Registered nested-support execution
Explicit top-level fitness selection
The same support-execution commands also accept a declaration whose format
is exactly top_level_fitness_execution_v1, paired with a descriptor prepared
by fitness-results prepare --protocol top_level_fitness_result_v1. This is a
separate, fixed instrument; a support descriptor cannot be used with the fitness
protocol or conversely. The existing support protocol and saved support reader
retain their original formats and behavior.
The original fitness request retains its captured model, field specification,
typed value and system prompt. The effective wire request separately records
removal of null temperature, stream: false, and any explicitly declared
thinking/effort options. A fitness registration additionally pins the fitness
result, schema-digest and fitness-instrument source modules. Captured recheck
uses the retained authority and does not inspect current source or reconnect.
Fitness execution reports identify axis: fitness and count only the
fitness_top_level stratum. Each admitted row contains its complete
saved_result with the strict fitness/failure/reason assessment, plus a separate
outer dispatch status. A valid score body received with an HTTP or transport
failure remains a failed execution; observed usage is still retained. Unstarted
requests and admitted requests with missing responses remain distinct.
All scheduling, single-attempt, raw-capture and caller-decision rules below also apply to fitness. Local fixture replies and mechanically accepted replies remain scientifically ineligible. No registration changes the original plan readiness, establishes calibration, combines fitness with nested support, or authorizes a paid campaign.
d4d evaluate support-execution explicitly opts into nested_support_execution_v1.
It consumes an existing nested_support_result_v1 descriptor and preserves its
original readiness blockers, selected records, context, schemas and requests.
The existing planner, saved-response commands and legacy evaluation calls do not
select this executor automatically. Top-level fitness requires its separate
descriptor and adapter; it is not a support verdict.
Registration and execution
Prepare a descriptor with support-results prepare, then provide a closed JSON
declaration. This example is a fictional local test configuration. All limits
are illustrative, not evaluation recommendations. The loopback endpoint must be
an explicitly started local fixture; this command does not start a server.
{
"format": "nested_support_execution_v1",
"registration_id": "fictional-local-example",
"purpose": "local_fixture",
"run_output": "/absolute/fresh/local-run",
"transport": {
"adapter": "native_message_http_json_v1",
"url": "http://127.0.0.1:8765/v1/messages",
"model": "EXACT_MODEL_FROM_DESCRIPTOR",
"auth": "none",
"anthropic_version": "2023-06-01",
"timeout_ms": 2000,
"thinking": null,
"effort": null,
"retries": 0,
"redirects": false,
"environment_proxies": false
},
"limits": {
"max_calls": 2,
"request_bytes": 100000,
"response_bytes": 50000,
"total_response_bytes": 100002,
"input_reservation_per_call": 100,
"input_scheduling_threshold": 200,
"output_scheduling_threshold": 834
},
"decisions": {
"instrument_review": null,
"context_review": null,
"controls_review": null,
"private_control_handling": null,
"paid_authorization": null,
"calibration_acceptance": null,
"canary_acceptance": null
},
"prices": null
}
d4d evaluate support-execution prepare --descriptor descriptor-dir \
--declaration declaration.json --output registration-dir
d4d evaluate support-execution run --registration registration-dir
d4d evaluate support-execution recheck --run /absolute/fresh/local-run
The actual dispatcher makes one nonstreaming messages POST per admitted request
using HTTPX, with no retries, redirects or environment proxies. The registered
wire transformation removes only the planned null temperature, adds
stream: false, and applies explicit thinking (null or adaptive) and effort
(null, low, medium, high, max). The model and requested max_tokens remain exact.
Both original and effective request bytes are retained. Unsupported request
shapes refuse before dispatch. timeout_ms is HTTP phase inactivity timeout,
not a total wall-clock deadline.
local_fixture requires a literal loopback HTTP address, rejects credentials and
approval references, and remains scientifically ineligible even when a response
is mechanically accepted. calibration and cohort require HTTPS, explicit
bearer or x-api-key authentication, and explicit decision references. Each
reference is { "reference": "...", "sha256": "64 lowercase hex characters" }.
Calibration requires the first five decisions; cohort requires all seven.
These declarations are not authenticated approvals or checks of scientific
truth. The software does not discover private controls or choose their handling.
The caller must review the exact captured request content and authorize any
export of selected private material before using a paid route.
Paid use still needs actual owner authorization; preparing a declaration does
not provide it. Credentials are supplied only at runtime through an explicitly
named --credential-env variable and are not saved.
Scheduling, accounting and failure capture
Input reservations are provisional scheduling thresholds. There is no tokenizer or count endpoint, so the first call can exceed its reservation. Unknown usage, a reservation overrun, an incomplete or rejected reply, and any transport failure stop further admission. This is not a hard input-token or monetary cap. Such unimplemented configuration fields are rejected. Requested output tokens, actual call count and captured byte limits are enforced. Zero-valued reported cache counters remain zero; absent counters remain unknown. Reported usage from failed calls remains in accounting rather than disappearing into accepted-only totals.
Optional prices have exactly currency, per_tokens, source, as_of, and
rates fields. Rates are finite nonnegative decimal strings keyed by any of
input_tokens, output_tokens, cache_read_input_tokens, and
cache_creation_input_tokens. A cost is calculated only with all four counters
and corresponding rates; otherwise it is unknown. No current prices are fetched.
These calculations use the declared rate basis and do not verify a provider bill.
Rates have at most 32 significant digits and a stored decimal exponent from -18 through 18; the positive integer denominator is at most 10^18. Costs use an explicit 64-digit ROUND_HALF_EVEN context independent of ambient decimal settings. Extreme rates refuse before admission. A complete failed HTTP reply can report usage and an estimated cost while its execution status remains failed; observed counters are not verified billing.
Admission is exclusively created and fsynced at the one registered run location
before HTTP dispatch. Reentry, including a copied registration pointing to that
location, refuses. A process interruption after admission cannot purchase a
replacement. Incomplete admitted evidence is spent_unknown; complete retained
raw evidence can be settled offline without another request. Recheck never
resumes unstarted calls. Symlink destinations and aliases are rejected.
Every later admission requires the preceding durable settlement. A missing
terminal settlement can still be reconstructed from its retained complete raw
response, without purchasing a replacement or accepting a later admission.
Raw HTTP body bytes are bounded during reading, with at most one extra byte to prove overflow. Identity encoding is requested; unexpected encodings fail and their bounded raw bytes are retained without decompression. HTTP status, request-id, transport stage, sanitized failure class and elapsed duration are saved before semantic parsing. Headers, cookies, credentials and exception text are not recorded. Raw message usage, reasoning and redacted reasoning remain available through the saved assessment; none establishes response correctness.
capture_run(directory) captures a closed ledger and all referenced artifacts.
recheck_captured(capture, ledger) reconstructs it using captured bytes only,
including after relocation. It does not verify the current runtime, contact a
provider, read original records or trust saved acceptance flags. This proves
internal evidence consistency and local invocation association, not cryptographic
provider provenance. Reports keep support edge/value counts separate and retain
scientific_eligibility: false; empirical calibration, reviewed human labels,
context choice and subsequent campaign authorization remain separate work.