Skip to content

Progressive governed decisions

Status: partially implemented; unimplemented flows below remain specifications.

Implemented integration and verification boundary

Section titled “Implemented integration and verification boundary”

Verification and independent-review collection now persist uh.decision-receipt.v0 artifacts under the owning mission’s decision-receipts/ directory and append decision.recorded lifecycle events. Receipts distinguish disabled credentials, unavailable providers, malformed responses, non-discriminating answers, advisory recommendations, and recommendations that changed the consumer’s disposition. They retain digests rather than raw prompts, source content, provider error messages, or provider responses. Human acceptance remains required; a semantic pass does not override a deterministic failure, and no result authorizes promotion, scope or spend.

TypeSafe System One questions are atomic. evaluateThreeVerdict asks one Noul per supplied non-deterministic criterion, points each question at its own state entry (for example criteria[2]), and adds a fixed battery of three report Nouls (work incomplete, blocker named, a check claimed passed that the state records as failed). The former broad acceptance choice and the tamper question are gone: criteria that already carry a deterministic result are not sent, and tamper comes from the deterministic Tool Guard, protected-path and diff state the caller supplies. The verdict is composed in code — any deterministic failure or any per-criterion Noul below REMEDIATION_THRESHOLD (0.5) yields needs-remediation; a verdict over no criterion is needs-attention, never pass; otherwise every criterion at or above PASS_THRESHOLD (0.8) with no report flag yields pass and anything else needs-attention. confidence is the minimum distance from 0.5 across the asked criterion Nouls, scaled to [0,1]; it is 0 when no criterion was asked, so three low report answers can never lend confidence to a verdict that judged nothing. Deterministic failure enters as an explicit input (state.deterministicFailure), never as a synthetic criterion: a failed or blocked required check, a failed acceptance check, or an overall failed status. It dominates every provider answer and is authoritative even when no non-deterministic criterion remains.

evaluateSystemOne bounds the request with timeoutMs (default 10 s) enforced through AbortSignal.timeout, retries only 429 and 529 up to three attempts honoring a numeric retry-after in seconds capped at 5 s and otherwise 250 ms then 1000 ms, and never throws for provider conditions. It returns a discriminated result: disabled, unavailable (timeout, transport, http), malformed (invalid_json, invalid_envelope) or ok. The zod-validated envelope requires one answer per asked question whose type matches the question (noul in [0,1], choice within its declared options with probabilities over exactly those options, score within the level range); a missing or mistyped answer is an invalid envelope. The requested model can be pinned through the call options or UH_TYPESAFE_MODEL; the receipt records the versioned model id that actually answered with its latency and bounded usage, so a future threshold can be tied to the version it was fitted on. Result kinds map onto the existing receipt statuses: disabled and unavailable record unavailable, malformed records malformed, and an answer set with no discriminating signal — zero confidence over the judged criteria with no deterministic failure or tamper to ground it — records uncertain; none of them is applied. A zero-criterion answer set has zero confidence, so it is uncertain and non-authorizing unless a deterministic failure or tamper grounds the outcome.

Provider-free regression coverage exercises timeout, 429 retry with retry-after, capped and non-numeric retry-after, 529 retries with default backoff, non-retryable statuses, invalid JSON, a missing answer, a choice outside its options, the composition table, the request shape (one question per non-deterministic criterion, no tamper question), tamper taken from deterministic state, disabled credentials, transport failure, malformed success responses, receipt validation, the all-abstention uncertain receipt, the zero-criterion non-pass composition, a caller-supplied deterministic failure with no criterion, the independent-review projection that asks no criterion, and omission of synthetic private-data sentinels. The verification integration also exercises a real failing subprocess with a synthetic provider answer: the deterministic failure remains failed and raw command output, diff text, and workspace paths are excluded from the provider request; only the established facts — expected-output paths and statuses and required-check names and statuses — reach it as evidence for the criteria the provider is asked about. It covers the observed vacuous-pass regression directly: a failing required check with no acceptance criteria and low report answers stays failed, records a deliberate available receipt rather than uncertain, and the finding states that no non-deterministic criterion was judged; with no deterministic failure the same shape records an uncertain, non-authorizing receipt.

uh verify now builds the per-criterion projection. Each declared acceptance criterion becomes one state.criteria entry: a criterion with a check_command carries its deterministic status (passed or failed), exit_code, and check_command, and is never sent to the provider; a criterion without one carries its id, description, severity, and an evidence object limited to facts the harness established — the relative path and verification status of each declared expected output, and the names and statuses of the required checks — never file contents, diff text, command output, absolute paths, or the mission prompt. tamper is deterministic: it is true only when the latest run control receipt stopped with the policy stop code. Verification also passes the deterministic failure it already established as state.deterministicFailure when any required check is failed or blocked, any acceptance check failed, or the overall status is failed, so a failed required check can never be overridden by a provider judging no criterion. The finding states what was judged — for example TypeSafe System One: no non-deterministic criteria to judge when no non-deterministic criterion remained — instead of a bare verdict. A deterministic failure stays failed regardless of the provider’s answer; a missing credential leaves native verification untouched, and unavailable, malformed, and non-discriminating responses authorize nothing.

The independent-review caller still projects only check dispositions, criterion identifiers and descriptions, severities, and review input/claim states, not per-criterion evidence states, so its live requests carry the fixed report battery alone and it is not yet migrated to the atomic per-criterion seam. With no criterion in its projection its answer set is zero-confidence and non-discriminating, so it can no longer compose a vacuous pass. Criterion descriptions remain task content, not a guarantee of anonymity; these projections are not yet a general-purpose privacy filter or a measured minimum-token design. Confidence-policy enforcement is still absent (thresholds compose the verdict, they do not gate application), and live judgment quality and the other decision flows specified below remain unverified or unimplemented.

Add a narrow semantic-recommendation layer to Ultimate Harness (UH) without turning a model into an authority. TypeSafe System One/JEV may recommend a runtime/model route, an acceptance disposition, a review escalation, a scope change disposition, or retry/stop. UH remains responsible for deterministic eligibility, tool guards, protected paths, resource admission, recovery limits, human approval, and artifact publication.

  • Telar owns intent, scope, route policy, human gates, stable identity, and cross-project governance. UH consumes only the execution-relevant mission contract and does not mutate Telar state.
  • UH owns live route resolution, adapter execution, supervision, recovery, verification, integration, and promotion mechanics.
  • JEV is a typed semantic recommender. It cannot grant a capability, widen a runtime/model allowlist, increase a budget, bypass a guard, downgrade a human gate, approve a scope expansion, or publish a policy.
  • Target policy (not implemented): missing, failed, malformed, or low-confidence responses must be non-authorizing; current callers do not enforce a confidence threshold.
  • Existing deterministic failures always win. A semantic pass cannot turn a failed check, protected-path stop, route mismatch, resource refusal, or required human gate into success.

Proposed, not accepted by the current mission schema: an optional decision_policy block:

decision_policy:
enabled: true
min_confidence: 0.75
allowed_runtimes: [oh-my-pi, command-code]
allowed_models: [openai-codex/gpt-5.6-luna]
require_provider_for_route: true
require_provider_for_retry: true

The block is additive and strict. allowed_runtimes is applied in deterministic adapter eligibility. allowed_models is the only model set from which a JEV recommendation may be applied. Provider requirements are opt-in so legacy missions retain their deterministic behavior while progressive missions can fail closed when semantic input is required.

src/schema/decisions.ts defines uh.decision-receipt.v0. A receipt contains only decision kind, typed recommendation (when valid), status, authorizer, human-required flag, confidence, provider/model, bounded usage and latency, input/response digests, reason, state transition, and timestamp. It contains no prompt, transcript, raw provider answer, command output, secret, or private source document. Receipts live under .harness/missions/<mission>/decision-receipts/; a compact allowlisted state transition is also appended to the mission event stream.

Receipt statuses are:

  • applied: a recommendation was used only within deterministic eligibility;
  • advisory: recorded but no authority-bearing effect was applied;
  • awaiting-human: a human decision is still required;
  • denied: deterministic policy rejected the recommendation or action;
  • unavailable: no configured provider or provider transport failure;
  • malformed: the provider response did not satisfy the typed answer contract;
  • uncertain: confidence was absent or below min_confidence.

uh mission run --auto first computes chooseAdapter, including capability, network, context, cost-class, and allowed_runtimes filters. JEV may select only one of those eligible adapters and, separately, one model in allowed_models. An explicit CLI model override wins over a semantic recommendation. Invalid, uncertain, or unavailable recommendations preserve the deterministic incumbent unless require_provider_for_route is true, in which case launch is blocked before process spawn.

uh verify runs required checks, acceptance checks, output verification, and TDD checks first. JEV receives only counts/IDs/statuses/severities and finding classes. A semantic remediation or tamper recommendation can harden a verification result to failure; semantic pass is advisory and never approves promotion. Deterministic check failures and blocked human checks are retained.

Independent review collection uses a compact JEV review/escalation question against hash-bound evidence summaries. JEV can request escalation or stop, but cannot accept a review, promote source work, or downgrade a required gate. Review provenance, captured-input hashes, and human acceptance remain mechanical requirements.

Not implemented as a command. The receipt schema has a scope-change kind, but the CLI has no uh mission decision command and nothing records one yet. When it is added it must hash the requested scope and send only bounded metadata to JEV, record every result as awaiting-human (or an explicit unavailable/malformed/uncertain/denied state) with applied: false, and change no mission packet, scope, allowlist, or budget; human approval stays with the existing authority surface.

Before a bounded native recovery resume, UH checks its existing deterministic stop-code/session/max-resume policy. JEV may narrow that action to stop or request human deferral; it cannot make a non-resumable stop resumable. A valid retry is applied only when deterministic recovery eligibility passes. With an optional provider, an unavailable provider leaves the deterministic recovery policy as the authorizer and is explicitly recorded; malformed or uncertain answers do not retry. require_provider_for_retry blocks retry when JEV is unavailable.

The implementation uses Node 20 Web APIs, existing artifact path guards, and existing adapter/recovery seams; it introduces no OS-specific process logic. Windows Job supervision, POSIX process-group cleanup, directory/git-worktree sandboxes, team resource waves, and container refusal behavior remain outside semantic decision code. No Runlayer SDK or cross-project dependency is added.

Runlayer lessons are limited to first-party documentation: identity and policy are separate, policy is least-privilege and evaluated before action, human approval adds a checkpoint rather than granting access, hooks default fail-closed, and bounded telemetry/outputs are appropriate. Those lessons map to UH’s deterministic guard plus semantic-receipt split; Runlayer does not become a UH authority or runtime dependency.

Telar and telar-demo remain read-only references. Telar owns planning, identity, route policy, and human authority; UH returns operational receipts. The implementation does not create a shared package, copy Telar decisions, or let a model act as a cross-project approver.

Required future coverage includes allowed-route/model enforcement, low confidence, mandatory human scope gating, deterministic-denial precedence, and recovery retry narrowing. These are obligations, not completed test or campaign claims. Existing provider-free checks are listed at the top of this document. Integrated live acceptance is incomplete. See the roadmap for unresolved policy, transport, privacy and evidence-integrity work.