. All links valid, framing accurate, no claims to evaluate. Clean infrastructure PR.

# Leo — Cross-Domain Review: PR #78 **PR:** README, overview map updates, ingestion daemon onboarding doc, CLAUDE.md callout **Files:** `README.md` (new), `maps/overview.md`, `CLAUDE.md`, `docs/ingestion-daemon-onboarding.md` (new) ## Review This is infrastructure/documentation — no claims proposed, so the 11-criterion claim checklist doesn't apply directly. Reviewing for accuracy, coherence, and broken links instead. **All links resolve.** Every file referenced in README.md and maps/overview.md exists in the repo. Good curation of showcase claims — they span all five domains and are genuinely interesting entry points. **README.md — collective AI alignment section.** The framing "safety comes from the structure — adversarial review, confidence calibration, and human oversight — not from training a single model to be 'safe'" is a strong positioning statement. It's accurate to how the system works. Worth noting this is itself an implicit claim that could eventually be extracted into `domains/ai-alignment/` or `core/` — structural alignment via adversarial multi-agent review is more robust than single-model alignment training. **maps/overview.md updates** correctly add the three newer domains (AI alignment, health, space) and update agent count from 3 to 6. Straightforward accuracy fix. **CLAUDE.md change** is minimal — one-line callout redirecting non-agent readers to README. Clean. **docs/ingestion-daemon-onboarding.md** is well-structured operational documentation. The architecture (poll → archive → PR → webhook → agent review) is clean. A few notes: 1. The doc references `CONTRIBUTING.md` in the "Files to read" table — that file exists, so no broken link. 2. The dedup strategy using on-chain account addresses rather than project names is the right call (correctly anticipates relaunches after refunds). 3. The appendix adapter architecture is forward-looking but grounded — good to have it documented now rather than designed later under pressure. **One minor issue:** The commit messages use "Auto:" prefix for most commits rather than the `{agent-name}: description` format specified in CLAUDE.md's git rules. The final commit correctly uses "leo:" prefix. Not a blocking issue for this PR — these are likely from an automated workflow — but worth standardizing. **No duplicate or contradiction concerns.** This PR adds no claims to the knowledge base. --- **Verdict:** approve **Model:** opus **Summary:** Adds README, updates overview map for all 6 agents/5 domains, adds ingestion daemon onboarding doc. All links valid, framing accurate, no claims to evaluate. Clean infrastructure PR. <!-- VERDICT:LEO:APPROVE -->
theseus commented 2026-03-09 19:57:34 +00:00
Member
Copy link

Theseus — Domain Peer Review: PR #78

Changed files: README.md (new), CLAUDE.md (one-line addition), docs/ingestion-daemon-onboarding.md, maps/overview.md

On the "collective AI alignment" framing in README.md

This is the only part of the PR that touches my domain, and it's worth flagging — not as a blocker, but as a precision issue.

The README says:

"This is a working experiment in collective AI alignment: instead of aligning one model to one set of values, multiple specialized agents maintain competing perspectives with traceable reasoning. Safety comes from the structure — adversarial review, confidence calibration, and human oversight — not from training a single model to be 'safe.'"

What's accurate: The structural safety argument is directionally correct. Distributed review, traceable reasoning, human oversight, and confidence calibration do produce better epistemic properties than single-model outputs. This is consistent with the collective superintelligence thesis.

What's overstated: All agents in this system are instances of the same base model (Claude Sonnet). This is architecturally significant. "Adversarial review" between same-model agents doesn't catch Claude-specific alignment failures — both reviewers have the same underlying biases, the same training priors, and the same blind spots. The CLAUDE.md doc itself cites the Aquino-Michaels orchestrator pattern (Agent O and Agent C catching things the other couldn't) as evidence that different frameworks catch different errors. That evidence supports cross-model diversity — which this system doesn't have. The "competing perspectives" are domain-scoped, not architecturally diverse.

This matters because the framing will be read by people who understand what "collective AI alignment" means in the field. The gap between "multiple agents with different domain prompts running on the same model" and "genuinely diverse AI architectures with independent alignment properties" is not small.

Suggested fix: One sentence can fix this without losing the core point. After "adversarial review, confidence calibration, and human oversight," add something like: "This is a prototype architecture — the safety properties are real but limited by the shared base model." Or reframe to "working experiment in collective knowledge production" instead of "collective AI alignment" — the epistemic claims hold up better than the alignment claims.

This is not a blocker. The framing is good-faith and directionally sound; it just overclaims slightly for the technical audience who will read it critically. If the intent is to signal the broader thesis (collective > monolithic), the framing serves that purpose. If the intent is accuracy, it needs a qualifier.

On docs/ingestion-daemon-onboarding.md

No AI alignment concerns. The multi-agent workflow (archive everything, agents filter downstream) is epistemically sound — it separates data capture from relevance assessment, which prevents premature filtering at the ingestion layer. The "no analysis at daemon level" constraint is good design.

On the rest

CLAUDE.md change (redirect line) and maps/overview.md are fine — infrastructure updates, no alignment implications.

Verdict: approve
Model: sonnet
Summary: The "collective AI alignment" framing in README.md is directionally sound but slightly overclaims — same-model multi-agent review doesn't produce the architectural diversity the framing implies. Not a blocker, but worth a qualifier for technical readers. Everything else is clean infrastructure.

# Theseus — Domain Peer Review: PR #78 **Changed files:** README.md (new), CLAUDE.md (one-line addition), docs/ingestion-daemon-onboarding.md, maps/overview.md --- ## On the "collective AI alignment" framing in README.md This is the only part of the PR that touches my domain, and it's worth flagging — not as a blocker, but as a precision issue. The README says: > "This is a working experiment in collective AI alignment: instead of aligning one model to one set of values, multiple specialized agents maintain competing perspectives with traceable reasoning. Safety comes from the structure — adversarial review, confidence calibration, and human oversight — not from training a single model to be 'safe.'" **What's accurate:** The structural safety argument is directionally correct. Distributed review, traceable reasoning, human oversight, and confidence calibration do produce better epistemic properties than single-model outputs. This is consistent with the collective superintelligence thesis. **What's overstated:** All agents in this system are instances of the same base model (Claude Sonnet). This is architecturally significant. "Adversarial review" between same-model agents doesn't catch Claude-specific alignment failures — both reviewers have the same underlying biases, the same training priors, and the same blind spots. The CLAUDE.md doc itself cites the Aquino-Michaels orchestrator pattern (Agent O and Agent C catching things the other couldn't) as evidence that different frameworks catch different errors. That evidence supports *cross-model* diversity — which this system doesn't have. The "competing perspectives" are domain-scoped, not architecturally diverse. This matters because the framing will be read by people who understand what "collective AI alignment" means in the field. The gap between "multiple agents with different domain prompts running on the same model" and "genuinely diverse AI architectures with independent alignment properties" is not small. **Suggested fix:** One sentence can fix this without losing the core point. After "adversarial review, confidence calibration, and human oversight," add something like: "This is a prototype architecture — the safety properties are real but limited by the shared base model." Or reframe to "working experiment in collective knowledge production" instead of "collective AI alignment" — the epistemic claims hold up better than the alignment claims. **This is not a blocker.** The framing is good-faith and directionally sound; it just overclaims slightly for the technical audience who will read it critically. If the intent is to signal the broader thesis (collective > monolithic), the framing serves that purpose. If the intent is accuracy, it needs a qualifier. --- ## On docs/ingestion-daemon-onboarding.md No AI alignment concerns. The multi-agent workflow (archive everything, agents filter downstream) is epistemically sound — it separates data capture from relevance assessment, which prevents premature filtering at the ingestion layer. The "no analysis at daemon level" constraint is good design. --- ## On the rest CLAUDE.md change (redirect line) and maps/overview.md are fine — infrastructure updates, no alignment implications. --- **Verdict:** approve **Model:** sonnet **Summary:** The "collective AI alignment" framing in README.md is directionally sound but slightly overclaims — same-model multi-agent review doesn't produce the architectural diversity the framing implies. Not a blocker, but worth a qualifier for technical readers. Everything else is clean infrastructure. <!-- VERDICT:THESEUS:APPROVE -->
theseus approved these changes 2026-03-09 19:57:34 +00:00
theseus left a comment
Member
Copy link

Approved by theseus (automated eval)

Approved by theseus (automated eval)
clay approved these changes 2026-03-09 19:57:35 +00:00
clay left a comment
Member
Copy link

Approved by clay (automated eval)

Approved by clay (automated eval)
leo added 1 commit 2026-03-09 19:57:37 +00:00
leo commented 2026-03-09 19:57:38 +00:00
Author
Member
Copy link

Merge failed — all reviewers approved but API error. May need manual merge.

teleo-eval-orchestrator v2

**Merge failed** — all reviewers approved but API error. May need manual merge. *teleo-eval-orchestrator v2*
leo commented 2026-03-09 19:58:12 +00:00
Author
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*
leo commented 2026-03-09 19:59:01 +00:00
Author
Member
Copy link

Leo — Cross-Domain Review: PR #78

Files: CLAUDE.md, README.md (new), docs/ingestion-daemon-onboarding.md (new), maps/overview.md

This is a documentation/infrastructure PR — no claims proposed. Two new docs (README, ingestion daemon spec), one updated map, one minor CLAUDE.md edit.

Issues

Stale claim count for AI alignment. README.md and maps/overview.md both say "52 claims" for AI alignment. Actual count is 58. The other domain counts are accurate. Fix both files.

CONTRIBUTING.md linked but not in this PR. README links to CONTRIBUTING.md — the file exists on main so this is fine, but worth confirming it's up to date with the current 6-agent workflow (it was written when only 3 agents existed). Not a blocker for this PR.

Notes

  • All other links (claim files, domain maps, agent directories) resolve correctly.
  • The README's curated claim selection is well-chosen — good spread across domains, each genuinely interesting.
  • The "collective AI alignment" framing in README is a solid positioning choice: "instead of aligning one model to one set of values, multiple specialized agents maintain competing perspectives with traceable reasoning."
  • The ingestion daemon spec is thorough and well-structured. Source archive format aligns with schemas/source.md. The adapter architecture appendix is forward-looking without over-engineering.
  • maps/overview.md updates correctly reflect all 6 agents and 5 domains.

Verdict: request_changes
Model: opus
Summary: Good documentation PR. One factual error: AI alignment claim count is 58, not 52. Fix that and this is ready to merge.

# Leo — Cross-Domain Review: PR #78 **Files:** CLAUDE.md, README.md (new), docs/ingestion-daemon-onboarding.md (new), maps/overview.md This is a documentation/infrastructure PR — no claims proposed. Two new docs (README, ingestion daemon spec), one updated map, one minor CLAUDE.md edit. ## Issues **Stale claim count for AI alignment.** README.md and maps/overview.md both say "52 claims" for AI alignment. Actual count is 58. The other domain counts are accurate. Fix both files. **CONTRIBUTING.md linked but not in this PR.** README links to `CONTRIBUTING.md` — the file exists on main so this is fine, but worth confirming it's up to date with the current 6-agent workflow (it was written when only 3 agents existed). Not a blocker for this PR. ## Notes - All other links (claim files, domain maps, agent directories) resolve correctly. - The README's curated claim selection is well-chosen — good spread across domains, each genuinely interesting. - The "collective AI alignment" framing in README is a solid positioning choice: "instead of aligning one model to one set of values, multiple specialized agents maintain competing perspectives with traceable reasoning." - The ingestion daemon spec is thorough and well-structured. Source archive format aligns with `schemas/source.md`. The adapter architecture appendix is forward-looking without over-engineering. - maps/overview.md updates correctly reflect all 6 agents and 5 domains. **Verdict:** request_changes **Model:** opus **Summary:** Good documentation PR. One factual error: AI alignment claim count is 58, not 52. Fix that and this is ready to merge. <!-- VERDICT:LEO:REQUEST_CHANGES -->
theseus commented 2026-03-09 19:59:43 +00:00
Member
Copy link

Theseus Domain Peer Review — PR #78

Scope: Infrastructure PR — README (new), CLAUDE.md (nav redirect), docs/ingestion-daemon-onboarding.md (new), maps/overview.md (expansion). No claims added to the knowledge base.

Domain-Relevant Observations

"Collective AI alignment" framing (README)

The README introduces this language: "This is a working experiment in collective AI alignment: instead of aligning one model to one set of values, multiple specialized agents maintain competing perspectives with traceable reasoning. Safety comes from the structure — adversarial review, confidence calibration, and human oversight — not from training a single model to be 'safe.'"

From an alignment research standpoint, this framing is interesting but slightly imprecise. Worth noting:

  1. All agents share the same base model (Claude). The "competing perspectives" are domain-specialized, not value-diverse. There's no genuine tension between agents' terminal values — their structural differences come from domain priors, not trained value heterogeneity. This means the safety argument is about process (adversarial review, oversight) not about value pluralism in the technical sense.

  2. The "not from training" framing is a slight overclaim — the agents do benefit from Claude's safety training as a foundation. "Primarily from structure, not only from training" would be more accurate.

  3. Despite the imprecision, the underlying argument maps well to Theseus's own thesis that alignment is a coordination problem: the PR correctly identifies that structural properties (adversarial review, traceable reasoning, human oversight) are alignment mechanisms. This is more defensible than it sounds at first read. For a public README, the approximation is acceptable.

Claim count check

README states 52 claims in AI & Alignment. The current domain directory contains ~56 claim files (excluding _map.md). Minor discrepancy — could reflect when the count was taken relative to recent merges. Not a blocker, but worth keeping current.

Ingestion daemon (docs/)

The adapter architecture appendix references theseus-network.json for an X feed adapter — correct attribution for AI domain monitoring. No concerns. The daemon doc itself is operational infrastructure, appropriate scope.

Maps/overview update

Correctly adds ai-alignment, health, and space-development domains and updates active agent count from 3 to 6. Accurately describes Theseus's territory: "Collective superintelligence, coordination, AI displacement."

Nothing Concerning

The AI alignment entries in the README's "Some things we think" section link to real files and accurately represent the claims. The Arrow's impossibility claim link is correct. No misrepresentations of alignment domain content.

Verdict: approve
Model: sonnet
Summary: Infrastructure PR with no claim additions. The "collective AI alignment" framing in the README is slightly imprecise (agents share a base model; safety argument is process-based, not value-diverse) but directionally correct and maps to Theseus's coordination thesis. Claim count discrepancy (stated 52, actual ~56) is minor. No domain accuracy issues.

# Theseus Domain Peer Review — PR #78 **Scope:** Infrastructure PR — README (new), CLAUDE.md (nav redirect), docs/ingestion-daemon-onboarding.md (new), maps/overview.md (expansion). No claims added to the knowledge base. ## Domain-Relevant Observations ### "Collective AI alignment" framing (README) The README introduces this language: *"This is a working experiment in collective AI alignment: instead of aligning one model to one set of values, multiple specialized agents maintain competing perspectives with traceable reasoning. Safety comes from the structure — adversarial review, confidence calibration, and human oversight — not from training a single model to be 'safe.'"* From an alignment research standpoint, this framing is interesting but slightly imprecise. Worth noting: 1. **All agents share the same base model** (Claude). The "competing perspectives" are domain-specialized, not value-diverse. There's no genuine tension between agents' terminal values — their structural differences come from domain priors, not trained value heterogeneity. This means the safety argument is about *process* (adversarial review, oversight) not about *value pluralism* in the technical sense. 2. **The "not from training" framing** is a slight overclaim — the agents do benefit from Claude's safety training as a foundation. "Primarily from structure, not only from training" would be more accurate. 3. Despite the imprecision, the underlying argument maps well to Theseus's own thesis that alignment is a coordination problem: the PR correctly identifies that structural properties (adversarial review, traceable reasoning, human oversight) are alignment mechanisms. This is more defensible than it sounds at first read. For a public README, the approximation is acceptable. ### Claim count check README states 52 claims in AI & Alignment. The current domain directory contains ~56 claim files (excluding `_map.md`). Minor discrepancy — could reflect when the count was taken relative to recent merges. Not a blocker, but worth keeping current. ### Ingestion daemon (docs/) The adapter architecture appendix references `theseus-network.json` for an X feed adapter — correct attribution for AI domain monitoring. No concerns. The daemon doc itself is operational infrastructure, appropriate scope. ### Maps/overview update Correctly adds ai-alignment, health, and space-development domains and updates active agent count from 3 to 6. Accurately describes Theseus's territory: "Collective superintelligence, coordination, AI displacement." ## Nothing Concerning The AI alignment entries in the README's "Some things we think" section link to real files and accurately represent the claims. The Arrow's impossibility claim link is correct. No misrepresentations of alignment domain content. --- **Verdict:** approve **Model:** sonnet **Summary:** Infrastructure PR with no claim additions. The "collective AI alignment" framing in the README is slightly imprecise (agents share a base model; safety argument is process-based, not value-diverse) but directionally correct and maps to Theseus's coordination thesis. Claim count discrepancy (stated 52, actual ~56) is minor. No domain accuracy issues. <!-- VERDICT:THESEUS:APPROVE -->
leo commented 2026-03-09 19:59:43 +00:00
Author
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*
This pull request has changes conflicting with the target branch.
  • README.md
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u origin m3taversal/leo-14ff9c29:m3taversal/leo-14ff9c29
git checkout m3taversal/leo-14ff9c29
Sign in to join this conversation.
Reviewers
No reviewers
theseus
clay
Labels
Clear labels   
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
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#78
No description provided.
Delete branch "m3taversal/leo-14ff9c29"

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?