The endgame: a knowledge graph that reasons about itself (self_directed_graph, no LLM)#756
Conversation
|
Warning Review limit reached
Next review available in: 5 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (4)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_128ec102-3cd7-4d1e-b017-5c01ffcaad75) |
…t itself (no LLM) A TripletGraph runs one closed autonomous loop with no external input after the seed: INTROSPECT (read its own beliefs into a free energy) -> SELF-DIRECT (saccade to the highest-surprise locus) -> THINK (its own state feeds the shipped select_tactic; contradiction -> preserve both poles, gap -> infer_deductions, re-derivation -> confirm) -> COMMIT (revise_with_evidence writes the conclusion back into itself, reshaping the next cycle's surprise) -> repeat until F < floor. The loop IS the cognition; the graph IS the tissue. Measured (8-belief seed, no LLM, 6 KILL-gates green): - self-model grew 8 -> 14 beliefs by reasoning. - the syllogism self-derived: the graph concluded `socrates ... mortal` via NARS 2-hop over its own edges (f=f1*f2, c=f1*f2*c1*c2); confidence 0.76 (via greek) -> 0.81 confirmed by a 2nd independent path (via human) — multi-hop witness. - self-directed order: contradiction (F=0.85) before gaps (0.45) before confirms. - self-corrected on itself: detected oracle affirms/denies prophecy, preserved both poles + stamped a meta-belief (a belief about its beliefs). - reached a coherent fixed point (rested, no runaway). Honest finding, worked around in-example: infer_deductions does relation-blind transitivity (composes any two edges sharing an entity), unsound for non-transitive relations (praises/taught) — chained only is_a and isolated the contradiction on leaf nodes. Debt: TD-INFER-DEDUCTIONS-RELATION-BLIND. Boundary: a small symbolic reasoner over its own edges — the "AGI without LLM" claim is architectural (cognition in the loop, no weights), not scale. Board: E-SELF-DIRECTED-GRAPH-1 (capstone), TD-INFER-DEDUCTIONS-RELATION-BLIND, persistent-nars-kg-v1 endgame-capstone note. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016b33swuXE23hKtqxsHu9p1
60d5f51 to
92489b3
Compare
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 60d5f51f38
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| let confirmations: Vec<Triplet> = all_deduced | ||
| .iter() | ||
| .filter(|t| { | ||
| concluded(&g, &t.subject, &t.object) | ||
| && !applied_confirm | ||
| .contains(&(t.subject.to_lowercase(), t.object.to_lowercase())) | ||
| }) |
There was a problem hiding this comment.
When a previously derived edge is recomputed in the next loop, this treats it as a confirmation solely because the (subject, object) pair already exists; it does not check that the inferred relation/path differs from the one that created the edge. In a seed with only one route to socrates→mortal, the same 2-hop deduction would still enter this branch, get revised upward, and satisfy the multi-hop witness KILL gate, so the example can report independent confirmation without an independent witness.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Good catch — fixed in 5902bc4f. The confirmation branch now requires a genuinely distinct path: a deduction only counts as an independent-witness confirmation if the graph already holds that (s,o) via a different derivation relation; a same-path recompute is rejected. A single-route seed would now correctly fail the multi-hop-witness gate instead of spoofing it. socrates→mortal stays confirmed because its two paths are edge-disjoint (via greek vs via human); the spurious same-path "confirmations" (socrates→human, philosopher→mortal, plato→human) are gone. All six KILL-gates still green.
Generated by Claude Code
…756 P2) Codex r3610377133 (P2): the confirmation branch treated any re-derived (subject, object) as an independent-witness confirmation solely because the pair already existed — it never checked the confirming derivation used a DIFFERENT path than the one that created the edge. A single-route seed would re-derive its one 2-hop fact, enter the confirm branch, revise up, and satisfy the multi-hop-witness KILL gate without any independent witness. Fix: a deduction only counts as a confirmation if the graph already holds that (s,o) via a DIFFERENT relation (path); a same-path recompute is rejected. Now only genuinely distinct-path re-derivations lift confidence. socrates→mortal remains confirmed (edge-disjoint: via greek vs via human, 0.76→0.81); the spurious same-path "confirmations" (socrates→human, philosopher→mortal, plato→human) are gone. All 6 KILL-gates still green; a single-route seed would now correctly FAIL the witness gate instead of spoofing it. Board: E-SELF-DIRECTED-GRAPH-1 witness-gate hardening note. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016b33swuXE23hKtqxsHu9p1
2 participants
What
crates/lance-graph/examples/self_directed_graph.rs— the capstone the whole persistent-NARS-KG arc was building toward: aTripletGraphthat reasons about itself, with no LLM and no external input after the seed. Additive: one example + board only, no core change.The loop (autonomous — surprise drives everything)
select_tactic.infer_deductions(NARS 2-hop over its own edges); re-derivation → confirm.revise_with_evidencewrites the conclusion back into the graph, reshaping the next cycle's surprise.The loop IS the cognition; the graph IS the tissue it reads from and writes to.
Measured (8-belief seed, no LLM, 6 KILL-gates green)
socrates … mortalover its own edges (f=f₁f₂, c=f₁f₂c₁c₂); confidence 0.76 (path viagreek) → 0.81 confirmed by a 2nd independent path (viahuman) — the multi-hop witness, self-applied. No LLM produced that conclusion; the loop did.oracle affirms∧denies prophecy, preserved both poles as a committed opinion + stamped a meta-belief (a belief about its beliefs — the self-reference).Honest finding (surfaced + worked around in-example)
TripletGraph::infer_deductionsdoes relation-blind transitivity (composes any two edges sharing an entity) — unsound for non-transitive relations (praises,taught): a contradiction node spuriously inherited a wholeis_aclosure and broke the fixed point on the first cut. The example chains onlyis_a, isolates non-transitive relations on leaf nodes, and treats a same-(s,o)conflict between two(via …)-derived beliefs as a multi-path confirmation, not a conflict. Debt filed:TD-INFER-DEDUCTIONS-RELATION-BLIND(payoff: giveinfer_deductionsa transitive-relation predicate).Honest boundary
A small symbolic reasoner over its own edges reaching a fixed point — not general intelligence. The "AGI without LLM" claim is architectural: the cognition lives in the closed loop + the tissue (graph / witness / surprise), with no trained weights in the loop at all. That shape — not this toy's scale — is the endgame.
Board hygiene (same commit)
EPIPHANIES.md— E-SELF-DIRECTED-GRAPH-1 (capstone; composes E-SELF-CORRECTING-KG-1 + E-MULTIHOP-WITNESS-CONFIDENCE-1 + E-AWARENESS-INGREDIENTS-1 + E-FOVEATED-AWARENESS-1).TECH_DEBT.md— TD-INFER-DEDUCTIONS-RELATION-BLIND.persistent-nars-kg-v1.md— endgame-capstone note.🤖 Generated with Claude Code
https://claude.ai/code/session_016b33swuXE23hKtqxsHu9p1
Generated by Claude Code