Anabasis Pipeline

The pipeline maps checked handoffs. It does not make a stage permanently correct.

PROMPT + CONTEXT → BUILD → ADMIT → MEASURE → LEARN → {AUTHOR | MEASURE | STOP}
                           │          ├─ Main Judge: control-calibrated census
                           │          └─ rebuild advice: deterministic packet from recorded rows
                           └─ host verifier alone owns correctness

Use the user request verbatim, with optional --context paths:

bun run fullrun -- --prompt "<request>" [--context <path>]...

Do not build a one-off driver around a test. Unknown flags refuse. The loop is unbounded by default; its budget, repeated-no-progress and typed operational stops still apply. The controller never stops on a reading of the tasks; which component hears a battery's placement is AGENTS.md "Goals and the climb", under "Who hears the placement, and what it drives". --max-iterations is an operator cap.

Input contract

The controller admits inputs; generated code cannot declare its own data public. Which bytes and facts may enter authoring or solving, under which identity:

Input	Admission	Consumer	Identity
user's exact request	unchanged controller input	Builder	prompt digest
explicit --context files	allowlisted, bounded, hashed	Builder	ordered context digest
public catalogue or task facts	provenance-checked projection	Builder or Built Harness	content digest
changing external fact	sourced observation with retrieval time	Builder research	source record
large relational data	bounded read-only query surface	Built Harness	database and schema digest
protected verifier output or reference result	never returned as authoring or solve input	host evidence readers only	recorded identity

Build public objects by allowlist; never clone a protected object and delete its known secret keys. Research may suggest an input or engine but cannot establish truth or grant a capability. Give the Built Harness the data it legitimately needs to choose while the material choice stays with the model, and refuse unsupported identifiers and unproven provenance instead of filling gaps from hidden state. The Builder authors tasks, controls and hidden operands in its own workspace. Protected execution detail stays out of later authoring and solve inputs; only the declared aggregate feedback may inform the next experiment.

For one-line prompt shape, launch anchors and predicted task families read references/prompt-shapes.md; for context files, catalogues, research, public databases, candidate sets and leakage checks read references/public-data.md.

Four owners

Owner	Owns	Does not own
Harness Builder	request, admitted context, public research, representation, task families, candidate bundle	verifier pin, case truth, adoption
Built Harness	solve work with public task data, one mutable draft, one submit path	hidden tasks, controls, verifier source, claim state
Measurement kernel	identity, verifier acquisition, isolations, controls, verification, non-results, denominators, claims, rollback, adoption	domain diagnosis
Review	the Judge census and the deterministic rebuild advice packet	case truth, denominators, claim fields, readiness

There is no fifth owner. A row with no owner is a defect.

Authoring and validation

A finding's owner is the bundle file at fault — one of the nine BUNDLE_FILES in src/author/feedback-routing.ts, from correctness-model/brief.json to agent/config.yaml — or environment, which no bundle edit repairs. The path's prefix says which half of the bundle a repair reopens. The files are not a sequence of separate model sessions: the Builder authors the whole bundle in one persistent session. A finding filed against a name that is not one of them reaches no owner.

The Builder inspects, writes and rehearses its bundle, then submits one captured candidate. A turn that spends AUTHOR_FIRST_TOOL_CALLS (16 in src/author/author-first.ts) tool calls without changing its owned authoring files is flagged as author-first-interrupt after the turn finishes; the provider turn is not aborted. Repeated unchanged turns can end the session. When a campaign stalls with no build output, read that record before the model's prose. The owner-file table in src/author/feedback-routing.ts identifies repair locations. Feedback guides the next repair; it does not restrict a product experiment to one file. Environment blockers remain typed stops.

Handoffs

1. Brief: capability, artifact, constraints, correctness (validateBrief).
2. Battery: fresh, conforming tasks at the published count.
3. Tools and representation: one draft, one submit path, public projection.
4. Controls: accept controls and a reject corpus that discriminates.
5. Verifier: closed predicates, grounded checks, no hidden solution advice.
6. Adoption: conformance, census, F2, task count, fingerprint. See harness-contract (adoption gates).
7. Measurement: recorded bundle, one battery, case record, censored denominators, claim.
8. Learning: IterationAnalysis, judge reviews, the rebuild advice packet, admission.
9. Decision: src/run/next-move.ts builds, measures, reopens the adopted product or stops, and reads no placement. --product-policy fixed permits only measure or stop (src/run/fixed-product-policy.ts). Adoption remains a separate decision. A candidate whose battery verified no case is held as candidate-zero-verified; one with no measurement at all is held as candidate-unmeasured.

Validate every handoff before downstream evidence depends on it.

Route a failure

Signal	Action
No verification path	research and pin an external verifier
Controls do not discriminate	reopen controls or verifier
Correct artifact cannot be expressed	reopen representation contract
Built Harness lacks legitimate data	reopen the input contract above for admission/projection or harness-contract (representation contract) for tool delivery
Tool cannot be called	reopen registration or backend projection
Runtime fails before attempt	fix environment or backend; record a non-result
Failures share a mechanism	name and patch the owner
Change cannot be isolated	make no update

Code owns schema, identity, capability exposure, conformance, submission, verifier calls, non-results, comparison, acceptance, and restoration. Models own representation, task authoring, diagnosis, and the next experiment. No model judge is in the verification path.

Invalidate and report

When an upstream handoff changes: record the counterexample, name stale identities and evidence, make the smallest coherent change, rerun the boundary and its consumers, and create a successor epoch when a decision-relevant binding changed. Do not pool evidence across changed tasks, verifier, prompt, toolkit, or policy.

Read the explicit operator plan and current Git state before acting. Report the handoff, evidence read, owner, change or no-update, invalidated evidence, verification, and next unproven boundary.