Release v0.7.0: build, evaluate, and improve decision-model systems

[24601]

Why 0.7.0

Augustus equips agents to find, build, evaluate and iteratively improve decision-model systems, including Software 3.0 workflows and prompt/program hill climbing. This release makes that capability explicit and operational instead of treating the skill as a survey or design-only guide. The earlier 0.6.1 patch plan was not published.

Changes

• Add typed composition contracts and conditional laws for branches/cascades, dependence-safe risk budgets, bounded-loss regret, information substitution and feedback.
• Add an executable paired-workflow outcome comparator: descriptive search versus fixed-sample confirmation, constraint observations, explicit unknown cost, no deployment side effects.
• Add a bounded build/eval/improvement loop, frozen incumbent and confirmation discipline, outcome capture and rollback.
• Complete the since-last-refresh GitHub/Hub/web research and all requested Jev-Omni, decision-model-testing, pijev and Jevusers sources. Directory traversal, metadata inspection, primary review and reproduction remain distinct.
• Re-derive all earlier delegated scopes under requested GPT-6 Astra/xhigh routing. Preserve historical reports and independently review actual new answers; requested routing is not backend attestation.
• Fix evaluator/collector/parser/checker edge cases and strengthen future research/release guidance.
• Align README, skill activation metadata, Codex UI, marketplace, changelog and Pages positioning. Record remaining scheduler and external-directory gaps honestly.

Evidence already run

• make check: 106 tests plus repository, numerical/fingerprint and shell checks; local Python 3.11.16, 3.12.13 and 3.14.7.
• Claude marketplace validation; production Pages build with github-pages 232/Jekyll 3.10.0 and rendered-site checks.
• 40 fresh actual answers across staged runtime snapshots, covering 31 catalog scenarios plus two additional requests; final 0.7.0 forward exercise built and executed a paired eval. Synthetic challenger lost, and unknown costs stayed unknown.
• Exact finite mathematical counterexamples and 18,225 bounded-regret configurations, not provider or product-benefit measurements.
• GitHub About read-back aligned to the mission; homepage/topics unchanged.
• Heavy read-only adversarial review of initial 9b4d86d (requested GPT-6 Astra/ultra) returned fix-first: finite workflow mean overflow/underflow, selective-cost overflow, and log-loss cancellation. All three corrected with eight additional regression tests, including exact-ratio randomized comparisons, schema mutations and fixed-sample null enumeration. Fresh independent approval of the corrected revision is still required.
• A second fresh Astra/ultra review rejected 977059f for early normalization erasing paired deltas. Exact paired accumulation and two further tests corrected it in 02f379a, then reviewed by Claude Code Fable 5.1/xhigh/YOLO at the user's request. Startup reported claude-fable-5-1 and bypassPermissions; effective effort is unobservable.
• That Fable review completed with ship (complete report). The coordinator addressed its six follow-up notes before release: strict-bound/report consistency, stable Brier/ECE aggregation, evidence-preserving violation labels, declared exclusions, build/improve activation examples and historical harness provenance. Candidate addf6c5 includes eight more regressions. A fresh final Fable 5.1/xhigh/YOLO review returned ship on that exact candidate, with independent numerical probes and all three Python-version suites passing.
• Existing-listing audit recommends amending Hellogumbo Fold peterpme/lev (educational Kev pointer head) #76 and five description-only updates, with generated/rejected/crawler surfaces distinguished. Nothing submitted externally.

Acceptance and release gates

Completed: coordinator integrated-diff inspection and checks, fresh final Fable ship verdict, exact-candidate/tag isolated installations and Skills CLI discovery, intended-head/main CI, merge/tag/release and deployed-commit/content read-back. See the publication receipt for exact revisions and limits. Published v0.7.0 at merge ef8e035; the subsequent user-requested website redesign is separate work.

Research/acceptance: research/decision-model-review-2026-09-22.md, research/decision-engine-2026-09-22.md, and research/audits/2026-09-22-refresh-acceptance.md.
Maintainer prompt: research/prompts/maintainer.md.

No third-party benchmark reproduction, external-directory submission, or claimed deployment gain. Model-based code review is separate from provider evaluation. The external research scheduler remains unverified. Whole-session API-equivalent cost unavailable: native tools did not expose complete usage.

[24601]

Claude Code adversarial review of 02f379a

Requested Fable 5.1 / xhigh / permission bypass; observed Claude Code model label claude-fable-5-1 and bypassPermissions. Effective effort and backend not independently attested.

Reviewed commit and checks

Reviewed 02f379a6af9542a28547619f1e0c5f920f7223e5 on branch refresh/2026-09-22, clean tree before and after review. I read AGENTS.md, CONTRIBUTING.md, SKILL.md, both helper scripts, the full 0.6.0-to-candidate diff, the new tests, the adversarial and acceptance audits, the engine-forward artifact, maintenance/listing/release-checklist docs, and the discovery-data schema. Nothing in the checkout was modified; temporary fixtures lived under mktemp -d and were deleted.

Checks actually run:

Check	Result
make check on Python 3.14.7	98 tests pass, both self-tests pass, shell syntax pass
Runtime manifest SHA-256 over 22 files	matches stated d2fac49f…
Test count at 0.6.0 base / candidate	53 / 98, as claimed
Scenario count at base / candidate	12 / 31, as claimed
Behavioral answers across four artifacts	19 + 11 + 4 + 6 = 40, as claimed
Jevusers inventory	400 tracked, 1,299 apps, 1,519 distinct names, as claimed
Engine-forward receipt re-run through shipped comparator	byte-identical to recorded output
Random paired-delta oracle (2,000 cases incl. subnormal/max)	0 mismatches
Random weighted-cost oracle (3,000 cases)	0 mismatches
Three-point-loss null enumeration, n=12, alpha .05	rejection rate 1.4e-4, under budget
CLI malformed inputs (BOM, NaN, 1e400, deep nesting, empty, null, directory)	explicit exit 2, no tracebacks

I independently re-derived the Hoeffding radius, the regret bound, the union bound, the XOR side-information example, the zero-error exact bound, and the Jensen claim. All match the runtime text.

Findings

No blocker. The four prior Astra defects are fixed and covered by tests that I could not break. Notes follow, most material first.

• P3, boundary decision flips against exact arithmetic. compare_workflows.py:151-154 decides strict_margin_supported in normalized space but reports upper_mean_loss_delta in loss units. A confirm receipt with 16 pairs, loss bound 0.7, incumbent loss 0.7, candidate loss 0.10670073329327479, margin 0.1649435713375823 and alpha 0.05 reports upper equal to minus the margin yet sets the strict flag true. Exact rationals with the same float radius give upper plus margin of +6.1e-18, so the strict claim is false. This is one ulp on a bound whose slack is 0.2, so it is not a practical false acceptance, but the JSON contradicts itself. Correction: decide with Fraction(exact_delta) + Fraction(radius) * Fraction(bound) < -Fraction(margin) and derive the reported upper from that same rational. Rounding the radius up would also make the direction conservative.

• P3, Brier, reliability and ECE are order-dependent on Python 3.11. evaluate_decisions.py:130, :171-172, :181 and :200-203 use the builtin sum, which is naive on 3.11 and compensated on 3.12+. One row with loss 1 followed by 10,000 rows with loss 1e-17 gives a Brier score 755 ulps from exact on 3.11 and a different value when the rows are reversed. Constant-mean preservation also fails at one ulp on every version, for example seven rows at p=0.7. CI supports 3.11, CONTRIBUTING names permutation invariance as a criterion, and the sibling log-loss path in the same file was converted to statistics.mean. Pre-existing, not introduced here, and invisible at the CLI's four decimals. Correction: use statistics.mean or math.fsum in those four places.

• P4, assessment label loses fixture provenance. compare_workflows.py:164-167 overrides fixture_evidence_only with observed_candidate_constraint_violation when a fixture receipt lists a violation. The word "observed" is then attached to a fixture. evidence_kind remains in the output, so nothing is hidden. Correction: name it candidate_constraint_violation or keep the kind prefix.

• P4, no way to declare missing outcomes. The receipt schema at compare_workflows.py:86-115 requires a loss on both arms of every pair, so attrition can only be handled by dropping pairs or imputing the bound. paired_units silently equals the complete-case count. The docs prescribe a missing-data policy but the artifact cannot record one. Correction: an optional excluded_units integer with a limits line, no behavioral change for existing receipts.

• P4, activation reference lags the new mission. references/activation-triggers.md:1-12 still describes placement questions only, with no build, eval-harness or hill-climb examples, while the SKILL.md description and marketplace copy now lead with those. Agents route mainly on the description, so the impact is consistency rather than discoverability. Correction: add two or three build/improve prompts and one non-trigger.

• P4, historical reproduction pins the old helper. The engine-forward artifact's harness at research/audits/2026-09-22-astra-engine-forward.md:307 pins comparator hash 9955ddba…; the shipped helper is aa47d8e4…, so the embedded command will refuse to run against 0.7.0. The adversarial audit discloses this. A one-line note in the artifact pointing at the current hash would prevent confusion.

Composition-law review: the laws are stated with their preconditions and I found no counterexample within those preconditions. One sentence would help: the "free observation cannot raise Bayes risk" statement holds for a single decision-maker with a correct joint model; estimated policies in finite samples, misspecified models, and strategic settings can lose value from extra information. The text already limits scope with "common joint model," so this is a clarity note.

Statistical design: fixed-sample one-sided Hoeffding with Bonferroni over a declared family is correct as written and conservative. Leakage, adaptive reuse, family declaration, causal identification, and label validity are explicitly delegated to the caller in the docstring, the reference, and the output limits array. Those are disclosed limitations, not defects. The helper cannot verify them and does not claim to.

Process items: the shell collectors contain no git, gh, push or publish verbs; refresh.py writes only with exclusive-create mode. maintenance.md states the scheduler is unverified. The listing audit authorizes no external action and contains drafts only. Release notes say no third-party submissions were made. These match the checkout.

Limits, scope, and identity

Not verified by me: make check on 3.11 and 3.12 could not run because PyYAML is absent from those interpreters and installs were out of scope; the parent's claim stands as parent evidence. Pages build, plugin validation, install smoke, CI on the intended head, and publication read-back were not run. External primary sources were not reopened; every third-party number remains Reported, and the research documents say the same. No provider or paid run occurred. The 40 behavioral answers were produced on earlier runtime bytes; only the two helper scripts changed after the six-answer forward test, so the prose those answers exercised is unchanged.

Scope and claims: a minor bump is justified. New executable capability, a new reference section, new schema, no breaking valid-input change. Public language limits itself to capability and disclaims measured superiority, deployment gain, growth, and native activation. The "build requests produce working adapters" wording rests on one synthetic fixture build, which the notes state. I found no overclaim.

Requested versus observable identity: the user requested Claude Code Fable 5.1 at xhigh effort with permission bypass. My environment reports model ID claude-fable-5-1. I cannot observe the effective effort setting or attest the backend beyond that string; treat the model ID as a runtime label, not independent attestation.

ASTRA REVIEW
VERDICT: ship
REASON: All four previously rejected numerical behaviors are fixed and survive independent exact-rational, randomized, and enumerated null probes; no new material defect in arithmetic, statistics, schema handling, composition laws, process authority, or public claims. Remaining findings are ulp-level or consistency notes suitable for a 0.7.1 patch.
FINDINGS: P3 strict-margin flag can flip against exact arithmetic at a one-ulp boundary (compare_workflows.py:151-154); P3 Brier/reliability/ECE order-dependent on Python 3.11 and not constant-mean exact (evaluate_decisions.py:130,171-172,181,200-203); P4 fixture violation label drops provenance; P4 no missing-outcome marker in receipt schema; P4 activation-triggers.md lags the new mission wording; P4 engine-forward harness pins the superseded helper hash.
RESIDUAL RISK: Hoeffding preconditions and confirmation isolation remain caller obligations the helper cannot check; 3.11/3.12 suite, Pages build, plugin validation, exact-tag install and publication read-back not verified in this review; behavioral evidence predates the final helper bytes; third-party results remain Reported; effective reviewer effort and backend not independently observable.

CLI completion metadata

{
  "type": "result",
  "subtype": "success",
  "is_error": false,
  "duration_ms": 760825,
  "num_turns": 48,
  "total_cost_usd": 7.5074775,
  "usage": {
    "input_tokens": 354,
    "cache_creation_input_tokens": 211790,
    "cache_read_input_tokens": 1114750,
    "output_tokens": 59789,
    "output_tokens_details": {
      "thinking_tokens": 38076
    },
    "server_tool_use": {
      "web_search_requests": 0,
      "web_fetch_requests": 0
    },
    "service_tier": "standard",
    "cache_creation": {
      "ephemeral_1h_input_tokens": 211790,
      "ephemeral_5m_input_tokens": 0
    },
    "inference_geo": "not_available",
    "iterations": [
      {
        "input_tokens": 32,
        "output_tokens": 3893,
        "cache_read_input_tokens": 213241,
        "cache_creation_input_tokens": 2506,
        "cache_creation": {
          "ephemeral_5m_input_tokens": 0,
          "ephemeral_1h_input_tokens": 2506
        },
        "type": "message"
      }
    ],
    "speed": "standard"
  },
  "modelUsage": {
    "claude-fable-5-1": {
      "inputTokens": 354,
      "outputTokens": 59789,
      "cacheReadInputTokens": 1114750,
      "cacheCreationInputTokens": 211790,
      "webSearchRequests": 0,
      "costUSD": 7.5074775,
      "contextWindow": 1000000,
      "maxOutputTokens": 64000,
      "thinkingTokens": 38076,
      "canonicalModel": "claude-fable-5-1",
      "provider": "firstParty",
      "costBasis": "list"
    }
  },
  "session_id": "69cefb52-f4c1-404b-b3ae-57c31d34f565",
  "permission_denials": []
}

[24601]

Final Fable acceptance — addf6c5

Requested Claude Code Fable 5.1 / xhigh / permission bypass. Observed startup claude-fable-5-1, bypassPermissions, no MCP servers; response labels agree. The effective effort/backend are not independently attested. Read-only was the task instruction, not a filesystem sandbox. No review edits were made.

Reviewed commit: addf6c5cb073a4019b28edf1b7fa650fd5be22d7, clean tree, no local v0.7.0 tag. Checkout was not modified. My scratch directory under /tmp was removed after use.

Requested vs observable model/effort: the session reports model claude-fable-5-1. Effort xhigh and backend were not independently observable. Permission bypass was never exercised; every action was a read or an in-memory/scratch probe.

What I ran

Check	Result
make check on Python 3.14.7, 3.11.16, 3.12.13	106 tests OK each, both self-tests OK, shell syntax OK
claude plugin validate .	passed
22-file runtime manifest hash, paths relative to the skill directory	aba7eaa2… matches
Manifest at 02f379a stage via git archive	d2fac49f… matches audit
Helper SHA-256 for both scripts	match the audit values
Historical harness pin 9955ddba…	equals compare helper at 9b4d86d
Test counts at 192faf0 / 02f379a / HEAD	53 / 98 / 106, matching public claims

Independent probes, run on all three interpreters:

• Historical receipt replay. The embedded outcomes block plus a trailing newline hashes to the recorded 614c6e1c…. The 9b4d86d helper reproduces the recorded stdout byte for byte. The HEAD helper output differs only by two added null fields and one added limits line. No prior data field changed.
• Strict-margin repro. On the 02f379a code the 16-pair case reports upper == -margin with strict_margin_supported = true. On HEAD it reports an upper one ulp above minus the margin and false.
• Randomized confirmation oracle, 6,000 receipts over bounds from 5e-324 to float max, alpha from 5e-324 to the largest float below 1, comparison counts up to 10**400, n from 1 to 39, including all-extreme and one-ulp cancellation deltas. Zero flag/report mismatches. Reported upper always at or above the exact rational value and never more than one ulp above it. Every supported case implied exact strict improvement, and setting the margin to the negated reported upper always cleared support. Output was permutation invariant and JSON-serializable. The same inputs produced 9 contradictions on the 02f379a code, so the old defect was reachable on ordinary data.
• Radius precision census against a 60-digit Decimal oracle: worst relative error 0.78 ulp, and the float radius fell below the true value in 219 of 392 cases. That is the disclosed binary64 limitation and it is sub-ulp, not a coverage problem in practice.
• Attrition metadata. All 15 malformed values rejected, including booleans, negative, float zero, string digits, empty and whitespace-only text. Big integers accepted and serializable. Declarations never altered the confirmation or delta.
• Provenance labels. All four evidence kinds survive the candidate-violation path in both phases. An incumbent-only violation does not mislabel the candidate.
• evaluate_decisions. 3,000 random row sets with subnormals, boundary values and ten-bin ties, six shuffles each: Brier, reliability, equal-width ECE and quantile ECE were permutation invariant and equal to exact rational oracles on Python 3.11, where the old code was order dependent. Bin assignment at k/10 and k*0.1 boundaries is unchanged. Extreme and subnormal cost paths through sweep, always_negative and selective_policy match the rational oracle. CLI runs end to end.

Fix-by-fix assessment

1. Strict bound. Correct and conservative. Rounding toward plus infinity moves a negative upper toward zero, which can only withhold support. Capping at the bound keeps float max and huge-radius cases finite. Hoeffding range 2B, one-sided tail, Bonferroni via log form: derivation checks. Fixed-sample interpretation is stated in the code, the reference and the output limits.
2. Means. Exact-ratio accumulation, order stable, constant means preserved. Tie and bin semantics untouched.
3. Labels. Retained through failure paths; the observed-kind string is unchanged.
4. Attrition fields. Honest wording in docstring, reference, output limits and release notes: declarations only, null means unknown.
5. References. Activation examples now cover build and hill-climb, with a refactor non-trigger. The information-value law now states single decision maker and correct model assumptions, which is the actual precondition for Blackwell monotonicity.
6. Historical harness. The pin paragraph is accurate and no old run is relabeled.

Findings

No material defect. Two low, non-blocking observations:

• Integer-typed bucket rate. statistics.mean preserves int input, so reliability() now returns Python int for rate when a bucket is all zeros or all ones, where 02f379a returned float. The CLI formats it identically and load() keeps mean_p float. Only programmatic JSON consumers would see 1 instead of 1.0. Worth a one-line note or a float() wrap at some point, not a release blocker.
• Reference budget headroom. The references total 179,459 bytes against a 180,000-byte ceiling. Any future addition forces consolidation. This is a maintenance note, not a defect.

Public claims in README, CHANGELOG, release notes, marketplace and citation are version-aligned at 0.7.0 and match what I measured. The README clone command and release-note links point at a v0.7.0 tag that does not exist yet; they become live only after the parent tags.

Not run by me

Jekyll/Pages build and check_site.py, plugin install into a temporary config, CI, external research sources, provider calls, and scheduler behavior. I did not reread the archived metadata rows.

ASTRA REVIEW
VERDICT: ship
REASON: All six follow-ups are correctly implemented and independently reproduced on Python 3.11, 3.12 and 3.14 with exact-rational and high-precision oracles across the declared input range; the historical replay, hashes, test counts and public claims are consistent; no new defect was found beyond the stated fixes.
FINDINGS: P4 reliability() bucket rate can now be Python int for all-0 or all-1 buckets, invisible in CLI output. P4 references sit 541 bytes below the 180,000-byte ceiling.
RESIDUAL RISK: Radius uses uncertified binary64 transcendentals with sub-ulp error; evidence is 40 earlier staged answers rather than independent benchmarks, and three references plus both helpers changed after the last behavioral exercise; no provider or deployed-benefit measurement; Pages build, install smoke, CI, tag existence and publication read-back remain the parent's gates.

Coordinator acceptance

Accepted for release after personally inspecting the full accumulated changes and corrections; running 106 tests on Python 3.11.16, 3.12.13 and 3.14.7; matched Pages build/rendered checks; plugin validation; isolated exact-candidate installation and one-skill/no-other-components read-back; installed helper execution from /tmp; and intended-head CI.

The remaining rate note changes numeric representation (1 versus 1.0), not the value or CLI output; no strict Python float-type contract is documented. The 179,459-byte reference total remains below the enforced 180,000-byte ceiling: future additions must consolidate, not raise the budget. Neither note is a release blocker. Binary64 transcendental precision is not a certified interval guarantee. Caller sampling/causal/confirmation assumptions remain unverified by the helper.

Frozen candidate: addf6c5cb073a4019b28edf1b7fa650fd5be22d7; Git tree 5eaf9ca3183f234ee3143c1c66c0c6ede4784c8e; 22-file runtime manifest aba7eaa218802dca982ce02f9080dc6396807f2c6b8fc10c8cd3f7414e829ba1. Earlier staged behavioral answers are not relabeled as final-byte tests. Parent owns merge/tag/remote installation/publication read-back; those will receive a separate receipt.

CLI completion metadata

{
  "type": "result",
  "subtype": "success",
  "is_error": false,
  "duration_ms": 473134,
  "num_turns": 32,
  "total_cost_usd": 4.347564500000001,
  "usage": {
    "input_tokens": 450,
    "cache_creation_input_tokens": 122723,
    "cache_read_input_tokens": 818818,
    "output_tokens": 33678,
    "output_tokens_details": {
      "thinking_tokens": 15921
    },
    "server_tool_use": {
      "web_search_requests": 0,
      "web_fetch_requests": 0
    },
    "service_tier": "standard",
    "cache_creation": {
      "ephemeral_1h_input_tokens": 122723,
      "ephemeral_5m_input_tokens": 0
    },
    "inference_geo": "not_available",
    "iterations": [
      {
        "input_tokens": 32,
        "output_tokens": 2602,
        "cache_read_input_tokens": 126604,
        "cache_creation_input_tokens": 2863,
        "cache_creation": {
          "ephemeral_5m_input_tokens": 0,
          "ephemeral_1h_input_tokens": 2863
        },
        "type": "message"
      }
    ],
    "speed": "standard"
  },
  "modelUsage": {
    "claude-fable-5-1": {
      "inputTokens": 450,
      "outputTokens": 33678,
      "cacheReadInputTokens": 818818,
      "cacheCreationInputTokens": 122723,
      "webSearchRequests": 0,
      "costUSD": 4.347564500000001,
      "contextWindow": 1000000,
      "maxOutputTokens": 64000,
      "thinkingTokens": 15921,
      "canonicalModel": "claude-fable-5-1",
      "provider": "firstParty",
      "costBasis": "list"
    }
  },
  "session_id": "4d96b9da-9369-4faa-a428-f4cfac40610d",
  "permission_denials": []
}

API-EQUIVALENT COST RECEIPT: the two disjoint Fable CLI sessions report USD 11.855042 at their reported list-price basis. This is delegated-review-only, not the whole task, subscription charges or all-Astra savings. Native parent/Astra usage and supported Fable/cache-write repricing are unavailable. No same-token comparison is claimed.

[24601]

v0.7.0 publication receipt — 2026-09-22

Published Augustus v0.7.0
at 18:13:12 UTC. It is the latest release, not a draft or prerelease. The
planned 0.6.1 patch was not published; no existing tag was moved.

Exact revisions and acceptance

• Reviewed candidate: addf6c5cb073a4019b28edf1b7fa650fd5be22d7.
• PR Release v0.7.0: build, evaluate, and improve decision-model systems #91 merged at 18:09:36 UTC as ef8e03596ef97a75c7fd8998b93e9473bd891a6c.
• Candidate and merge have identical Git tree 5eaf9ca3183f234ee3143c1c66c0c6ede4784c8e; an exact diff is empty.
• Annotated tag object: 09079550382f53234aef428ee067370e58d0d72f; v0.7.0^{commit} resolves to that merge commit, verified locally and through GitHub's tag API.
• Final 22-file runtime manifest SHA-256: aba7eaa218802dca982ce02f9080dc6396807f2c6b8fc10c8cd3f7414e829ba1.
• Final independent Fable ship verdict and coordinator acceptance, following two Astra fix-first rounds and the first Fable review. All six first-Fable follow-ups were addressed before the final review. Its two remaining P4 notes are documented in the acceptance receipt.

The earlier delegated-scope replacement ledger is complete; historical reports,
40 staged behavioral answers and their hashes remain historical. Final review
and executable checks do not relabel earlier answers as final-byte evidence.

Verification

• Parent and final Fable independently passed 106 offline tests plus self-tests/shell checks on Python 3.11.16, 3.12.13 and 3.14.7. Parent inspected accumulated diffs and ran plugin validation and the matched production Pages build/rendered-site checker.
• Intended-head PR CI passed. Main Quality, Pages checks and Scorecard passed.
• A fresh HTTPS clone of the published tag resolves to the intended merge and runtime manifest. Exact-tag Claude installation in an isolated configuration passed: version 0.7.0, one skill, zero agents/hooks/MCP/LSP servers. Installed runtime files exactly match the clean tag; both offline helpers execute independently from /tmp with python3 -I.
• The advertised remote Claude marketplace/install commands passed in a different isolated configuration. That installed runtime also exactly matches the tag. Normal user plugin settings were not modified.
• Skills CLI --list passed against both the exact-tag checkout and 24601/Augustus, discovering one correctly described skill; no Skills CLI installation or implicit activation is claimed.
• An initial tag-local discovery attempt ran before its asynchronous clone finished and failed with missing files. Both local checks were repeated after clone completion and passed; the premature checks are not counted as package evidence.

Published website and discovery surfaces

Pages deployment
completed successfully. The Pages build API reports built at the exact merge
commit. Direct HTTPS read-back of the homepage and 0.7.0 notes matches the
locally checked production output byte for byte:

index.html                    1c1293c348709f64da2a41a4ed06811c5a1e8c0a32f96b41c9cf6066dda1bf84
release-notes-v0.7.0.html       716ec3a0a23766d2f1e245f0b8203a86b50225250b68fcaf93907c6b8ae448bc

GitHub About read-back matches the find/build/evaluate/improve mission;
homepage and 20 topics remain correct. Existing-listing positioning drafts are
recorded in the audit, not submitted. Scheduler liveness remains unverified.
No provider benchmark, user-growth, native activation or deployment-benefit
measurement is claimed. Historical pre-publication records are closed by this
receipt, not rewritten to impersonate deployment evidence.

The user has now requested a separate Stitch/OpenDesign-led website redesign;
that work starts after this release and is not part of the frozen 0.7.0 tree.

API-EQUIVALENT COST RECEIPT: two disjoint Fable review sessions reported
USD 11.855042 total on the CLI's list-price basis. This is review-only coverage,
not subscription charges. Whole-task native usage and supported same-token
Astra repricing are unavailable; no savings claim is made.