Conversation
…tegration Add file-explorer entry points so users can start DAAF without typing terminal commands, as thin shims over the existing control panel scripts: - scripts/host/daaf.bat (Windows): cd-to-self + same-console PowerShell handoff to daaf.ps1 with per-process ExecutionPolicy bypass; window held open on any nonzero/negative exit; UNC-path guard; committed with CRLF bytes via a new *.bat -text gitattributes rule (install.ps1 serves the raw blob, so the committed bytes must be CRLF) - scripts/host/DAAF.command (macOS, mode 100755): Bash-3.2-safe cd-to-self + exec bash daaf.sh, with pre-handoff guards only - scripts/host/daaf.ico: DAAF icon (6-size ICO, 16-256px, RGBA) - install.sh / install.ps1: place the launchers (Darwin-gated .command chmod; non-fatal icon download); install.ps1 creates a draggable in-folder DAAF.lnk shortcut (never writes to the Desktop; icon via Test-Path fallback; warn-and-continue on any failure) - update_daaf.sh / update_daaf.ps1: platform-filter whitelists so existing installs receive the launchers and icon on update, with exec-bit preservation for .command - Docs: quickstart, host README, and technical FAQ cover the per-OS double-click paths, the one-time Gatekeeper note for browser-downloaded copies, and Linux's terminal-only path (no launcher by design); terminal commands remain the universal fallback - Tests: new launcher_shims.bats (11 tests incl. a runtime handoff/env passthrough exercise) plus updater filter coverage (+4 bats, +2 Pester); lint gains a binary-asset exemption for the icon; CI bash -n gate extended to *.command; CLAUDE.md records the .command exec-bit convention and the .bat CRLF exception Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TPeEQJdorcZxyimrErNgvu
…n standard Encode a mode-agnostic rule that every substantive turn end briefs the user on all work since their last message, taken altogether — translational, action/insight-oriented, attention-economy-governed (no fixed length/item quotas). Master statement added under SKILL.md "User-Facing Communication Standards"; pointers propagated into all nine mode files following the Wave Barrier Discipline define-once pattern. User-facing paragraph added to user_reference/02. remind-orchestrator.sh GPT-gated register guidance removed (user-deployed), superseded by the model-agnostic standard. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CZmJomwUAb5AYaanHpd5t4
…, v3 billing reconciliation - models.yaml: add deepseek/deepseek-v4-flash-0731:novita/fp8 (pricing verified online 2026-08-02; wire_id OBSERVED via probe run, purity verified); retire the undated DeepSeek V4 Flash entry per the REMOVED-block precedent; new retired_model_pricing: section preserving list rates for retirees with corpus history (viewer-only, harness ignores it) - README: new 10-step "Adding a New Model" checklist (S4); billing reconciliation pipeline documentation in S7 (v1-stale/v2/v3 lineage, +/-0.26 correction rule, dated-permaslug gotcha); 2026-08-02 reconciliation snapshot changelog entry; stale example key and count re-sweeps - test_chatgpt_route.py: key-set partition updated for the add/retire; Terra/Sol now asserted against their own published tiers (registry was corrected 2026-07-29 but the test still froze Luna's copied schedule) - module green 27/27 - generate_results_viewer_v2.py: load_model_pricing() reads retired_model_pricing so retired models with corpus history keep cost estimates (30/30 priced) - viewer_template.html: Key Takeaways refocused on DeepSeek V4 Flash 0731 (name-only swaps in prose + fillTakeaways lookups) Benchmark battery (4 phases x 3 reps, zero failed runs): DC 27/36 (old Flash 22/36), MC 42/45, PC 24/27, SR 5/45. Reconciliation: obs/pred 0.639 flagged; rate correction deliberately deferred pending a second campaign. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YJgLy49vixnTBbe3w7Pat5
Existing Windows installs predate install.ps1's DAAF.lnk block and never
received an in-folder shortcut. Sync-HostScript now creates (or refreshes,
preserving user customizations like hotkeys; dragged-away copies untouched)
the DAAF.lnk after daaf.bat/daaf.ico sync, with the installer's non-fatal
warn-and-continue semantics; the absolute TargetPath also self-repairs a
stale shortcut after a folder move. Dry-run gated so CI never invokes
WScript.Shell COM (unavailable on Linux pwsh). FAQ updated: running an
update is now the easiest way to recreate a missing shortcut.
+3 static Pester tests (COM cannot be exercised on Linux; the dry-run
gating test pins branch nesting by requiring an '} else {' token between
the dry-run message and the COM call, not just textual order).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TPeEQJdorcZxyimrErNgvu
Normalize the three newest CHANGELOG version headings (v3.0.1, v3.0.0, v2.1.0) from '--' to em dashes so all six Table-of-Contents anchors resolve, and rename the v2.0.0 '### Painfully Detailed Changelog' subsection to '### Detailed Notes'.
Repo-wide cross-link sweep fixed 65 broken heading anchors: 1 in user_reference/07_faq_technical.md (renamed 'GPT Fast Mode' heading), 15 stale/typo ToC anchors across crdc/fsa/pseo/fixest/statsmodels skill references (including removing one dead 'Status Flags' entry), and 49 kramdown {#custom-id} attributes stripped from six synthetic-data-workflow references with their ToCs repointed to GitHub-computed slugs. Independently verified: 14/14 files pass, 0 residual {#id}.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HuVMk979GU4JVMoSorBtfh
… all-perfect Consistency rate (v3.7.3)
T4 previously ran on the v3.7.0 rate_agree agreement variant, which is a
predictability measure (a model failing identically every rep scores high)
and had no counterpart anywhere in the data displays. It now reads the
all-perfect consistency rate — the same metric as the leaderboard's
Consistency column — so the takeaway's claim is verifiable in the table,
and the contrast is starker (Fable 5 88.2% vs DeepSeek V4 Flash 0731
52.9%, vs 92.2%/80.4% under agreement).
- viewer_template.html: T4 prose reworked (headline narrowed to
"reliability"; mechanism sentence names the leaderboard tie-in); spans
renamed kt-t4-{topagree,budgetagree} -> kt-t4-{topcons,budgetcons};
fillTakeaways reads .rate; anchor comment restamped (supersedes the
2026-07-29 hand-edit protection per user decision)
- generate_results_viewer_v2.py: v3.7.3 changelog entry, version bump,
dev-guide/inline comments corrected (rate_agree stays in the payload
under the schema-additive contract but is display-orphaned)
- test_viewer_schema_loading.py: generator_version pin updated to 3.7.3
(was stale at 3.7.1, already failing against the committed 3.7.2)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011D99zruSSS3nVqQFmVmCQ8
…Changelog Align viewer_template.html with the 2026-08-03 site-wide navbar change on the Open Augments website: Support is now a dropdown (Overview + Changelog at daaf-product/changelog.html). Regenerated viewers no longer clobber the hand-patched website copy. Reuses the existing button-based nav-dropdown pattern; no CSS/JS changes needed. ../daaf-product/ link count is now 14, matching the website deploy check. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LB4tLa3LpWWCgemZp6hEqV
…al in 7defcc8 Commit 7defcc8 removed the GPT-gated collaborator-guidance branch from remind-orchestrator.sh (superseded by the model-agnostic Turn-End Briefing standard) but left the BATS suite asserting the removed guidance, breaking 8 tests in CI. Flip guidance assertions to refutations (regression guard: the marker string may never reappear), collapse the four guidance-positive model-matching tests into one, and drop the unused GUIDANCE_* clause variables. All 21 tests pass. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LB4tLa3LpWWCgemZp6hEqV
…th verbatim relocations Relocate the Defense-in-Depth table to BOUNDARIES.md and context-threshold profile-selection mechanics to new CONTEXT_THRESHOLDS_REFERENCE.md; backfill code-style rules, file-naming tables, and the stage-directory table into SCRIPT_EXECUTION_REFERENCE.md (previously existing only in CLAUDE.md). Condense Versioning & Naming and Shell Script Permissions to summaries with pointers. All cited headings preserved; stale cross-references updated in revision-and-extension-mode.md, FRAMEWORK_INTEGRATION_CHECKLIST.md (H3/HM6/ HR3), and ERROR_RECOVERY.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LB4tLa3LpWWCgemZp6hEqV
Use medium as the fallback response verbosity, update operator guidance, and align focused regression expectations and shim version metadata. Co-Authored-By: Claude <noreply@anthropic.com>
…ing WebFetch Add a local-first web retrieval system: scripts/web_fetch.py + web_fetch.sh (httpx + trafilatura) save raw bytes, a deterministic .md extract, and an atomic MANIFEST.jsonl provenance row (SHA-256s, redirect chain, dual byte counts) per fetch attempt, failures included. Safety guards live inside the tool: env-secret exfiltration refusal, DEST_DIR containment under /daaf, http(s)-only, size/URL caps, untrusted-content fencing with injection-shape redaction, enumerated exit codes (3 reserved for run_with_capture.sh), and graceful raw-only degradation when trafilatura is absent. New web-retrieval skill encodes the doctrine: WebSearch = discovery only, web_fetch.sh = retrieval; permitted contexts (search-agent, data-ingest, debugger, orchestrator); size rules; failure table; untrusted-content rule; recorded stances (honest UA + --browser-ua escape hatch, robots.txt not honored, hook-only enforcement). Block the built-in WebFetch via the instructive block-webfetch.sh PreToolUse hook (no permissions.deny entry, by design - the deny-rule would suppress the teaching redirect). Correct stale WebFetch guidance in deny-claude-code-guide.sh and enforce-explore-model.sh. Register the hook in settings.json. Sweep WebFetch out of the instruction surface: remove from three agents' tools frontmatter; update agent protocols, agents README, CLAUDE.md, orchestrator SKILL.md, four mode/workflow references, and the deploy-smoke T2.3 pass signal; rewrite the framework-development-mode false claim about framework-engineer web tools. Pin trafilatura 2.2.0 + extraction stack in the framework Python layer of the Dockerfile. Live-verified post-rebuild: trafilatura extraction, dedup detection, thin-extract warning, network-failure audit row, guard refusals, and the WebFetch deny hook all exercised successfully. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PcXVBetVpXGZHYVtesTRtV
Add Trafilatura (Adrien Barbaresi) to the Open Source Software list -- it powers the web-retrieval layer's content extraction -- with a pointer to the Barbaresi (2021) ACL citation, and add Barbaresi to the software authors in Methodological Foundations. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PcXVBetVpXGZHYVtesTRtV
Add a Responsible-pillar note to the provenance section: cite Barbaresi (2021, ACL-IJCNLP System Demonstrations) when retrieved web content materially supports findings, mirroring data-source skills' citation guidance. The manifest already records the trafilatura version per row. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PcXVBetVpXGZHYVtesTRtV
…ol, and Ad Hoc wiring New capability piloted in Ad Hoc Collaboration mode: user-directed analytic triangulation that turns researcher concerns into competing hypotheses, bounded discriminating tests, and evidence-calibrated interpretations in a pre-committed hypothesis/evidence ledger — behavioral, output-level verification replacing line-by-line code review. - .claude/skills/creative-verification/: SKILL.md + 4 references (12 challenge families over Shadish-Cook-Campbell validity types; sensitivity-statistic catalog with container-verified executable vs flag-and-describe split; six interpretation statuses, budgets, stopping rules; interactive-artifact QA) - agent_reference/TRIANGULATION_PROTOCOL.md: ledger schema, test lifecycle, provenance chain, pre-analysis variant, scope-change approval rules - ad-hoc-collaboration-mode.md: trigger signals, direct-load wiring, ledger duties, promotion path to Revision and Extension - README.md, user_reference 02/07: discoverability blurbs Design grounded in a deep prior-art survey (multiverse/spec-curve, validity frameworks, adversarial collaboration, graphical inference) and an observed container package inventory. Three-angle review passed; CLAUDE.md registration and Full Pipeline wiring deliberately deferred until post-pilot. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019v8Uemwff3V62fWYW9Qcta
…h revision-pinned fetches Introduce the versioned-mirror architecture: mirrors.yaml gains a vintage block pinning the 2026q3 HF mirror to immutable commit 0ad00ce0 (predecessor v0.24.0 repo frozen and pinned at 10fea9c7); fetch-patterns thread the revision through every Python/R data, discovery, and codebook URL. datasets-reference gains the vintage note, discovery-vs-fetch forward pointer, and verified year rows; vintage-drift.md documents what changed between vintages (grad-rates dedup and revaluation, cosmetic missing-code re-encoding classes). education-data-context and tidyverse io.md updated to the pinned-URL pattern. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bg1F7KtnSSUdmVn57DrwGA
…vintage (v0.26.1) Refresh the education-data-source-* family against the pinned 2026q3 mirror and ~460 live-verified data-doc claims: updated year coverages (grad-rates 1996-2023, MEPS 2.0 through SY2022-23, 2024 CCD/SAIPE/IPEDS directory years), corrected CRDC id-dtype framing (per-file heterogeneity, no universal contract), EADA opeid softening, hate-crimes crime_type roster, csafety codebook-partialness note, CCD teachers_fte API-rounding note, reproducible probe citations, and mirror-first explorer cross-reference wording. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bg1F7KtnSSUdmVn57DrwGA
…skill
Replace the API-endpoint-centric explorer (28 of 83 documented routes dead,
41 of 62 matched endpoints carrying fabricated variable names per the
2026-08-07/08 probe audit) with a mirror-first discovery layer for read-only
agents:
- SKILL.md (591->177 lines): question->mirror-file routing, grain logic,
emit-currency guidance; deletes the REST URL spine and live-fetch samples
- New routing references: colleges/districts/schools-datasets.md (85 dataset
families: path, grain, verified years, real key variables, joins, caveats)
- 14 variable-dictionary-{source}.md: mechanically generated from the Portal
varlist captured at v0.26.1 (2,235 per-source variables), zero hand
transcription
- variable-codes.md: corrected award_level scheme, disability codes, CCD sex 9,
sentinel roster (-99 in, -4/-9 out), local-filter examples, legacy-name
alias table
- maintenance-live-api.md: live-API knowledge fenced as maintenance-only, with
the four API-vs-bulk divergence classes and regeneration pointers
- Deletes *-endpoints.md and metadata-api.md (claim-by-claim migration ledger
in the session workspace)
All facts trace to probe parquets or quoted live probes; verified by three-angle
review (119/120 traceability sample) plus a functional-equivalence audit (10/10
jobs, none degraded) and a live Stage-2 smoke run.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bg1F7KtnSSUdmVn57DrwGA
…ic Dataset Identifier Fix the discovery->planning->acquisition currency mismatch: Stage-2 output and the plan-completeness gate previously spoke Portal endpoint names while acquisition consumes canonical dataset paths. The contract is now "Candidate Datasets" keyed by "Dataset Identifier" - defined generically as the identifier the domain's retrieval skill consumes - with the education mirror path kept strictly as a labeled exemplar, so non-mirror domains (e.g. Dataverse-based sources) fit the same contract. Applied consistently across WORKFLOW_PHASE1 (both contract copies), WORKFLOW_PHASE2, PLAN_TEMPLATE, full-pipeline-mode, data-discovery-mode, and research-synthesizer's section-lookup table. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bg1F7KtnSSUdmVn57DrwGA
…ion tooling Commit the scripts/mirror_maintenance/ toolchain (immutable, log-bearing per DAAF script versioning): mirror audit/build/validation (01-36), the endpoint ground-truth probe battery (37-44), parity-expansion harnesses (45-52), and the regeneration pipeline referenced by the explorer skill's maintenance-live-api.md - endpoint-file bridge (53), variable-dictionary generation (54), and the EDFacts race-code probes (55-56, currently blocked by an Urban-side outage). Appended execution logs are retained as the audit trail. Scripts reference session-workspace input paths; next-vintage runs create dated revisions per the documented regeneration pattern. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bg1F7KtnSSUdmVn57DrwGA
…metadata - Convert 112 prose em dashes in rendered copy and JS display strings to the house `--` convention; retain 52 "no data" placeholder glyphs and all comment/CSS em dashes (comment-aware rule-based sweep, node --check verified on all script blocks) - Metadata titles (title/og:title/twitter:title) and og:image:alt use a single hyphen per maintainer ruling 2026-08-10 (metadata-only exception to the body-copy `--` convention) - Add missing head fields per cross-site field-presence audit: theme-color, og:site_name, apple-touch-icon, og:image width/height/alt companions, twitter:image; flip twitter:card to summary_large_image (no static og:image — deploy script injects exactly one) - Replace inline SVG favicon with the shared site favicon-32 (absolute URL; bench page deploys under /bench/) - REPLACE-WITH-FINAL-URL.html deploy token verified byte-identical Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VNFFUrT3m2iNUn7ZZzQTAc
- Add openrouter registry entries for thinkingmachines/inkling and /inkling-small, both pinned to deepinfra/fp8 (shared-provider pair per maintainer decision; together endpoint degraded on inkling at fetch time). Endpoint facts verified online 2026-08-10; wire_ids OBSERVED (bare slugs) from the dc-01 probe archives; full purity pinning - Pricing billing-reconciled 2026-08-10: effective rates calibrated to obs/pred 1.00 on each model's billed token mix (heavy DeepInfra cache discounting); published list rates preserved in comments - Full 4-phase battery completed at 3 reps (306/306 runs, no stalls or timeouts); results under results/20260810_* - reconcile_openrouter_costs.py: default output JSON now dated at run time so reruns never overwrite a prior dated snapshot (the hardcoded 2026-06-11 default silently clobbered the June historical file) - README model-matrix counts 36/24; test_chatgpt_route key partition and count assertions updated (27/27 passing) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VNFFUrT3m2iNUn7ZZzQTAc
…collision (1.3.10 -> 1.3.18) Root cause: auto-start failed when a stale PID-derived FIFO (shim.stream.<pid>) collided with a reused supervisor PID. The repair replaces PID-derived transport with private mktemp-allocated stream workspaces and, across bounded review cycles, hardens the full lifecycle protocol: - kernel flock lifecycle/log locking with stable private lock dirs - queue-only signal handling retained through natural shell exit, with durable SHIM_RESTART_RESULT append as the irreversible commit boundary - typed fail-closed PID evidence via the new standard-library helper scripts/provider_shim/lifecycle_capability.py (no-follow reads, exact decimal grammar, role adjudication, start-time-bound TERM->KILL) - inode-aware workspace cleanup with quarantine-rename + re-stat + identity-bound unlink; documented as best-effort hardening against same-UID substitution (Linux has no inode-conditional unlink) - exact legacy four-field supervisor argv compatibility for live transitions; exact W1 capability token grammar - benchmark wire-ID contract narrowed to the observed exception; stale ModelConfig guidance corrected Tests: lifecycle Bats grown 52 -> 108; helper contract suite added (7 tests); provider suite 287/287; shim-route 36/36; benchmarks 182/182. Live daemon validated at 1.3.18 (strict ready, smoke-clean on all shim-relevant probes). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016Lyjxj4He482bpUfuKzba2
…erdict triage Voice-pass follow-up to the lifecycle repair (eb11074). The technical lifecycle-guarantees paragraph had accreted in the Option F setup flow across repair cycles; per the voice-pass standard (shim internals do not belong in setup steps), it now lives in the lane's optional "Under the hood" reference as "Lifecycle safety and its limits," with every load-bearing fact preserved verbatim, and the setup flow keeps one plain-language sentence in its place. Also adds SHIM_RESTART_RESULT triage guidance to the technical FAQ's shim troubleshooting entry: how to read the verdict line and its stage vocabulary (stop/launch/readiness), verified against the manager source. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016Lyjxj4He482bpUfuKzba2
…erage and display-retirement fixes (viewer v3.8.1) Pricing (2026-08-11 schedules, verified against live OpenAI/OpenRouter pages): - models.yaml: GPT-5.6 Terra/Luna api-equivalent blocks repriced to OpenAI's new schedule (Terra ~-20%, Luna 5x cut; Sol re-verified unchanged); OpenRouter lane entries updated to OpenRouter's listed rates (Terra cached_input inferred, marked provisional); dated provenance comments - cost_estimator.py: LUNA_*_CONTEXT_RATES constants + source/accessed stamps - README dual-ledger table and prose; install-guide slug table; rate-derived test expectations recomputed (test_chatgpt_route, test_subscription_accounting, test_artifacts_and_preflight) - Viewer battery costs re-price all historical runs at current list rates (existing build-time behavior); archived billing records untouched Changelog feature (viewer v3.8.0): - New curated, public, date-keyed benchmarks/CHANGELOG.md (16 entries, 2026-06-12 onward) parsed at build time (load_changelog, fail-soft) into __CHANGELOG_HTML__/__CHANGELOG_LATEST__ tokens; hero pill button (teal border, inline with the report date) opens a mobile-friendly modal (scrim/Escape/close dismiss, focus trap + return, scroll lock, full-screen sheet on narrow viewports); wrapped-bullet continuation-line parsing; README §8 format contract + regeneration workflow Cost coverage + display retirement (viewer v3.8.1): - reconcile_openrouter_costs.py: PERMASLUG_BASE_OVERRIDES for the dated-slug gotcha (deepseek-v4-flash-20260731 -> -0731); removed the stale static gemini-3.5-flash exclusion — both models restored to the battery-cost table via the new 2026-08-11 reconciliation snapshot - retired_display_exclusions (models.yaml + generator): undated DeepSeek V4 Flash fully removed from all report surfaces (168 runs; on-disk runs untouched); obsolete retired_model_pricing entry retired - Public prose no longer cites the OpenRouter Terra/Luna discount (report GPT-5.6 costs are OpenAI api-equivalent) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VfzTapWcwjWWwCk8j3rMMd
Drop the "Relative Duration" price-basis toggle from the Cost vs. Performance COST_FORMS registry and the "duration multiplier, precisely" methodology footnote (durationDisclosureHtml + call site) — both functionally deactivated. PRECOMPUTED.duration payload and the now-unreachable isDur branch remain (schema-additive, v3.7.0 precedent). Deliberately no public CHANGELOG.md entry (dead-UI removal, not audience-relevant); recorded in the generator's v3.8.1 docstring notes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VfzTapWcwjWWwCk8j3rMMd
…en 3.8, harden onboarding process Model onboarding (benchmarks/README.md § Adding a New Model, followed end-to-end): - config/models.yaml: add x-ai/grok-4.6:xai/zdr (wire_id OBSERVED, purity verified), deepseek/deepseek-v4-pro-0813 (bare slug; undated Pro kept active for side-by-side comparison), and qwen/qwen3.8-2.4t-a95b (day-zero endpoint not yet benchmark-viable: 3x instant 400 then 502 on full-payload traffic, $0 billed; temporarily display-quarantined via retired_display_exclusions to unbreak the range-quartile tier breaks). - Full 3-rep four-phase batteries for Grok 4.6 + DS Pro 0813; billing reconciled (billed/harness 1.06 / 1.04, list rates retained). New PERMASLUG_BASE_OVERRIDES entry: dated permaslug deepseek-v4-pro-20260813 otherwise misattributes to the still-active undated Pro. - tests/test_chatgpt_route.py: key partition + count assertions (39 total / 27 openrouter); fixed pre-existing count-comment drift (27/27 tests pass). Rep-count hygiene + editorial refresh (viewer v2, bundle daafbench_2026-08-12d): - Exactly-3-reps corpus contract verified: 33 models x 51 cases, 0 deviations (scratch/verify_rep_counts_2026-08-12.py, uses the generator's own loader). Onboarding probe sets quarantined (_quarantine_* rename + notes); DS 0813 dc-11 topped up at --timeout 1500. - viewer_template.html: hero + Key Takeaways re-ratified for the new frontier (five models, four creators; Grok 4.6 joins T1 prose); span contract now 27 with two-way span<->setter verification. - build_rerun_queue.py: dc-11/dc-12 are valid, retained cases (maintainer decision 2026-08-12, supersedes the 2026-07-18 audit exclusion); DEFAULT_EXCLUDED_CASES emptied. - README: new mandatory onboarding steps 7b (rep-count hygiene) and 8b (editorial review of narrative/takeaways/headlines). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EEPQtVXw9qcjVtVd374DzC
…n install.sh The DAAF.command fetch failure previously degraded to a quiet NOTE while the install reported success — the same fault aborts loudly on Windows (daaf.bat is load-bearing in install.ps1), so Mac gaps went invisible. The failure branch now emits a WARNING with the curl exit code, exact URL, and effective branch, plus remediation guidance (update_daaf.sh existence-heal; bash daaf.sh fallback), and removes any partial download so a broken file cannot shadow the later heal (Tier A only copies absent files). Still non-fatal; dry-run branch, Darwin gate, and success path unchanged. Diagnosed alongside: field report was installer bootstrap version skew (main install.sh executing with DAAF_BRANCH=daaf_dev content), resolved by branch-adjusted bootstrap; deferred follow-ups recorded in session notes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XbZcyZY2nw7mrPQg56wFMQ
…e pass Version sweep across the 9-file touchset (CITATION.cff first, README badge/citations, four agent_reference templates). New CHANGELOG entry: Double-Click to Launch, incl. two-updater-runs note and Windows AV known-issue note. Doc-currency fixes: two-run parenthetical precision, ~57 pinned Python packages, 12 R library skills. Voice pass over README, CONTRIBUTING, and user_reference 01/02/03/04/07 per five-reviewer scoping and three-angle review; all load-bearing facts preserved verbatim. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XbZcyZY2nw7mrPQg56wFMQ
Appended by run_all_smoke_tests.sh during v3.0.2 release validation (exit 0, igraph/tidygraph/ggraph all pass). Matches the committed-log convention of the other 18 smoke scripts. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XbZcyZY2nw7mrPQg56wFMQ
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Merge daaf_dev updates through v3.0.2!