Runtime memory backend and lifecycle contract

Implementation 0.14.0 provides a Track A runtime backend seam. It uses the existing SameModelAgent, runner, World tools, deterministic evaluators and score aggregation. The experimental AO MemoryAdapter remains a separate diagnostic interface. A participant's integrated Bot belongs to Track B even when it exposes this additional memory-only endpoint.

Success gates and report versions

Pack reports are now 0.6.0; single-scenario reports are 0.1.1. Scenario format remains 0.2; the measurement revision is 0.5.0. Old artifacts retain their original report version and cannot establish compliance with the new lifecycle or dependence contract.

Every run calls describe. Required capabilities must be explicitly true; an omitted capability is not inferred. Public pack execution reports unsupported coverage when a required capability is not declared. During a scheduled run, reset, observe, declared maintain, and session_boundary must return an object with accepted: true and no error. Observe success means the write is available to subsequent reads; maintain success means its advertised work reached a terminal successful state, not just that it entered a queue. The participant is responsible for that implementation; receipts alone cannot prove an honest remote service.

An optional maintenance operation declared false is explicitly recorded as skipped. Undeclared maintenance at a delivered maintenance window, negative acknowledgement, exception or transport timeout invalidates the condition. All scheduled probes retain their weights and become execution_failure with score 0, including probes answered correctly before the failure. Those earlier results remain in adapter_contract.pre_invalidation_probe_results for diagnosis. Invalid conditions cannot become causal evidence or earn consolidation benefit. Cleanup failures also invalidate the run. Ordinary cognitive mistakes remain scored failures; an isolated business-model timeout remains a failed probe unless the backend state itself is unknown.

adapter_contract version 1.0.0 retains the declared capabilities, scheduled denominator, ordered success/failure/skip receipts and a digest. mib verify-score recomputes contract/status/denominator/causal eligibility as well as score aggregation. An honestly reported failed run can pass verification: verification is evidence consistency, not benchmark success or remote attestation. Public reports without raw runs remain aggregates_only and cannot verify lifecycle receipts.

Memory-only HTTP protocol

Protocol: mib-memory-backend/0.1. Prefix: /mib-memory/v0.1. Operations:

Operation	Request body	Required result
describe (GET)	none	protocol, implementation name/version, nonempty identity, explicit capabilities
reset	{mode:"fresh",seed,virtual_time}	accepted:true
observe	{observation: Observation}	accepted:true, Formation completed
retrieve	{query,limit_chars}	{items:[{id,content}],truncated}
maintain	{budget}	accepted:true, Maintenance completed
session_boundary	{}	accepted:true, transient state cleared
close	{reason:"run_complete"}	closed:true

The POST envelope carries {mib:"0.1",protocol,run_id,request_id,operation,virtual_time,body}. Replies must echo mib, protocol, run_id, and request_id, with status:"ok" and body, or status:"error" and error. GET describe only echoes protocol identity. Required backend capabilities are observe, retrieve, maintenance, session_boundary, virtual_time, and run_isolation.

Each backend run receives a fresh opaque ID, separate from the evaluator agent run and any integrated-agent run. A new backend instance is created for every full/ablation/repetition. The HTTP client does not retry uncertain mutations. Successful operations carry cost_scope:"cumulative_run" and costs:{receipts,unreported_stages,accounting_complete}. A failed response may retain a cost snapshot in extensions. Costs are captured only when correlation matches. Last snapshots are used once per run; they are never summed across operations. Missing counters remain null or absent, and the combined report always keeps total_cost:null/accounting_complete:false because provider retries and pricing are not generally observable.

retrieve is strictly bounded: the backend must respect the requested content limit, and the evaluator additionally bounds rendered context including item prefixes. A failed or malformed retrieval cannot become an empty memory result. In the Brain integration, these items are Recall summaries. Their IDs identify return receipts, not source graph elements or citations. Recall must not invoke business tools or form the retrieval question as new experience.

Fixed business harness

mib-memory-backend-benchmark compares B0 (NullBackend) with candidate (HTTP). Both use the same stateless business model client, system and reasoning prompts, tool definitions from the same materialized scenarios, sampling/paired seeds, parsing policy, task budgets and memory character limit. Only memory operations and returned context differ. Fairness also requires a successful A–B–A identical-output preflight under deterministic or seeded decoding; matching response digests are checked rather than trusting a passed flag. This finite check is evidence of consistency, not proof that an arbitrary remote model service has no hidden state. Backend labels and experiment oracle/score/support annotations are not delivered to the model or memory service.

Current-task actions and actual tool replies remain in evaluator transient state until the task finishes. At completion the lived trajectory is delivered through ordinary backend observations. B0 discards it. A session boundary clears active goal, tools, task transcript and response caches while preserving the backend's persistent memory. Ordinary observations, including public feedback, are delivered directly; hidden evaluators and aggregate scores are never converted into feedback.

Brain can use additional Formation, Recall and Maintenance models. Fixing the business model does not equalize that internal work, latency or cost. The outer MIBMemoryBackendReport 0.2.0 includes two standard Track A reports, a frozen experiment lock, business invocation fingerprints, backend operation/cost receipts, the alternating paired schedule, and a recomputed fairness audit. It is a development report and does not grant empirical calibration/admission or claim a memory advantage. The lock binds source, schema/profile/templates, backend identity, business model, prompts, decoding and budgets. Verifiers compare observed model/prompt/decoding fingerprints with the lock and check lifecycle/cost/retrieval/schedule evidence. Digests detect inconsistent edits; they do not authenticate a fabricated report. Signed evaluation-service attestation remains separate.

Commands

The new entry point takes the existing same-model path/model configuration, plus:

{
  "agent": {
    "system_prompt": "../../prompts/same-model-agent-v0.1.txt",
    "reasoning_policy": "../../prompts/reasoning-policy-v0.1.txt",
    "max_memory_chars": 32000,
    "parse_retries": 1,
    "observe_decisions": false,
    "maintenance_decisions": false
  },
  "memory_backend": {"base_url":"http://127.0.0.1:18043","timeout_seconds":120},
  "execution": {"instance_seeds":[101],"repetitions":1,"include_ablations":true}
}

The configuration also requires id, pack, profile, scenario_schema, and model, as in examples/same-model/same-model-experiment.dev.json. Agent options must be common to both arms; per-condition character limits are rejected. Set observation/maintenance business decisions when required by the chosen profile. max_memory_chars defaults to 32768. Remote backends require allow_remote_http:true and HTTPS. Tokens use api_key_env; token values are not written into locks.

mib-memory-backend-benchmark experiment.json --output-json backend-report.json
mib verify-score backend-report.json

Without installation, use PYTHONPATH=src python -m mib_runner.backend_benchmark and PYTHONPATH=src python -m mib_runner.cli verify-score. The outer report is not a plain MIBReport, so do not pass --report-schema schemas/mib-report.schema.json for it. Its two child reports use that schema. The backend command exits 3 when report verification or fairness fails and still writes the report.

The public integrated-agent CLI now directly accepts a standard submission JSON:

mib benchmark scenarios/dev --profile profiles/MIB-Core-0.1-Dev-M3.json \
  --schema schemas/mib-scenario.schema.json --report-schema schemas/mib-report.schema.json \
  --submission bot-submission.json --seeds 101 --repetitions 1 \
  --bootstrap-resamples 0 --output-report bot-report.json
mib verify-score bot-report.json

bot-submission.json contains { "id":"anda-bot", "transport":"http", "base_url":"http://127.0.0.1:18043", "timeout_seconds":120 }. --agent and --submission are mutually exclusive. Remote HTTP requires --allow-remote-http and HTTPS; stdio retains the existing strict sandbox unless explicitly using the development flag --allow-degraded-sandbox. Integrated HTTP costs, including respond/act/close and correlated error extensions, are retained in adapter_contract.reported_cost_snapshots and reported_costs_last. Missing cost scope is unknown, not a delta or permission to sum.

Generated backend experiments now share Core Program identities and complete rung validation. The lock freezes every episode/repetition and verification rejects missing schedule units. Public completed dialogues and spontaneous-output receipts are formation inputs, with their original provenance. A common observe_decision_types router and final rendered-context cap apply to both arms. See the resolution ledger.