AI_SPEND_AUDIT v2, can anything replace $200/mo of subscriptions?

Run 2026-08-28 · 15 candidates · 377 validated citations · workload observed from 30 days of local logs · every figure tagged observed or modelled · no external requests on this page.

1 · Verified sources (377 citations, re-fetched this run)

Every link below returned HTTP 200 during validation on 2026-08-28 and its page body still contained the cited evidence snippet (L1). Unverified sources are in the appendix.

2 · Verdict

Winner: OpenRouter, 60/100. It is the only candidate that pairs a frontier coding model (Opus 5; SWE-bench Verified 79.20 via an independent leaderboard entry) with every cost-control property this audit measured: prepaid credits, per-key spend limits, a real-time activity dashboard, pass-through per-token pricing plus a 5.5% credit fee, and zero recorded price or limit changes in the last 90 days. At the observed workload its modelled cost is $27,628/month, parity 0.0072, meaning $200 buys 0.7% of the volume consumed today. The parity hard-cap (≤60) encodes that: no scored option at $200/month reproduces the observed 33.4B-token month. The incumbent Claude Max 20x (56/100) is the only parity-1.0 row, it demonstrably served that volume for $200, at the cost of the lowest cost-control score in the top half (quotas not published in tokens, pool shared with chat, 14 throttle events in week 32). Runner-up: hybrid-router, 60/100 (LiteLLM proxy, 95/5 token split: DeepSeek V4 Flash default, Opus 5 fallback, hard monthly budget at the proxy). It is the better pick than OpenRouter when a routing policy is enforceable: the same money then covers parity 0.1013 (14× more) and every dollar is capped by LiteLLM max_budget before it is spent. The single risk most likely to change this answer: the Claude Max +50% weekly boost lapses on 2026-08-31. If delivered Max capacity drops in week 36, the parity-1.0 baseline measured here no longer exists, and the hybrid becomes the only path that keeps both capacity and control.

^† B3 tie-break: raw scores within 3 points of the top (61 / 60 / 60) are ordered by the Cost-control dimension, then Reliability. Kimi Vivace holds the highest raw score (61) but control 65 vs 100/100 for the two winners; its new signups have also been paused since 2026-07-20.

3 · What changed since last run

First run, no out/prev/ exists.

4 · Score breakdown (hover a cell for the arithmetic)

Hard caps fired:

• agg-openrouter: cap 60: parity 0.0072 < 1.0
• hybrid-router: cap 60: parity 0.1013 < 1.0

Capability uses independent benchmarks only (SWE-bench Verified, Terminal-Bench 2.1, Artificial Analysis Intelligence Index), normalised to best-in-set (83.8 TB / 63.05 AA / 79.2 SWE); fewer than 2 independent entries caps the dimension at 70. The S4 bake-off was skipped (no funded key), so no 50/50 blend was applied.

5 · Cost model, all cells modelled; workload mix observed

Observed 30-day mix (Claude Code logs, 2026-07-29 → 2026-08-28): input-miss 250M, cache-write 935M, cache-read 32.1B, output 122M tokens; cache-hit rate 96.4%. Formula: cost = in_miss·p_in + cache_read·p_hit + cache_write·(p_cw or p_in) + out·p_out, × peak adjustment × (1+fee) + fixed. DeepSeek peak (Mon–Fri 01–04 & 06–10 UTC = 09–12 & 14–18 Kuala Lumpur) carries 2× prices; 23.6% of observed tokens fall inside it. The off-peak-shift column halves that fraction (batch work moved off-peak).

Sub break-even = token volume at which the same models on PAYG API cost $200/month; the sub is cheaper above it. Observed volume is 33,398M tokens, 131× Claude Max's break-even of 255.1M.

6 · Observed usage & throttle timeline (tool: claude_code; per ISO week, total billed tokens)

Throttle events observed in logs: 14 (all in week 2026-W32). Codex CLI and Kimi app stores held no token logs on this machine (Codex sqlite threads table: 0 rows; ~/.kimi: config only), their usage is not in these bars; Kimi K3 tokens routed through Claude Code (1,567M input) are.

7 · Candidate cards

8 · Migration plan (winner: OpenRouter, with the hybrid as its enforcement layer)

1. Day 1, spend cap first. Create an OpenRouter account; buy $50 of credits, no auto-top-up. Create one provisioning key per tool with a per-key credit limit: claude_code=$20, opencode=$15, experiments=$15. Total exposure: $50, hard.
2. Stand up LiteLLM (MIT, self-hosted) as the router: default route deepseek/deepseek-v4-flash (off-peak aware), fallback anthropic/claude-opus-5 via OpenRouter; set proxy max_budget=200, budget_duration=30d.
3. Point tools at the proxy: ANTHROPIC_BASE_URL=http://localhost:4000 for Claude Code (LiteLLM /v1/messages passthrough); OpenAI-compatible base URL for opencode.
4. Days 1–7, parallel run. Keep Claude Max 20x active. Route 3 real projects through the proxy. Compare per task: pass rate, wall time, and router-metered USD vs the sub's flat $200. The S4 fixture in out/S4/fixture/ (hash 2d78…8cb0) is the controlled comparison set: 3 tasks, 2 attempts each.
5. Decision gate, day 7: if proxy pass-rate ≥ sub pass-rate on the fixture AND projected 30-day proxy spend ≤ $200 for the workload actually routed, downgrade one sub tier ($200→$100), never all at once. Observed data says full replacement will NOT fit $200 (parity 0.0072–0.10); the realistic outcome is sub + capped overflow lane.
6. Watch 2026-08-31 (Max boost lapse) and week-36 throttle counts in /usage; if delivered capacity drops, shift the routing split rather than buying a second sub.
7. Rollback: unset ANTHROPIC_BASE_URL (tools revert to the sub), leave remaining OpenRouter credits parked, they do not expire on a schedule the audit found, and the per-key caps keep them inert.

9 · Assumptions & gaps (how missing data moved scores)

• S4 bake-off skipped for every candidate, DeepSeek key present but balance −$0.04 (is_available:false); Cloudflare token lacks Workers AI scope; no other keys in env. Effect: capability came from public benchmarks alone; candidates with <2 independent entries (api-deepseek, api-moonshot, sub-kimi-top) were capped at 70 and may be understated.
• Codex CLI and Kimi app usage unobserved (no local token logs). Observed workload = Claude Code only; if the other tools carry similar volume, every parity figure here is an overestimate of what $200 buys.
• Sub quotas in messages/credits, not tokens (Max, Codex Pro, Kimi), parity for Codex Pro and Kimi is unquantifiable (marked n/a, escaping the <1.0 cap); their cost-control scores already carry the −10 "not in tokens" deduction, and Max's is capped at 40.
• Modelled-profile cache-write assumption: all cache-miss input billed at the cache-write rate where a provider prices writes (worst case for agentic loops). Observed rows use the real 4-way split.
• Self-host throughput proxied from Artificial Analysis median API-host speed (69.2 tok/s GLM-5.2) + assumed 5,000 tok/s prefill and 1.2× warm-idle; no published self-host throughput for these models passed validation. Self-host costs are order-of-magnitude, tagged modelled.
• Hybrid splits are policy assumptions (80/20, 90/10, 95/5 by token volume), not measurements.
• E4 disagreements recorded: Claude Max context (200k table vs 1M "newest models" page); GLM-5.2 params 753B (HF) vs ~743B (vLLM recipe); OpenRouter Kimi-K3/DeepSeek page-vs-API price rows. Each kept both citations, confidence LOW.
• 147 null fields across artifacts, each with a reason recorded in its candidate JSON; the notable ones are listed per candidate below.

• agg-openrouter, confidence LOW, 10 nulls
  full null list

  • benchmarks (current-gen) (swebench.com Verified leaderboard (rendered) lists GLM 5/Kimi K2.5/Dee)
  • benchmarks[].evidence_snippet (Leaderboard table cells are tag-separated; only the bare score string )
  • changes_last_90d (OpenRouter API changelog is OpenAPI-generated (docs/changelog.md); no )
  • fixed_fees_usd_month (No monthly subscription; OpenRouter monetizes via 5.5% ($0.80 min) car)
  • integrations.kimi_code (No OpenRouter doc for Moonshot's kimi_code CLI in docs/llms.txt index )
  • limits.rpm (paid models) (Docs publish rate limits only for :free model variants (20 RPM; 50 RPD)
  • limits.shared_pool_with_chat (No published statement fetched on whether openrouter.ai chatroom usage)
  • limits.tpm/concurrency/session_window/weekly_cap (Not published on openrouter.ai/docs/api-reference/limits (fetched toda)
  • models[].open_weights (License/HF repos for GLM-5.3, Kimi K3, DeepSeek V4 Pro not fetched tod)
  • models[].price (kimi-k3, deepseek-v4-pro) (E4: model page and /api/v1/models disagree (kimi-k3 $2.55/$12.75 page )
• hybrid-router, confidence LOW, 9 nulls
  full null list

  • benchmarks (no benchmark applies to a routing proxy; capability comes from routed )
  • changes_last_90d (OSS software fee $0 unchanged; no price/limit change events apply to a)
  • fixed_fees_usd_month(infra) ($0 is the software license fee (MIT 'free of charge'); self-hosting co)
  • limits.rpm/tpm/concurrency (operator-set per key via tpm_limit/rpm_limit/max_parallel_requests in )
  • limits.session_window/weekly_cap (not a LiteLLM concept; budget_duration (e.g. 30d) is the reset window,)
  • models[].context_tokens/max_output_tokens (modelled in S3 from other candidates' artifacts)
  • models[].price_usd_per_1M.* (modelled in S3 from other candidates' artifacts (brief instruction); r)
  • reliability.status_page (self-hosted LiteLLM: the status page is your own deployment; OpenRoute)
  • throttling_incidents_180d (self-hosted: no vendor throttling; no official OpenRouter throttling s)
• sub-kimi-top, confidence LOW, 8 nulls
  full null list

  • integrations.codex_cli (no official membership doc for Codex CLI; membership docs list Kimi Co)
  • limits.rpm/tpm (membership limits published as requests per 5-hour window range and cr)
  • models.kimi-for-coding.context_tokens (context window for kimi-for-coding alias not published on membership p)
  • models.kimi-for-coding.price_usd_per_1M (subscription benefit; no per-token USD price published for membership )
  • per-tier 5-hour request quotas (only the range 300-1,200 requests/5-hour published; per-tier (Vivace) )
  • primary_urls.changelog (no public changelog page found for kimi.com membership)
  • reliability.data_retention_url (kimi.com consumer privacy policy not fetched within budget; platform p)
  • sub_boost_lapse_dates (no boost/lapse promotions found on fetched pages)
• sub-codex-pro20, confidence LOW, 8 nulls
  full null list

  • credit_price_usd (No fixed USD-per-credit price published on any fetched vendor page tod)
  • limits.rpm/tpm/concurrency (Not applicable/published for ChatGPT-plan Codex; quotas published as p)
  • limits.weekly_cap (Vendor states 'Additional weekly limits may apply' with no published n)
  • models[].price_usd_per_1M (Subscription overflow bills in credits per 1M tokens (rate card cited))
  • models[gpt-5.3-codex-spark].context_tokens (Changelog states text-only 128k context at Feb 2026 launch; current va)
  • models[gpt-5.6-luna].context_tokens (Model page not fetched within budget.)
  • primary_urls.data_retention (Consumer-plan data retention page not fetched within budget.)
  • sub_boost_lapse_dates.pro100_10x_codex_launch_promo (Secondary sources report a 10x-of-Plus Codex promo on the $100 Pro tie)
• sub-claude-max20, confidence LOW, 9 nulls
  full null list

  • benchmarks (SWE-bench Verified for plan models) (swebench.com (fetched 2026-08-28) lists no Opus 5 / Sonnet 5 / Fable 5)
  • integrations.codex_cli / integrations.kimi_code (no official or community doc for using a Claude Max subscription in th)
  • integrations.opencode.official=false (opencode docs (fetched 2026-08-28): 'plugins that allow you to use you)
  • limits.rpm / limits.tpm (quotas not published in tokens or requests; vendor states 'no fixed me)
  • limits.weekly_cap (quantity) (weekly limit exists but no token/message quantity is published on vend)
  • models[].context_tokens (two primary sources disagree: claude.com/pricing comparison table says)
  • models[].max_output_tokens ('higher output limits' stated for Max but no number published on fetch)
  • models[].price_usd_per_1M.* (subscription plan; Anthropic publishes no per-token price for included)
  • open_weights (proprietary hosted models; not applicable)
• api-openai, confidence LOW, 7 nulls
  full null list

  • benchmarks SWE-bench Verified for gpt-5.3-codex / gpt-5.6 (swebench.com leaderboard (fetched today) has no submission for these m)
  • integrations.claude_code/kimi_code (No official documented OpenAI-API path found in fetched docs; communit)
  • limits.concurrency (OpenAI publishes no concurrency limit; documented limit dimensions are)
  • limits.usage_dashboard_realtime (Set false as unproven: no fetched primary page states dashboard latenc)
  • models[gpt-5.3-codex].price_usd_per_1M.cache_write (No cache-write price listed for gpt-5.3-codex on pricing page (dimensi)
  • models[gpt-5.6-luna].context_tokens/max_output_tokens (Luna model page not fetched within budget; Artificial Analysis (indepe)
  • throttling_incidents_180d (Status page reports 'fully operational', APIs 99.94% uptime May-Aug 20)
• self-rented-gpu, confidence LOW, 16 nulls
  full null list

  • benchmarks: GLM-5.2/Kimi-K3/Kimi-K2.7-Code on swebench.com (absent from swebench.com embedded leaderboard data fetched 2026-08-28)
  • benchmarks: Kimi models on tbench.ai 2.1 (no Kimi entries in tbench.ai terminal-bench 2.1 page source fetched 20)
  • changes_last_90d (GPU rental price-change history not published on fetched pricing pages)
  • gpu_rentals.lambda.H200 (H200 per-GPU-hr not listed in tables on lambda.ai/service/gpu-cloud fe)
  • gpu_rentals.vast median (30-day median chart is canvas-rendered; only the page's current-offer )
  • limits.rpm/tpm/session_window/weekly_cap (not applicable: operator-controlled self-hosted endpoint)
  • limits.usage_dashboard_realtime (provider console usage views not fetched)
  • models[*].max_output_tokens (not stated as a separate limit on model cards fetched)
  • models[*].price_usd_per_1M (self-host: no per-token vendor price; cost basis is GPU rental per hou)
  • models[GLM-5.2].params_total_b (E4 disagreement: HF safetensors counter says 753B params; vLLM recipe )
  • models[Kimi-K2.7-Code].published_throughput (no tokens/s figure in vLLM recipe or HF model card fetched 2026-08-28)
  • models[Kimi-K3].published_throughput (no tokens/s figure in vLLM recipe or HF model card fetched 2026-08-28)
  • models[Qwen3-Coder-30B].min_serving_config.gpus (HF card gives vllm serve command but names no GPU model/count)
  • models[Qwen3-Coder-30B].published_throughput (no tokens/s figure in HF model card fetched 2026-08-28; no recipes.vll)
  • reliability.status_page (provider status pages not fetched within budget)
  • throttling_incidents_180d (no official incidents found on fetched pages; dedicated incident searc)
• agg-coding-plans, confidence LOW, 16 nulls
  full null list

  • atlas_data_retention (Atlas docs index lists Privacy Policy / Data De... pages but they were)
  • atlas_status_page (status.atlascloud.ai returns a JS-only shell with no status content wh)
  • benchmarks[].evidence_snippet (swebench.com leaderboard cells are tag-separated; only bare score is v)
  • changes_last_90d (No changelog located for either vendor's coding plan; unverified, not )
  • fixed_fees_usd_month (Multiple plans, one field: Atlas $10/$20/$50/$100/$200/$500 per month;)
  • integrations base_url/env (Atlas quick-start names first-class Claude Code/Codex/OpenCode support)
  • limits.published_in_tokens_or_requests (TRUE for Novita (quota in tokens + requests/min). Atlas quota is in PO)
  • limits.rpm (Value 45 = Novita Lite tier only (Pro 150, Max 450, cited in plans[]).)
  • limits.spend_cap_settable (No user-settable cap found; Atlas plans hard-pause on exhaustion (cite)
  • limits.usage_dashboard_realtime (Novita page shows 'Unified Console Centralized billing, quota, and API)
  • limits.weekly_cap (Value 16500000 = Atlas Starter 16.5M POINTS/week (not tokens); each ti)
  • models[].max_output_tokens (Not published on the fetched Atlas/Novita pricing pages.)
  • models[].open_weights (HF repos/licenses not fetched today (fetch budget).)
  • novita_quota_reset_period (Tier cards show token totals with monthly billing; explicit reset word)
  • seed_claim_45pct (VERIFIED for Atlas: multiplier table lists '45%' savings vs official p)
  • throttling_incidents_180d (status.novita.ai shows 'All services are online'; incident history not)
• sub-zai-coding, confidence LOW, 7 nulls
  full null list

  • first-month discount (no first-month/first-order discount shown on z.ai/subscribe render 202)
  • fixed_fees_usd_month (single field cannot express 3 tiers x 3 billing cycles; Lite monthly l)
  • limits.concurrency (vendor publishes no numbers: 'dynamically adjusts ... Max > Pro > Lite)
  • limits.usage_dashboard_realtime (Charge Type page documented but refresh latency unverifiable without a)
  • models[].price_usd_per_1M (subscription bills in credits, not USD per token; multipliers recorded)
  • models[glm-5.3].open_weights (GLM-5.3 weights not published; see api-zai.json)
  • throttling_incidents_180d (no official incident feed (no status page found); systematic user-repo)
• api-anthropic, confidence LOW, 8 nulls
  full null list

  • benchmarks (SWE-bench Verified, Opus 5) (swebench.com fetched 2026-08-28 has no Claude Opus 5 / Sonnet 5 / Fabl)
  • benchmarks (Terminal-Bench, Opus 5) (tbench.ai Terminal-Bench 2.1 leaderboard fetched 2026-08-28 lists Fabl)
  • data_retention detail (API inputs/outputs auto-deleted within 30 days of receipt or generatio)
  • integrations.codex_cli / integrations.kimi_code (no official Anthropic-provider doc for these CLIs found in fetched pag)
  • limits.concurrency (Anthropic publishes RPM/ITPM/OTPM per tier, no concurrency limit on fe)
  • limits.session_window (not applicable to PAYG API; limits are per-minute token buckets plus m)
  • models[].price_usd_per_1M.peak_multiplier (no peak-window pricing published; optional fast mode (Opus 5 $10/$50, )
  • open_weights (proprietary hosted models; not applicable)
• api-moonshot, confidence LOW, 8 nulls
  full null list

  • benchmarks.swe_bench_verified/tbench.ai (tbench.ai leaderboard rendered empty (client-side data); swebench.com )
  • limits.session_window/weekly_cap (not applicable: API limits are RPM/TPM/TPD by recharge tier; TPD Unlim)
  • limits.spend_cap_settable (prepaid recharge model; docs document spending alerts, not settable ha)
  • models.*.cache_write (context caching is automatic; no cache-write fee documented on caching)
  • models.*.max_output_tokens (per-model max output tokens not published on pricing or model paramete)
  • models.*.peak_multiplier (no peak/off-peak pricing documented on international pricing pages)
  • models.kimi-k2.6/kimi-k2.5.open_weights.license/params (model cards not fetched within 40-fetch budget; repos confirmed on HF )
  • models.kimi-k2.7-code.min_serving_config.gpus (HF card names engines (vLLM/SGLang/KTransformers) but no GPU count; de)
• api-deepseek, confidence LOW, 9 nulls
  full null list

  • benchmarks SWE-bench Verified / Terminal-Bench official leaderboards for V4 (no DeepSeek V4 entry on swebench.com (newest: DeepSeek V3.2 high 70.00)
  • limits.rpm/tpm (DeepSeek publishes no RPM/TPM limits; the documented scheme is per-acc)
  • limits.shared_pool_with_chat (no public statement; status page lists API service and chat service as)
  • limits.spend_cap_settable (platform.deepseek.com console is login-gated; no public doc on spend c)
  • limits.usage_dashboard_realtime (usage page is behind login at platform.deepseek.com; realtime granular)
  • models[].open_weights.min_serving_config.gpus (HF cards give vLLM/SGLang serve commands but no GPU count/type)
  • models[].price_usd_per_1M.cache_write (pricing page lists no cache-write or cache-storage fee; only cache-hit)
  • models[deepseek-v4-flash].open_weights.params_active_b (not stated on HF model card text; safetensors gives total only; AA pag)
  • throttling_incidents_180d (status page shows no incidents flagged (99.92%/99.81% uptime May-Aug 2)
• cf-workers-ai, confidence LOW, 11 nulls
  full null list

  • benchmarks.swe_bench_verified (No GLM-5.x entry extracted from swebench.com within budget; AA Intelli)
  • confidence (LOW solely because tpm/concurrency/session_window/weekly_cap are unpub)
  • limits.spend_cap_settable (Negative claim: docs describe threshold usage notifications only; no p)
  • limits.tpm/concurrency/session_window/weekly_cap (Cloudflare publishes per-minute request limits and a daily neuron allo)
  • limits.usage_dashboard_realtime (Dashboard monitoring documented on the pricing page; refresh latency/r)
  • models[*].max_output_tokens (Not listed on Cloudflare pricing or model pages fetched today)
  • models[*].price_usd_per_1M.cache_write (Cloudflare lists cached-input-read price only; no cache-write price pu)
  • models[@cf/zai-org/glm-5.2].open_weights (No HF/license link on the Cloudflare glm-5.2 model page; HF not fetche)
  • reliability.alt_hosts (Alternative hosts for the same models not fetched within budget)
  • seed.context_served_262k_not_1M (Seed claim matches glm-5.2 (Cloudflare serves 262,144; AA lists 1M nat)
  • throttling_incidents_180d (cloudflarestatus.com history.rss covers only ~2 weeks; no Workers AI i)
• api-zai, confidence LOW, 10 nulls
  full null list

  • benchmarks for glm-5.3 on swebench.com/tbench.ai (no GLM-5.3 entry on either official leaderboard as of 2026-08-28; newe)
  • data_retention duration (privacy policy: retained 'as long as you have an account'; processed i)
  • limits.rpm/tpm/concurrency (published only in login-gated console (z.ai/manage-apikey/rate-limits))
  • limits.shared_pool_with_chat (no public statement found on pooling between chat.z.ai and API)
  • limits.spend_cap_settable (console login-gated; no public doc on spend caps; prepaid recharge mod)
  • models[].price_usd_per_1M.cache_write (pricing page bills 'Cached Input Storage' currently 'Limited-time Free)
  • models[glm-5.3-flash].open_weights.min_serving_config.gpus (HF card lists frameworks (SGLang/vLLM/KTransformers) but no GPU count)
  • models[glm-5.3].open_weights (GLM-5.3 weights not on HF zai-org (org page fetched 2026-08-28); AA la)
  • reliability.status_page (status.z.ai does not resolve (NXDOMAIN on 2026-08-28); no status link )
  • throttling_incidents_180d (no official incident feed found (no status page); systematic user-repo)
• self-local, confidence LOW, 11 nulls
  full null list

  • benchmarks: GLM-4.5-Air on swebench.com/tbench.ai (no GLM-4.5-Air entries in swebench.com embedded data or tbench.ai 2.1 )
  • changes_last_90d (no price/limit change history applicable; Apple price history not publ)
  • fixed_fees_usd_month (one-time hardware purchase; no recurring vendor fee. Electricity not p)
  • hardware.availability (Apple page lists M5 Mac Studio as pre-order, 'Available starting 9.22')
  • limits.rpm/tpm/session_window/weekly_cap (not applicable: local operator-controlled server)
  • limits.usage_dashboard_realtime (not applicable: no vendor dashboard for local inference)
  • models[*].max_output_tokens (not stated as a separate limit on model cards fetched)
  • models[*].min_serving_config.gpus (no primary doc fetched names a specific GPU/Mac model; only quantized )
  • models[*].price_usd_per_1M (self-host on owned hardware: no per-token price; cost is hardware cape)
  • models[GLM-4.5-Air].context_tokens (context length figure not captured from fetched card text; not re-fetc)
  • throttling_incidents_180d (not applicable: local hardware)

10 · Appendix — unverified / blocked sources (14)

USER_LINK resolved via mirror (x.com blocks direct fetch; retrieved through Twitter's syndication CDN, mirror:true): @zebassembly, 2026-08-27 — “i was so excited that local models like qwen 3.8 27B exist, and then I did the math on how much it costs on my power bill.... looks like i'll be using glm 5.3 flash on Cloudflare instead” Author carries a Cloudflare business label on the platform; the audit scored Cloudflare Workers AI on its own artifacts (B1): final 39, parity 0.1658, rpm 20.

Run fb2d5acf-a818-4a52-9c17-b4da13b7f15b · inputs_hash 6ac7f23a16c9727d… · S0–S7 per AI_SPEND_AUDIT v2 · every figure tagged observed/modelled · generated 2026-08-28.