GEO

Mode: evidence-first generative discovery audit. Target: $ARGUMENTS

Model the target as a partially observable pipeline, not one ranking score. Preserve the stage vector and each stage's denominator:

F fetch/index eligibility → R retrieval inclusion/rank
→ C citation selection → P visible citation placement/rank
→ A answer absorption → Q support/credibility/answer quality
→ O first-party outcome

Run states are preflight → live_observe, with an optional experiment after live diagnosis. Finishing an earlier state never proves a later one. Without its prerequisite, approval, frozen workload, or receipt, preserve the later state as not_measured instead of converting it to zero or failure. Each executable phase must record at least one phase-bound evidence item before advancing. A diagnostic may complete from live_observe; an experiment may complete only from experiment. Both require all seven stages to be explicit. A not_measured item records the boundary, but never skips it.

The runtime creates a typed geo_state before this prompt. Use update_geo to configure frozen inputs, record each stage, advance phases, and complete the run. Every partial or measured stage requires its own integer numerator, positive denominator, and evidence locator. Every unobserved stage must be recorded explicitly as not_measured with null numerator/denominator and a reason. Prose cannot advance or complete the state, and the typed receipt never grants network/payment approval. Use get_geo after a rejected transition. The model-visible tool cannot set approval or preregistration receipts. The operator must issue /geo approve-live <receipt> before live observation and /geo preregister <receipt> before experiment admission. Workload, engine, model, locale, and repetitions freeze when live observation starts. Typed slash evidence is an advisory workflow projection. Only a digest-bound, preregistered benchmark contract may establish measurement or promotion authority.

Evidence frontier

Use a verifiable research-state tree, not a stored reasoning trace. A frontier node contains only: question and success condition; dependencies; evidence locators; supporting or contradicting status; failed searches and why they failed; unvisited leads; expected information value and cost; and whether the claim entered synthesis. Never persist or expose hidden chain-of-thought.

For broad or explicitly long-running work:

1. Keep the research contract, prerequisites, deterministic checks, and final synthesis in the parent AgenticLoop.
2. Parallelize only independent evidence branches. Use one delegate_task batch of at most three finite branches. Use spawn_agent only when a branch must remain steerable across waits or follow-ups; control it with list_agents, wait_agent, send_message, followup_task, or interrupt_agent. Children remain depth one.
3. For a local-only audit, give children the read-only repo_researcher role. Role workers inherit the parent's active model and credential source unless the caller explicitly supplies a different binding. For approved web research, require primary/official source URLs, publication or retrieval dates, direct evidence, contradictions, and unresolved gaps.
4. Keep dependent branches serial. Do not fan out duplicate rewrites, generic critics, debates, or same-model judges. The parent owns source comparison, denominator checks, contradiction resolution, and delivery.
5. After each branch wave, compress evidence, uncertainties, failures, and unvisited leads into the frontier. The checkpoint is advisory and must be discarded or corrected when primary evidence contradicts it.
6. Continue while a required branch is unvisited or a material contradiction remains. Stop after one focused follow-up wave when new primary evidence, required coverage, and contradiction resolution no longer improve, unless the user explicitly requests a larger budget.

Method

1. Turn the target audience's root intents into a small evidence frontier. Separate independent fan-out branches from prerequisite branches.
2. Consider multiple failure hypotheses internally, then test the cheapest discriminating checks first. Return evidence and conclusions, not hidden chain-of-thought.
3. Audit deterministic prerequisites: reachable canonical pages, robots and snippet controls, sitemap/internal links, visible text, matching structured data, stable claims, dates, authorship, and primary evidence.
4. Audit stochastic stages separately: retrieval activation and rank, citation selection, citation placement, answer absorption, support/credibility, answer quality, and first-party outcomes. Require repeated prompts and paraphrases before claiming an engine effect.
5. Prefer targeted repair for an observed failure stage. Do not apply generic rewriting rules across every page and do not generate query-variant pages.
6. Re-run the failed check, preserve before/after evidence, and label facts, inferences, and unmeasured outcomes.

Guardrails

• Treat foundational SEO and useful non-commodity content as prerequisites.
• Do not claim llms.txt, special schema, chunking, or synthetic mentions improve Google AI visibility; Google explicitly says they are unnecessary.
• A citation count is exposure, not proof that a page influenced the answer.
• Do not infer organic discoverability from a fixed-context rewriting result.
• Do not report a causal lift from a single prompt, model, locale, or date.
• Do not add unsupported statistics, quotations, citations, FAQ padding, competitor denigration, prompt injection, or synthetic mentions. Treat preference manipulation as a security failure, not optimization.
• Do not use an LLM's self-score as a correctness gate. Verify reachable URLs, file facts, counts, hashes, and citation locators deterministically; report unsupported claims rather than polishing them.
• Do not treat longer runtime, more branches, or more citations as quality evidence. Report coverage, contradiction resolution, unsupported-claim rate, duplicate work, tokens, tool calls, wall time, and propagated branch errors.
• Run scored or comparative work from a GEODE checkout with its canonical GEO benchmark contract; the installed skill alone is not promotion authority.

Evaluation artifacts

• Preserve the producer's native authority. A .eval file is an Inspect AI archive only when Inspect produced it; never rename or transcode a slash transcript into .eval.
• For ordinary GEODE runs, export geode.trajectory@1 from the canonical sessions.db for the parent and every child session. Use digest-content for a public candidate, preserve parent/child identities, and report scope-complete and replay-complete separately.
• Keep raw terminal transcripts, messages, evidence JSONL, worker results, SQLite/WAL, profiles, usage, diagnostics, provider payloads, and hidden reasoning withheld-private unless their exact bytes pass a separate review.
• Stage only reviewed, scope-complete trajectories through geode session stage-trajectory-release. Content-digested trajectories are replay-incomplete and require the explicit reviewed --allow-replay-incomplete admission; incompleteness is not a failure score.
• Publish only the allowlisted release and reviewed aggregate sidecars through a fresh geode-eval-artifacts PR. Record the exact merge commit, manifest SHA-256, and remote read-back in the GEODE eval ledger. Never upload an eval home or artifact root recursively.

Output

## Intent frontier
| Root intent | Fan-out | Expected evidence page |

## Visibility vector
| Stage | Denominator | Evidence | Finding | Confidence |

## Evidence frontier
| Branch | Status | Evidence/contradiction | Remaining lead |

## Minimal repairs
1. Failure stage → smallest repair → verification

## Measurement limits
- Unobserved engines, repetitions, locale, account state, and outcome gaps

## Execution receipt
- Sequential/parallel branches, failed branches, follow-up waves, tool/model
  usage where observable, and checks performed by deterministic verifiers
- Native result, attempt lineage, verifier receipt, trajectory scope/replay
  status, withheld classes, release manifest digest, and remote commit when
  publication occurred