The same budget now does 8.4 times the work.
Think of a car that suddenly goes many times further on the same tank. Same tank, same road, same driver, and the same tests checking the work.
Cost per worker job, 25 matched jobs, API-equivalent pricing. Updated 2026-10-08.
Every figure on this page comes from a dated record of our own machine and our own AI workers. Each one is labelled MEASURED or MODELLED, and each one says how it was taken. Figures still being taken show as measuring, and are filled in when the record exists.
21 figures have both a before and an after, 27 have at least one side filed, 32 are tracked. Round 1.
Before and after, every vector
Machine
| Figure | Before | After | Change | Label |
|---|---|---|---|---|
| Memory in use on the build machine | 28 GB | measuring | measuring | MEASURED |
| CPU busy while 'doing nothing' | 84% | measuring | measuring | MEASURED |
| Windows VM memory (vmmem) after a cache trim | 13 GB | 7.4 GB | -43% 1.8x | MEASURED |
| Load the worker fleet was allowed to put on 8 cores A ceiling, not an observed load. Observed load is a round-2 figure. | 24 | 6 | -75% 4.0x | MEASURED |
| Observed load average (24 hour mean) | measuring | measuring | measuring | MEASURED |
| Memory per headless worker spent on tool servers | 550 MB | 0 MB | -100% | MEASURED |
Cron
| Figure | Before | After | Change | Label |
|---|---|---|---|---|
| Cron work per 24 hour day Job-hours can exceed 24 because jobs overlap. | 43.3 job-hours | measuring | measuring | MEASURED |
| Failure rate of the 5-minute state snapshot job | 89.2% | measuring | measuring | MEASURED |
| 5-minute snapshot, time per run | 224 s | 3.1 s | -99% 72x | MEASURED |
| Cloud mirror, time per run Before is the midpoint of a 30 to 45 minute range. | 2,250 s | 14 s | -99% 161x | MEASURED |
| Progress ping, time per run | 222 s | 0.07 s | -100% 3171x | MEASURED |
| Repository bundle step, time per run | 51 s | 3.6 s | -93% 14x | MEASURED |
| Most jobs firing in one minute | 22 | 5 | -77% 4.4x | MEASURED |
Backups
| Figure | Before | After | Change | Label |
|---|---|---|---|---|
| Backup success rate | measuring | measuring | measuring | MEASURED |
| Files the cloud sync client tracks After is approximate. | 235,274 | 142,700 | -39% 1.6x | MEASURED |
Fleet
| Figure | Before | After | Change | Label |
|---|---|---|---|---|
| Fleet throughput (jobs completed per day) | measuring | measuring | measuring | MEASURED |
| Jobs per hour | measuring | measuring | measuring | MEASURED |
AI cost
| Figure | Before | After | Change | Label |
|---|---|---|---|---|
| Tokens a worker reads before its first action | 215,000 tokens | 29,000 tokens | -87% 7.4x | MEASURED |
| Cost per worker job (API-equivalent) | US$0.973 | US$0.116 | -88% 8.4x | MEASURED |
| Daily spend, modelled ceiling | US$467 per day | US$130 per day | -72% 3.6x | MODELLED |
| Interactive session start | 118,369 tokens | 78,105 tokens | -34% 1.5x | MEASURED |
| Session start inside the main codebase | 139,579 tokens | 101,493 tokens | -27% | MEASURED |
AI quality
| Figure | Before | After | Change | Label |
|---|---|---|---|---|
| Recall test, 50 questions | 50/50 | 48/50 | measuring | MEASURED |
| Held-out recall test, 25 questions | 25/25 | 24/25 | measuring | MEASURED |
| Fully good results on 25 matched jobs | 23/25 | 25/25 | measuring | MEASURED |
| Kernel compression gate | bar: 48/50 | passed, 48/50 | measuring | MEASURED |
Build
| Figure | Before | After | Change | Label |
|---|---|---|---|---|
| Main codebase test suite run time | 975 s | 579 s | -41% 1.7x | MEASURED |
| Kernel gate run time | measuring | measuring | measuring | MEASURED |
| Editor hook time per edit | 1.7 s | 0.1 s | -94% 17x | MEASURED |
| Reply hook time per reply | measuring | measuring | saves 3.5 s | MEASURED |
| Kernel repository size | 235 MB | 35 MB | -85% 6.7x | MEASURED |
| Disk freed from stale worktrees | measuring | measuring | saves 2.94 GB | MEASURED |
How each figure was taken
MEASURED means read from a dated record or a timed run. MODELLED means real measured quantities repriced under a stated scenario, which is a what-if and never a result. Where a before or after is not in yet, the table says measuring.
How: Windows memory reading while the fleet ran idle-looking; 28 of 31 GB in use before the work. Round 2 re-reads the same counter.
Source: PC optimisation report, 2026-10-08
How: Windows CPU reading before the work. Round 2 re-reads it with the fleet running under its new ceilings.
Source: PC optimisation report, 2026-10-08
How: Measured around one cache-trim run: 13.0 GB before, 7.4 GB 20 seconds later. A scheduled trim now runs every 10 minutes.
Source: PC optimisation report, 2026-10-08
How: Configured ceiling, read from the fleet governor settings: allowed load 24 before, 6 after. Workers dropped from 12 to 3 per account.
Source: PC optimisation report, 2026-10-08
How: Round 2: mean load average over 24 hours before and after, from the machine's own counters.
Source: Round 2 (measuring)
How: Each worker started about 7 tool servers (about 550 MB); a strict config now starts none.
Source: PC optimisation report, 2026-10-08
How: Summed run time of every scheduled job over 3 days of run records, per day. Round 2 repeats the sum on the same record.
Source: PC optimisation report, 2026-10-08
How: 256 of 287 runs failed (89.2%), so state was not reaching the remote copy. The fix is in; the after rate needs a fresh run record.
Source: PC optimisation report, 2026-10-08
How: Run record average before, a timed run after. The secret check now scans only files that could be committed.
Source: PC optimisation report, 2026-10-08
How: Before: 30 to 45 minutes a run (midpoint shown, 37.5 minutes). After: 14 seconds, incremental with a daily full pass.
Source: PC optimisation report, 2026-10-08
How: Ping re-ran a test suite that always timed out. The result is now cached until the code changes.
Source: PC optimisation report, 2026-10-08
How: One bundle per repository keyed on content, instead of re-bundling about 12 worktrees.
Source: PC optimisation report, 2026-10-08
How: Counted from the schedule. A lint check now fails any minute with more than 5 jobs.
Source: PC optimisation report, 2026-10-08
How: Round 2: successful runs over total runs for the backup jobs, before and after, from run records.
Source: Round 2 (measuring)
How: File count before; about 142,700 after a duplicate rolling copy (92,538 files) was excluded and moved out.
Source: PC optimisation report, 2026-10-08
How: Round 2: completed worker jobs per day from the fleet ledger, before and after the ceilings.
Source: Round 2 (measuring)
How: Round 2: completed jobs divided by active hours, from the fleet ledger.
Source: Round 2 (measuring)
How: Counted at session start across 198 worker sessions (before, about 215k, because the rule book loaded twice) against sessions on the compact core (about 29k).
Source: Sigil ceiling report, 2026-10-08
How: 25 matched jobs: same task text as earlier runs on the full rule book, re-run on the compact core. Cost is API-equivalent token pricing, not a bill.
Source: Kernel compression gate record, 2026-10-08; Sigil ceiling report, 2026-10-08
How: The same measured token counts repriced under five layers of change. The last layer needs a decision not yet taken. A scenario, not a forecast.
Source: Sigil ceiling report, 2026-10-08
How: 50 questions about the rule book, read-only tools, judged. Before is the full rule book, after is the compact core. Bar was 48 or better.
Source: Kernel compression gate record, 2026-10-08
How: Questions written by an independent agent that never saw the compact core.
Source: Kernel compression gate record, 2026-10-08
How: Same 25 tasks. Two full-rule-book results claimed clean lint on files that fail lint. The compact core had none.
Source: Kernel compression gate record, 2026-10-08
How: The compact core is 17,143 bytes. It must clear the bar before it is allowed to load.
Source: Kernel compression gate record, 2026-10-08
How: Tokens before the first reply with the old and new kernel layout.
Source: Kernel compression gate record, 2026-10-08
How: Tokens before the first reply after 18 reference files moved to on-demand loading. The full test suite stayed green.
Source: Sigil ceiling report, 2026-10-08
How: Wall-clock seconds for the full suite, before and after.
Source: Round-2 brief, 2026-10-08 (raw log attaching)
How: Round 2: wall-clock seconds for the kernel verify gate, before and after.
Source: Round 2 (measuring)
How: Seconds of hook time per file edit, before and after.
Source: Round-2 brief, 2026-10-08 (raw log attaching)
How: Seconds removed from every reply. Saved time is the figure supplied; before and after are filed with the log.
Source: Round-2 brief, 2026-10-08 (raw log attaching)
How: Size of the repository directory before and after history and run-state cleanup.
Source: Round-2 brief, 2026-10-08 (raw log attaching)
How: Gigabytes freed by removing finished worktrees.
Source: Round-2 brief, 2026-10-08 (raw log attaching)
Timeline
Hardening
Speed means nothing if it breaks things. These are the counts of the checks that stand behind the figures above.