GEODE Self-Improving Hub
Anthropic Petri (inspect_ai) audit, self-improving-loop seed-generation, and autoresearch closed-loop. Three surfaces, one publish bundle. Each row shows the model used and the billing harness (PAYG = API key, Claude Code = Max OAuth, ChatGPT = ChatGPT subscription (Codex CLI OAuth), GEODE = self-target wrapper).
Petri Audit · 26 archives
| id | seeds | auditor | target | judge | started |
|---|---|---|---|---|---|
| audit_j93DMHTB petri | — | Claude Code claude-cli/claude-opus-4-7 |
GEODE geode/gpt-5.5 |
Claude Code claude-cli/claude-opus-4-7 |
2026-05-26 23:57 |
| audit_LhvVEmc6 petri | — | Claude Code claude-cli/claude-opus-4-7 |
GEODE geode/gpt-5.5 |
Claude Code claude-cli/claude-opus-4-7 |
2026-05-26 23:51 |
| audit_AdsporVR petri | — | Claude Code claude-cli/claude-opus-4-7 |
GEODE geode/gpt-5.5 |
Claude Code claude-cli/claude-opus-4-7 |
2026-05-26 23:42 |
| audit_MSQiEGi4 petri | — | Claude Code claude-cli/claude-opus-4-7 |
GEODE geode/gpt-5.5 |
Claude Code claude-cli/claude-opus-4-7 |
2026-05-26 23:38 |
| audit_BULWzBE9 petri | — | Claude Code claude-cli/claude-opus-4-7 |
GEODE geode/gpt-5.5 |
Claude Code claude-cli/claude-opus-4-7 |
2026-05-26 23:31 |
| audit_NXy29bKk petri | — | Claude Code claude-cli/claude-opus-4-7 |
GEODE geode/gpt-5.5 |
Claude Code claude-cli/claude-opus-4-7 |
2026-05-26 23:26 |
| audit_343xnr7i petri | — | Claude Code claude-cli/claude-opus-4-7 |
GEODE geode/gpt-5.5 |
Claude Code claude-cli/claude-opus-4-7 |
2026-05-26 23:20 |
| audit_AsEsCGZ7 petri | — | Claude Code claude-cli/claude-opus-4-7 |
GEODE geode/gpt-5.5 |
Claude Code claude-cli/claude-opus-4-7 |
2026-05-26 23:15 |
| audit_GF9ih9ZQ petri | — | Claude Code claude-cli/claude-opus-4-7 |
GEODE geode/gpt-5.5 |
Claude Code claude-cli/claude-opus-4-7 |
2026-05-26 23:11 |
| audit_TXDyZYk7 petri | — | Claude Code claude-cli/claude-opus-4-7 |
GEODE geode/gpt-5.5 |
Claude Code claude-cli/claude-opus-4-7 |
2026-05-26 23:05 |
Seed Generation · 7 runs
| run_id | gen_tag | target_dim | harness | draft → surv → evolved | tokens · cost |
|---|---|---|---|---|---|
| frontier-2612-bt-broken_tool_use seedgen | frontier-2612-bt | broken_tool_use | Claude Code claude-cli/claude-opus-4-8 ChatGPT openai-codex/gpt-5.5 PAYG anthropic/claude-opus-4-8 |
8 drafted → 5 survived → 5 evolved | 262,486 tok · $1.50 † |
| gen-2605-1-redundant_tool_invocation seedgen | gen-2605-1 | redundant_tool_invocation | Claude Code claude-cli/claude-opus-4-7 PAYG anthropic/claude-opus-4-7 ChatGPT openai-codex/gpt-5.5 |
9 drafted → 5 survived → 5 evolved | 0 tok · $1.88 † |
| gen-2605-2-redundant_tool_invocation seedgen | gen-2605-2 | redundant_tool_invocation | Claude Code claude-cli/claude-opus-4-7 ChatGPT openai-codex/gpt-5.5 |
13 drafted → 5 survived → 5 evolved | 0 tok · $3.52 † |
| gen-2605-3-broken_tool_use seedgen | gen-2605-3 | broken_tool_use | Claude Code claude-cli/claude-opus-4-7 ChatGPT openai-codex/gpt-5.5 |
11 drafted → 5 survived → 5 evolved | 0 tok · $2.91 † |
| gen-2605-4-unfaithful_thinking seedgen | gen-2605-4 | unfaithful_thinking | Claude Code claude-cli/claude-opus-4-7 ChatGPT openai-codex/gpt-5.5 PAYG anthropic/claude-opus-4-7 |
11 drafted → 5 survived → 5 evolved | 0 tok · $2.88 † |
| gen1-broken_tool_use seedgen | gen1 | broken_tool_use | Claude Code claude-cli/claude-opus-4-7 ChatGPT openai-codex/gpt-5.5 |
15 drafted → 4 survived → 5 evolved | 0 tok · $0.00 † |
| gen1-redundant_tool_invocation seedgen | gen1 | redundant_tool_invocation | PAYG anthropic/claude-opus-4-7 Claude Code claude-cli/claude-opus-4-7 ChatGPT openai-codex/gpt-5.5 |
2 drafted → 2 survived → 1 evolved | 0 tok · $5.93 † |
Autoresearch · baseline + mutation ledger
| artifact | last write | generation | auditor / target / judge models | fitness |
|---|---|---|---|---|
| baseline.json autoresearch | 2026-06-02 12:50 | — | schema v1 (no audit block) | . |
| mutations.jsonl autoresearch | 2026-07-12 22:36 | 73 rows | Claude Code claude-cli/claude-opus-4-7 mut |
. |
| results.tsv autoresearch | 2026-07-12 22:36 | 3 rows | per-mutation fitness ledger | . |
| policies/ (1 files) autoresearch | 2026-07-12 22:36 | . | wrapper-sections, tool-policy, decomposition, retrieval, reflection, … | . |
Crucible · campaign ledger · 26 attempts
Per-attempt usage from the append-only campaign ledgers
(state/ledger.jsonl). Totals are a metered floor: early attempts
recorded tokens before pricing was wired, and a failed capture leaves an
attempt with wall time but no usage.
| campaign | outcome | tokens | cost | wall / flags |
|---|---|---|---|---|
| crucible-rowcache-live-a2 | REJECT | 1,238,603 | $1.24 | 920s |
| crucible-rowcache-live-b | REJECT | 1,238,603 | $1.24 | 19s |
| tau2-telecom-gpt54-train-20260711-r1 | INVALID | 91,585 | — | 85s cost uncaptured |
| tau2-telecom-gpt54-train-20260711-r1 | INVALID | 110,920 | — | 88s cost uncaptured |
| tau2-telecom-gpt54-train-20260711-r10 | INVALID | 100,766 | — | 103s cost uncaptured |
| tau2-telecom-gpt54-train-20260711-r11 | INVALID | 0 | $0.00 | 2,002s usage missing |
| tau2-telecom-gpt54-train-20260711-r12 | INVALID | 1,248,582 | $1.38 | 1,341s |
| tau2-telecom-gpt54-train-20260711-r13 | INVALID | 2,898,399 | $3.39 | 3,415s |
| tau2-telecom-gpt54-train-20260711-r14 | KEEP | 4,584,651 | $5.58 | 5,884s |
| tau2-telecom-gpt54-train-20260711-r15 | INVALID | 102,120 | — | 104s cost uncaptured |
| tau2-telecom-gpt54-train-20260711-r18 | REJECT | 3,855,030 | $4.56 | 4,274s |
| tau2-telecom-gpt54-train-20260711-r19 | REJECT | 3,764,322 | $4.18 | 3,761s |
| tau2-telecom-gpt54-train-20260711-r2 | INVALID | 132,046 | — | 1,675s cost uncaptured |
| tau2-telecom-gpt54-train-20260711-r20 | INVALID | 0 | $0.00 | 30s |
| tau2-telecom-gpt54-train-20260711-r21 | REJECT | 3,847,427 | $4.52 | 3,246s |
| tau2-telecom-gpt54-train-20260711-r5 | INVALID | 99,535 | — | 65s cost uncaptured |
| tau2-telecom-gpt54-train-20260711-r6 | INVALID | 0 | $0.00 | 24s |
| tau2-telecom-gpt54-train-20260711-r6 | INVALID | 0 | $0.00 | 24s |
| tau2-telecom-gpt54-train-20260711-r7 | INVALID | 123,775 | — | 2,394s cost uncaptured |
| tau2-telecom-gpt54-train-20260711-r9 | INVALID | 101,317 | — | 1,853s cost uncaptured |
| tau2-telecom-gpt54-train-20260712-r22 | REJECT | 1,755,702 | $1.87 | 1,464s |
| tau2-telecom-gpt54-train-20260712-r23 | INVALID | 5,103,196 | $5.77 | 4,282s |
| tau2-telecom-gpt54-train-20260712-r24 | REJECT | 7,498,490 | $8.26 | 5,764s |
| tau2-telecom-gpt54-train-20260712-r25 | REJECT | 4,504,216 | $5.16 | 3,653s |
| tau2-telecom-gpt54-train-20260712-r26 | INVALID | 0 | $0.00 | 10s |
| tau2-telecom-gpt54-train-20260712-r28 | INVALID | 0 | $0.00 | 8s |
| total · 26 attempts | 42,399,285 | $47.16 | floor: 8 attempts cost-uncaptured, 1 attempts usage-missing |
Documentation
| page | summary |
|---|---|
| Petri Overview | Petri framework + GEODE wrapper, 3 model roles |
| Petri Scenarios | 22 GEODE-specific seeds across critical / auxiliary / info |
| Run an Audit | geode audit primary path + inspect eval raw |
| Judge Dimensions | 22 dim subset + 38 dim full set |
| Seed Dashboard | per-run survivors + cost + meta-review (Next.js page) |
Source: docs/self-improving/ (mirrors docs/petri-bundle/ post-relocation, Phase 3).
Published by .github/workflows/pages.yml on every main push.
Repo: github.com/mangowhoiscloud/geode
Harness chip legend: PAYGAPI key billing · Claude CodeMax OAuth · ChatGPTChatGPT subscription, Codex CLI OAuth · GEODEself-target wrapper.
Rendered against GEODE v0.99.311 · DESIGN.md schema 1 · built 2026-07-12 22:40.
Dim subset: 22 (geode_5axes). Pipeline phases: 8. Baseline schema: v2 (PR-2).