D17 synthetic ensemble held-out study
The D17 study module is experimental and default-off.
D17 synthetic ensemble held-out study
Prompt-first procedure: Describe the outcome you want in your agent conversation. The agent should select and load the appropriate AIWG assets, explain material changes, request any needed approval, and report verification evidence. Exact commands and flags appear only in the CLI reference.
The D17 study module is experimental and default-off. It prepares a new, deterministic synthetic corpus, preregistration, operator forms and offline analysis on top of the shared held-out collector. It changes no existing decision workflow. No live Jev collection, fitted D09 calibration, completed human audit or production qualification is supplied.
The module prepares one corpus under two preregistration scopes:
- UNCALIBRATED diagnostic (v1, the prepare option): preregistration declares
`calibration: { scope: 'uncalibrated-diagnostic', allowedModes: ['uncalibrated-diagnostic'] }`, and the approval declares `calibration: { mode: 'uncalibrated-diagnostic' }`. Study reports carry `calibrated: false`, `d09Qualified: false` and `calibratedGate: false`, and stay HOLD or preserve an upstream ROLLBACK. Artifact and staged calibration approvals are refused under this preregistration. It cannot close AC7/AC14.
- Staged D09 calibration (v2 analysis, the prepare-staged option): a calibration
phase on the tuning and calibration splits, an operator-reviewed and D09-qualified calibration set fitted on calibration rows only, then a separate test-phase approval and calibrated scoring of the test split. This is the path that makes AC7/AC14 evaluable; see Staged D09 calibration.
Both scopes share the same corpus, split manifest, gold, review template and native templates for a given seed. No genuine D09 artifact exists until a staged calibration phase is collected, fitted and reviewed; a fixture digest must never stand in for one.
Frozen population and arms
`src/decision/ensemble-study/corpus.ts` allocates 1,800 fictional subjects: 200 development (`tuning`), 400 calibration and 1,200 final-test subjects. Each split balances `yes`/`no` within direct facts, multi-fact relationships, negation/insufficient evidence and authority-style distractors. Final-test support is 300 per slice. Twelve authored grammar families are allocated to splits before drawing parameters; different split grammars use different state/relation vocabulary. Complete families stay in one split. Worlds and parameter draws are independent within these shared templates; shared latent operations and template effects limit transfer to real work.
The world-draw hash-counter prefix is `aiwg-holdout-2497b51d-v1:D17:<split>:<familyId>:<row-index>:<k>`, with the seed digest in the family ID and rejection sampling for bounded draws. A separately keyed Fisher–Yates shuffle assigns balanced labels within each split and slice; SHA-256 digests of a different seed/item domain form opaque record IDs. Neither the ordinal seed nor the label is rendered in provider payloads or blind review IDs. State words, relation verbs and names have fixed lengths within each split so payload length does not reveal gold. Each row uses the closed `<corpus-seed>:<index>:single` seed and the registered `d17-entailment/v1` generator. The shared collector re-derives every row, its output digest, the corpus seed and the generator source digest before dispatch. Labels follow an executable latent-world oracle. A separately implemented text oracle checks development rendering. Gold and latent worlds live in `gold.json`, separate from provider payloads. Split manifests bind IDs, family memberships, slices, input digests and duplicate checks. They do not expose per-item gold hashes. Never copy existing fixtures or internal issue text into this population.
The champion receives one fresh invocation; the challenger receives three fresh invocations of pinned `jev-1.13.0`. Champion output is never reused as an ensemble member. The native policy template declares `repeated-sample`, `samples: 3`, `mean-probability-v1`, three required successful members, Jensen–Shannon disagreement at most 1,000 bps, tie/disagreement deferral, member concurrency one and fallback depth zero. Native distributions remain native; derived aggregate distributions are marked as derived stability signals. Agreement does not establish correctness or calibration.
Offline preparation from source
No build or credential is needed:
Use AIWG to complete this documented outcome: No build or credential is needed
Have it inspect the current state, explain the plan, ask before material
changes, and report the result with verification evidence.Choose `OUTPUT_DIR` immediately below the routed artifact root reported by the second command. Preparation exclusively creates the directory and its files; it will not overwrite an existing freeze. The first command writes nothing, projects all inputs through the existing D10 helper and prints budget estimates and canonical digests without exposing test labels. The documented `d17-2611-v1` seed is an offline demonstration and must not be approved for a paid held-out run: the source and seed are public enough to regenerate gold. Use a fresh uninspected operator seed for any paid study and approve its newly computed pins. Development review must precede the external preregistration approval and test access.
The files include `corpus.json`, `preregistration.json`, `gold.json`, `analysis.json`, `splitManifest.json`, `nativeTemplates.json`, `reviewTemplate.json`, `guide.json`, `approval-template.json` and `dryRun.json`. The preregistration binds the analysis digest, source-module byte digest and corpus. It records the pre-collection gold-label leak correction, the three superseded pins and an explicit no-live-observations declaration. The previous corpus assigned labels by index parity and rendered that index in a record ID; its pins are invalid for collection. Analysis binds every runtime source/schema file, dependency lock, source runner and split/template digest. Source changes invalidate these pins. The fixed protocol timestamp is a version marker; the separate immutable operator record establishes the actual review/approval time.
Both native templates are deliberately incomplete: calibration, execution identity, alias/rollback, eligibility and approval pins stay null until supplied from real immutable records. The collection approval remains `approved: false` and is invalid for dispatch. Diagnostic collection needs no calibration artifact or calibration digest. The disabled native templates describe a future qualified workflow and must remain incomplete under this scope. A future calibrated study needs a new preregistration and genuine D09 evidence from calibration data only; never invent calibrated flags, fixture artifact pins or eligibility records.
Preregistered analysis
The closed v1 protocol fixes two-sided 95% intervals. All native comparisons use candidate minus champion and require 1,200 complete test pairs:
| Metric | Native point bound | Additional study gate |
|---|---|---|
| Quality | At least −0.03 | Newcombe-10 lower bound at least −300 bps; positive lower bound for a benefit claim |
| Calibration | Brier increase at most 0.02 | Paired percentile-bootstrap upper bound at most 0.02; seed 2611, 20,000 resamples; ECE reported separately |
| Risk-coverage | Accepted-error increase at most 0.01 | Wilson accepted-error upper bound at most 5%; coverage lower bound at least 60% |
| Abstention | Increase at most 0.05 | Paired interval upper bound at most 0.05 |
| Latency | Mean increase at most 10,000 ms | Paired distribution and p95 increase; p95 cap +15,000 ms |
| Tokens | Mean increase at most 8,000 | Retry overhead included; unknown usage remains missing |
| Cost | Mean increase at most 800 USD micros | Conservative reservation prices, including failures |
| Slice | Mean quality delta at least −0.03 | Four actual slice tables, 300 per slice, harm flagged |
Slice minimums establish support, not per-slice −3-point non-inferiority or rare-error certification. Abstentions fail unconditional correctness. Missing required endpoints block native reporting. Reports retain failure records, complete-case diagnostics and a worst-case failure-as-error analysis; they never substitute fabricated distributions or zero usage. Bootstrap intervals are deterministic conditional on the frozen paired data. Latency is the sum of collector attempt durations within each arm, including failed retries; it excludes inter-dispatch pacing and journal persistence. The p95 comparison is for that declared active-duration measure, not end-to-end session wall time.
The scorer uses native ensemble aggregation, paired helpers and the existing ensemble report builder. The source study wrapper reproduces the prepared pins before scoring. The shared `scoreHeldoutStudy` boundary rereads evidence and validates externally digest-bound approval, scorer, gold and upstream integrity. `buildD17NativeReport` also needs independently anchored report and integrity digests plus a real champion/challenger record bound to the baseline and statistical evidence. Scoring is a trusted host/library operation; the collector CLI collects observations, not promotion approval. Replay a completed protected run locally with:
nice -n 19 node tools/decision/d17-score.mjs SCORE_CONFIG.json OUTPUT.jsonThe closed config contains `run`, `trustedEvidenceDigest`, `trustedApprovalDigest`, `goldFile`, `integrityFile` and `trustedIntegrityDigest`. Supply the trusted digests from protected external records; do not derive a new trusted pin from untrusted edited evidence. The output must reside within the routed artifact root and is exclusively created. This command performs no collection or credential resolution.
Economics compares candidate reservations with champion reservations and reports net savings with its sign. Extra ensemble cost normally makes net savings negative. D17 preregisters an additional-cost tolerance, not a savings claim: an intended operational benefit additionally needs a positive lower quality bound and a separately accepted cost/benefit tradeoff. A statistical pass alone does not produce PROMOTE. The Brier, ECE and risk/coverage calculations are descriptive diagnostics, not D09 qualification or calibrated gates. D09 compatibility, blind review, protected integrity evidence and separate operational approval remain external inputs for a future qualified study; they cannot enable promotion under this scope.
Staged D09 calibration
`src/decision/ensemble-study/staged.ts` reuses the diagnostic corpus, gold, split manifest, review template and native templates unchanged; the corpus module and its corpus pin do not move. Only these differ:
- Preregistration `calibration: { scope: 'calibrated', allowedModes: ['staged'],
calibrationPhaseSplits: ['tuning', 'calibration'] }`, bound to a `decision-d17-analysis/v2` analysis.
- The v2 analysis adds the frozen `D17_CALIBRATION` protocol
(`schemas/decision/D17StudyCalibration.v1.schema.json`).
- The approval form starts at `{ mode: 'staged', phase: 'calibration' }`.
- The dry run (`decision-d17-dry-run/v2`) splits the same 7,200 first attempts
into a 2,400-attempt calibration phase (600 rows) and a 4,800-attempt test phase (1,200 rows).
The v1 schemas are unchanged; the staged artifacts use new schema versions (`D17StudyAnalysis.v2`, `D17StudyDryRun.v2`, `D17StudyReport.v2` and `D17StudyCalibration.v1`).
Calibrators
Two calibrators are fitted on calibration-split rows only. Tuning rows are collected and sealed with the phase but never fitted; test rows are outside the calibration phase and the collector refuses them. The library fit also refuses any attempt pinned to another study, corpus or preregistration.
| Calibrator | Input | Fitted on | Applied to |
|---|---|---|---|
| `d17-single-call-isotonic` (member) | One fresh call's native P(yes) | All four single calls of every calibration row | The champion |
| `d17-three-sample-mean-isotonic` (aggregate) | Raw mean P(yes) of the three ensemble members | One mean per calibration row | The challenger |
Both use `isotonic-pav-laplace-v1`: pool-adjacent-violators on the observed frequencies, blocks of at least 10 observations, Laplace-smoothed block frequencies `(yes + 1) / (n + 2)` and a final monotone pass. The result is a non-decreasing step function. Fitting requires at least 380 observed rows for each calibrator (the 5% failure tolerance of 400).
At test time a calibrated probability of exactly 0.5 defers. The challenger is accepted only when the native aggregate accepts (three members, Jensen–Shannon disagreement, tie rule) and its calibrated value is not a tie.
Qualification metrics are out-of-fold
An isotonic fit is nearly perfectly calibrated on the rows it was fitted on, even for uninformative scores, so in-sample ECE cannot gate anything. The deployed mapping is fitted on the full calibration split, but its qualification metrics are cross-fitted:
- Calibration rows are assigned to 5 folds, stratified by slice: record-ID order
within each slice, index modulo 5. The assignment is deterministic.
- Each row is scored by calibrators fitted on the other four folds.
- ECE, selective risk and the Wilson intervals are computed on those
out-of-fold probabilities.
Each mapping records both the out-of-fold and the in-sample metrics. The `CalibrationArtifact.v1` metrics, and therefore D09 qualification, use the out-of-fold numbers.
The artifact profile (at least 380 samples, 95 per slice, Wilson 95%, ECE at most 0.1, selective-risk Wilson upper bound at most 0.1, 30-day expiry) is frozen in the analysis. The approval carries one `calibrationArtifactDigest` slot, so it binds the digest of a `decision-d17-calibration-set/v1` record that pins both artifacts, both mapping digests, the phase seal and the calibration-phase approval.
The scorer re-derives the calibration set
Test scoring trusts no registered file. Before any row is scored, the staged option:
1. Re-reads the calibration phase through the verified collector seal. 2. Re-fits both calibrators deterministically, recomputes the out-of-fold metrics, and re-applies the anchored development and calibration reviews. 3. Requires the resulting set digest to equal the approved `calibrationArtifactDigest`. 4. Requires the registered set, artifacts and mappings to be canonically identical to that re-derivation.
A mapping fitted on other observations, an artifact hand-marked approved, or a set without its calibration review is therefore refused.
The scoring clock (`evaluatedAt`) cannot be in the future. It must also be no earlier than the calibration review and no earlier than the collector's recorded first test access (`frozen.testPhaseAccessAt`). Both artifacts must qualify at that clock. The D09 registry also refuses any resolution before an artifact's `effectiveAt`.
Development review precondition
The 40 development assessments must be complete before calibrated work can proceed: reviewer, preregistration review, review time, rationale, and a gold audit label that agrees with both the generator gold and the independent text oracle. Any `goldAmbiguousOrIncorrect: true` refuses and requires a revised generator and a new holdout. Blind-test and repeat assessments must still be blank at that point.
The calibration review's `reviewer` must be the reviewer the approval names, and it is bound into each approved artifact's approval reference, so the same reference and time under another reviewer produce a different calibration set and are refused.
Approvals carry no timestamp. So ordering is proved by digest: the calibration-phase approval's `approvalReference` must cite the completed review digest. The review therefore existed before the approval that authorized any spend. The review and this binding are enforced against the operator's anchored review digest at four points:
1. `d17-study.mjs the bundle option` assembles the calibration-phase bundle only with a completed review that the approval cites. 2. the fit-calibration option additionally requires every development review time to precede the phase seal. 3. the register-calibration option re-checks the review. 4. Staged test scoring re-checks it before scoring any row.
The shared collector cannot enforce the review itself. It is study-agnostic and its `decision-heldout-approval/v1` schema is closed. A hand-assembled bundle can therefore still collect calibration-phase observations, but without the cited review no calibration can be fitted or registered and no calibrated report can be produced.
Seeds, ledgers and both-phase budgets
The staged corpus is the diagnostic corpus for the same seed. A seed whose corpus already has observations under the diagnostic preregistration cannot be reused:
- the collector refuses it (`changed-preregistration`);
- the dry-run-phases option refuses it before any spend.
Always prepare a staged study from a fresh, uninspected seed.
Both phases are checked before the calibration phase spends anything:
- the bundle option checks the combined worst case of both phases against the
approval's allowance on a fresh ledger.
- `the dry-run-phases option BUNDLE APPROVAL_DIGEST ARTIFACT_ROOT` checks it against the
real ledger, after every prior charge.
- the dry-run-staged option prints the computed per-phase plan.
At the current request sizes and the USD 0.10/1M reservation floor:
| Phase | Worst-case attempts | Reserved tokens | Reserved USD |
|---|---|---|---|
| Calibration (600 rows) | 4,800 | 6,493,848 | 0.528776 |
| Test (1,200 rows) | 9,600 | 12,997,056 | 1.058248 |
| Both | 14,400 | 19,490,904 | 1.587024 |
The staged approval template carries the operator-authorized budget of 24,000 calls, 96M tokens and USD 8; the USD 8 study cap is unchanged. On a fresh ledger the allowance at the 80% stop is 19,200 calls, 76.8M tokens and USD 6.4.
The combined worst case fits with 4,800 calls, 57.3M tokens and USD 4.81 to spare. Attesting the USD 0.099052 already spent on the retired diagnostic corpus as the study floor leaves USD 4.73. The previous 18,000-call budget left zero call headroom.
A test-phase bundle additionally needs the calibration context (calibration run, registered directory and calibration review). the bundle option re-derives the set and requires it to be reviewed, qualified and unexpired at assembly time.
Gates and the promotion route
D17 has no `GateBinding`. The D29 gate-pack pattern does not apply, and the calibrated report uses the frozen D17 native protocol (the table above), evaluated on the calibrated champion and challenger measurements (`gates.source: 'd17-native-protocol'`, `gateBinding: null`). The uncalibrated v1 statistics are retained alongside as `uncalibratedStatistics`, a descriptive comparison only.
The v2 report carries `calibrated: true`, `d09Qualified: true` and `calibratedGate: true`. These mean that both re-derived, approved artifacts resolved as `allow` in the D09 registry at the scoring clock and stayed within the frozen profile and the out-of-fold selective-risk interval. The collector's `scoreHeldoutStudy` wrapper reflects this at its top level only when the scorer names the exact approved calibration binding; other studies, including D29, keep the collector's `not-performed` default.
The study report's decision stays HOLD or a preserved ROLLBACK. Promotion is a separate, preregistered route (`D17_CALIBRATION.promotion`):
1. the native-handoff option derives the eval-integrity gate from the locked artifact snapshot, with no HOLD cap. A PROMOTE gate survives only if the calibrated study gate passes, the quality lower bound is positive and an anchored `decision-d17-tradeoff-approval/v1` approves this exact report. That approval must accept at least the measured additional reservation cost. Otherwise PROMOTE is held, with each missing piece listed as a reason, and the integrity metadata binds the tradeoff digest. 2. `buildD17NativeReport` accepts an optional D09 `PromotionEligibility`, which must be canonically equal to the registry's stored record for `record.eligibilityId`. It also accepts an upstream PROMOTE only with that bound tradeoff approval. 3. `promoteChampionChallenger` moves the alias through the D09 gateway.
The diagnostic v1 report keeps refusing PROMOTE and eligibility exactly as before. The the native option CLI has no persisted D09 registry, so it serializes without eligibility (HOLD, `d09-eligibility-missing`). The host D09 gateway supplies the eligibility and performs the promotion.
Operator procedure
Use one fresh artifact root for both phases. The collector re-validates every frozen bundle in a ledger at the current source. The 2026-10-01 diagnostic ledger (`.claude/worktrees/d17-live`) was written at the pre-D29 generator digest and is refused there (`corpus-provenance`). Keep it as evidence, and attest its spend as the new approval's floors: USD 0.099052 study, and the program's current portfolio total. Every approval repeats the original budget (24,000 calls, 96M tokens, USD 8 for the staged study) and floors.
Use AIWG to complete this documented outcome: Use one fresh artifact root for both phases. The collector re-validates every frozen bundle in a ledger at the current source. The 2026-10-01 diagnostic ledger (.claude/worktrees/d17-live) was written at the pre-D29 generator digest and is
Have it inspect the current state, explain the plan, ask before material
changes, and report the result with verification evidence.The the staged option config adds `calibrationRun`, `calibrationDir`, `trustedCalibrationSetDigest`, `calibrationReviewFile`, `trustedCalibrationReviewDigest`, `developmentReviewFile`, `trustedDevelopmentReviewDigest` and `evaluatedAt` to the diagnostic keys.
the native-handoff option adds `report`, `trustedReportDigest`, `operator` (`alias`, `aliasRevision`, `eligibilityId`, `integrityReportId`, `approvalReference`, `approvedAt`), `tradeoffApprovalFile` and `trustedTradeoffApprovalDigest`. The two tradeoff fields may be null.
the native-handoff option re-scores under a locked artifact snapshot and requires the anchored report digest to reproduce. It then emits the champion/challenger record, with each role citing its D09 artifact, and the eval-integrity metadata.
Price, budget and live operator handoff
The template records the task-supplied USD 0.042 per million input-token rate, free output and a zero request fee, with these evidence references:
- https://www.eesel.ai/blog/typesafe-jev-pricing
- https://www.mindstudio.ai/blog/jev-pricing-cost-per-token
- `roctinam/aiwg#2613 comment 153093`
These are supplied attestation references, not fetched or verified live here. The operator must sign the price approval reference and confirm the fee/rates. Reservations charge projected serialized UTF-8 request bytes plus the preregistered 512-token provider overhead at the greater of the attested input rate and USD 0.10 per million. The approval attests zero output price and a zero request fee. The expected retry scenario below conservatively reserves the full 3,744-token input allowance for each call. These are planning bounds, never provider-reported charges:
| Quantity | Expected | Worst-case reservation |
|---|---|---|
| Attempts, all three splits | 7,380 (7,200 initial plus 2.5% retries) | 14,400 |
| Tokens | 27,630,720 reserved input; output usage unknown | 57.6 million total, including output allowance |
| USD | 2.7675 | 5.4 |
Approval ceilings are 18,000 calls, 72 million tokens and USD 8 study spend. The USD 48 portfolio cap and scanned/attested earlier spend also apply. The worst case fits the collector's 80% stop threshold when the full budget remains. Each request reserves before dispatch, with no refund. The 4,000-token bound includes output/hidden allowance; conservative projection estimates are not measured token usage. Thirty-minute sequential sessions require checkpoints and resumptions. No soak or concurrent load is authorized.
Complete the approval with a clean exact source commit, matching CI evidence, real reviewer/reference, titan workspace, region, scoped resolver reference and byte digest, diagnostic calibration mode, execution digest, provider terms and actual prior study/portfolio spend. Bind the completed approval digest in an external immutable record. Assemble the collector's closed bundle from only `corpus`, `preregistration` and that approval. Its digest is canonical JSON via `heldoutDigest`, not a formatted-file hash. Then the operator can run:
nice -n 19 node tools/decision/heldout-study.mjs \
the dry-run option BUNDLE.json APPROVAL_DIGEST ARTIFACT_ROOT
AIWG_DECISION_HELDOUT_LIVE=1 nice -n 19 node tools/decision/heldout-study.mjs \
the collect-approved option BUNDLE.json APPROVAL_DIGEST ARTIFACT_ROOTThe second command is an operator handoff, not authorization from this document. Approval must precede any spend. The public-seed dry-run emits this placeholder approval text; a paid run needs the form generated from its fresh pins:
I, roctinam, approve D17 synthetic-only UNCALIBRATED diagnostic preregistration PREREGISTRATION_DIGEST and the separately completed priced approval digest APPROVAL_DIGEST, with USD 8 study/USD 48 portfolio caps and the frozen 88-assessment review protocol; no D09 qualification, calibrated gates or promotion are authorized.
Operator review and remaining evidence
The deterministic review template supplies exactly 88 assessments:
- 40 development gold/guide checks, ten per slice, before freeze.
- 40 stratified blind test gold/result audits, ten per slice, after collection.
- Eight delayed repeat audits, two per slice, with prior answers hidden.
First audit test gold with model outputs hidden; then inspect blinded results. Record labels, ambiguity/correctness flags, rationale and review time. An ambiguous or incorrect gold item invalidates the affected preregistered analysis. Never quietly relabel, replace or remove it: revise the generator and use a new independent holdout. One reviewer supplies intra-rater repeats, not independent or inter-rater agreement. Also review the preregistration and final disposition; neither artifact review is one of the 88 assessments.
Offline tests establish deterministic generation, isolation, transport/budget failure handling, statistical/report gates and the staged path: phase sealing, calibration-only fitting with out-of-fold qualification, review-gated registration, D09 qualification, seal-re-derived calibrated scoring, the AC7/AC14 native record and the D09 promotion route. AC7's measured ensemble report and AC14's live integrity report remain pending a staged live run: actual Jev observations for both phases, the operator-reviewed calibration set, externally anchored integrity and completed blind review are still missing. The diagnostic collection cannot close AC7/AC14 or qualify calibrated gates. Production risk-tier benefit, independent-provider ensembles, operational drift/rollback, actual promotion and rollout remain open. These synthetic worlds cannot establish representative production performance.
Offline handoff snapshot
For seed `d17-2611-v1`, the source-only dry run made zero provider calls and produced these canonical JSON pins. These identify proposed artifacts, not an operator approval, durable freeze or collected evaluation:
| Artifact | Digest |
|---|---|
| Corpus | `sha256:e268bf40c8f03d3b2b30a9919520705535bce6104b9fd892412388d20d2216b7` |
| Preregistration | source-derived: emitted by the dry-run option at the approved source commit |
| Unapproved priced template | source-derived: emitted by the dry-run option at the approved source commit |
| Split manifest | `sha256:09e2934e06781d8d64c65d104b8ee72ee87f3c48952eb2fe9333bb345862540f` |
| Analysis | source-derived: emitted by the dry-run option at the approved source commit |
| Staged preregistration and v2 analysis | source-derived: emitted by the dry-run-staged option at the approved source commit |
| Private gold | `sha256:865f5f28321f93a10be477a2b9095cbb2fa97cb3f5ee33b8f0fcd66b089d68db` |
Corpus rows, split and gold are data pins and stay fixed, and are identical under both scopes. The corpus digest moves only through `provenance.generatorDigest`, which pins the D17 corpus module bytes; the staged scope lives in a separate module so that it does not move this pin. Preregistration, the priced template and analysis embed the digest of every source byte the study runs, so they change with any shared decision-source change; they are bound at approval time to the exact approved source commit, and the test suite checks that they are well-formed and reproducible rather than pinning stale values. This source-only dry run contains no live observations.
Pins will be regenerated at the new source commit before any operator freeze or approval; the source dry-run values above were regenerated from the revised source and identify only this offline demonstration. Recompute after any source/schema change. Complete and attest the approval separately; its digest will differ from the unapproved template. The projected maximum was 1,179 reserved input tokens, including the 512-token provider overhead; a further 256-token output/hidden allowance fits under the 4,000-token bound. This is not observed provider usage.
The workspace artifact router currently resolves to the worktree's `.aiwg` directory. No durable study was written there. Configure/verify the intended protected artifact store before preparation and preserve the external approval record, source CI evidence and actual remaining budget before collection. The shared collector now anchors study and portfolio baselines to a durable hash-chained spend counter and independent head. A missing or altered counter requires operator reconciliation; this source-only dry run creates no ledger.