Nothing in the workspace compared any number to 100,000 events/s while
the evidence file reported "AM-6 | met, 16.5x". Now a test does — a test,
not a bench, because Criterion reports throughput and asserts nothing,
which is why this row measured nothing for six passes.
Measured on bnt-lap001: 341,280 ev/s in debug (3.4x the target), ~2.4-3.1M
in release. The spec target holds even in an unoptimized build, so the
gate needs no cfg split and runs in the ordinary `make test`.
The trap this task named — loosening a flaky timing assertion until it
never fires — is avoided by construction. The threshold is the spec value,
untouched; the constant says lowering it requires an ADR; and the failure
message repeats that, states measured headroom, and names reference
figures, so an agent hitting a red AM-6 is told not to tune it in the
place they are actually reading. Robustness comes from best-of-N, not from
a lower bar: a throughput floor asks whether the machine is capable, so
transient load should not fail the build.
Two positive controls in the test: a run that applied fewer than 50,000
events, or measured zero elapsed time, fails rather than scoring as
infinite throughput.
Verified by a PROPERTY mutation — 4,000 black_box iterations injected into
GroundState::fold, the hot path — not a threshold tweak, which would only
prove the comparison runs.
And the FA class found last pass is now gated. mutation-check rows gained
an `expect` field: the mutant's output must contain the row's stated
failure string or the verdict is WRONG-REASON, not red. Without it a
mutation that merely failed to compile would credit its row with an
assertion it does not have. Verified by pointing expect at a string the
verifier never prints and watching the verdict flip. This is remedy (2)
from the CB-WP-0005 retrospective, built a task earlier than planned
because the class it guards is the newest and most dangerous.
M-D1-MUT: 4 -> 5 of 14.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
For each acceptance row in GameKernel §5, invert the property and require
the verifying command to go red. adapted:mutation-testing, with the
denominator changed from source lines to acceptance rows.
M-D1-MUT: 4/14 rows enforced
PARTIAL 2 (AM-7, AM-8 — some clauses live, some inert)
unmutatable 8 (no property to invert, reason stated per row)
SURVIVED 0
Two corrections to our own numbers. There are 14 rows, not the twelve
ADR-0005 and CB-WP-0005 both asserted — AM-4 splits into a/b/c. And the
prediction of 9-of-12 (75%) becomes >=10 of 14; measured 4 (29%), badly
unmet. No target moved in this commit.
The second correction matters more. My first run reported two SURVIVED
rows and both were my own no-op mutations: `pub struct NullRng;` ->
`pub struct NullRng {}` is semantically identical, and renaming
max_age_days does nothing because CA-17 reads it with a default of 90.
Both would have been published as "this row asserts nothing" — a false
accusation against code that is fine. Replaced with real inversions (a
per-construction counter in the ChaCha seed; reverting AC-9's output
resolution to the first-wins bug it was fixed for), after which both go
red. T08 asks whether writing a weak mutation is the new grep. It is,
demonstrably, on the first attempt.
The finding is larger than the workplan assumed. 8 of 14 rows are
unmutatable — AM-2, AM-3, AM-4c, AM-5, AM-6, AM-9, AM-10, AM-11 have no
instrument at all. AM-6 is the sharpest: nothing in the workspace
compares any number to 100,000 events/s, the headline throughput claim.
The problem is not three unimplemented rules, it is that more than half
the acceptance table has nothing behind it.
Harness controls: a stale find-string reports HARNESS-BROKEN rather than
scoring the baseline as the mutant; a red baseline reports inconclusive
rather than red; the tree is restored in a finally and the restoration is
verified. Not in `make all` — it rebuilds the workspace once per row.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
AGGREGATE becomes a list of source roots and rule patterns become
per-spec, so the link runs over every numbered spec x every crate rather
than GroundRules.md x games/ground/src/lib.rs.
The prediction held on the first run:
AM-1b kernel spec->code link: 15/18 (83%) across 10 source files
unlinked: K10 K14 K18
Kernel rules are link-only by design, and the output says so: they are
kernel invariants with no aggregate, setup preset or command vocabulary,
so scenarios/kernel/*.yaml with covers: [K11] would be a tag in a
directory the runner cannot dispatch. Claiming scenario coverage for them
is the inflation this gate exists to prevent.
Per ADR-0005 §5 the kernel arm reports without feeding the exit code
until 2026-08-31, then binds — the date in the tool, not in prose, with
days remaining printed every run, because open-ended "gate it later" is
how AM-4's targets went unratified for four workplans. The self-test
asserts the gate returns 0 before that date and 2 after.
The zero-rules positive control is replicated on the new denominator: a
kernel regex that stops matching aborts rather than printing 0/0 as
though it were 100%.
The self-test passed while the tool was completely broken. A print(
inside say() became say(), so every real `make coverage` died with
RecursionError while --self-test reported all-ok — it only ever called
kernel_arm(quiet=True) and never executed the reporting path. The control
named the behaviour and did not assert it, which is precisely what this
workplan is about. Fixed by exercising the loud path and asserting it
prints, then verified by re-breaking say() and confirming both new checks
go red. Seventh instance of the harness-does-nothing shape, in the tool
written to find that shape.
Also caught by its own gate: a self-test label that printed "0 K-ids"
beside a passing ">5" assertion, because the detail string rebuilt the
pattern with different escaping. A label that contradicts its own check
is worse than no label.
k_rules, k_linked and k_unlinked are registered facts under facts-check.
A limit of that checker is recorded rather than patched: it is
line-based, so a tagged value that prose-wraps fails.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
cb-cost gains --since, so the baseline (--pin 578dcbe, 662 responses,
$135.60) and this pass (--since 578dcbe, 83 responses, $7.82) are two
disjoint windows over the same transcripts. Every verdict is on share of
pass, since the windows differ 17x in size.
Test 1 — did mechanical turns disappear? Partly. Mechanical share fell
38.2% -> 32.4%. Six points against a predicted 25-30; reported unmet, and
no target was moved in this commit.
environment setup 11.5% -> 0.6% (85 turns -> 1) met, decisively
hub + workplan 8.5% -> 0.0% (46 turns -> 0) met
text patching 10.4% -> 9.6% not met
orientation 5.1% -> 21.2% not met, worse
The confound is stated before any defence of the numbers: this is the
pass that built the tools, and the two categories that missed are exactly
the two whose tools were under construction. The clean test is the next
pass, and it is carried forward rather than waived.
Test 2 — did the work relocate? Not into prose. Output tokens per
response fell 896 -> 681. Output's rising share of cost is a shrinking
denominator, not more writing. Cost per response halved and the evidence
refuses to claim it: that is compaction (mean context 232,982 ->
117,822), and attributing it to tooling would repeat CB-WP-0002's
original error in a new direction.
Test 3 — did quality hold? Yes, recorded as explicit judgment. make all
green with two gates that did not exist before, and four findings
surfaced this pass, three caught by controls written this pass — one of
them a 5.2x attribution error that would have reached the hub as a
measured number.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Duplicated-fact drift is the fourth error class and the only one with no
executable rule. No positive control catches it (both copies are
internally consistent) and re-derivation does not either (the copy
reproduces whatever it was copied from). It is caught only by reading a
copy against its source, which nothing in the loop required.
facts.toml holds 15 facts and is GENERATED by `make facts-gen` from
cb-cost, dep-weight and rule-coverage. The trap this task named — a
hand-maintained registry that becomes another drifting copy — is closed
by facts-check re-running the instruments and failing when the committed
registry disagrees with them. A stale registry cannot certify stale
artifacts.
An artifact quoting a fact tags it: **$93.15** <!-- fact --> with the key.
17 occurrences across 5 artifacts are now checked.
Falsified before being believed: changing CostAccounting.md line 158 from
$93.15 to $92.87 — the exact historical drift — produced exit 1 naming
the file, the line and the expected value. Tested against the class it
exists to catch, on a real artifact, not only in its self-test.
It then caught a live tag inside its own documentation example in
InnerLoop.md within the hour. Third time a gate has failed on its own
pass's work.
What it does not close is stated rather than implied: 22 untagged literal
copies remain and are reported, not failed. Tagging is opt-in, a number
can legitimately recur, and a gate that fires on coincidence gets routed
around. Naming the uncovered surface beats claiming the class is closed.
InnerLoop single-source-of-fact moves from prose to executable — v1.3.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Replaces the three hand-done steps of a task close (46 turns, $11.52 per
CB-RES-0003): the heredoc flipping status in the workplan file, the
hand-written hub call, and the hand-typed token counts.
The third is the reason this task exists. Every update_task_status this
repo produced carried estimated tokens_in/tokens_out — in a project whose
central finding is that estimated token counts are worthless. task-done
reads the measured figure from the transcripts, or refuses; there is no
path through it that emits an estimate.
cb-cost gains by_task_detail: cost, response count, model histogram and
token components per task. task-done imports cb-cost rather than parsing
its printed table, so the hub figure is not a copy that can drift from
its source.
The positive control found a real defect before the tool ran once.
Attribution keyed on a bare T\d\d from the commit subject, so CB-WP-0002
T01, CB-WP-0003 T01 and CB-WP-0004 T01 shared a bucket: the self-test
reported $12.10 for "T01" where the qualified figure is $2.33. That 5.2x
overstatement would have been pushed to the hub as a *measured* number —
the same fiction in a new form. task_label() now keys qualified subjects
on the full id and leaves unqualified ones bare rather than
retro-assigning them to a workplan. The pinned $93.15 benchmark is
unchanged, so historical attribution was not disturbed.
Fourth instance of trusted arithmetic: a number believed because a
program produced it rather than a hand.
Refusals, all exercised by --self-test: unknown id, typo'd id,
already-done task, missing state_hub_task_id, no measured spend, and a
status flip that produced no change.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
CB-RES-0003 measured 84 turns and $15.33 — the largest mechanical
category — spent prefixing commands with `cd` and
`export PATH="$HOME/.cargo/bin:$PATH"`. Both causes are now fixed once
instead of per-leaf.
tools/repo.py resolves the repo root from __file__ and cargo from PATH
then the standard rustup locations. Every tool imports ROOT from it, so
the repo path is stated once rather than redefined in four files —
single source of fact, the rule DFD earned in InnerLoop v1.2.
rule-coverage and dep-weight now call enter_root(), which is why their
relative paths did not need rewriting one by one.
The Makefile derives REPO from MAKEFILE_LIST and resolves CARGO the same
way, so `make -C <repo> <target>` works from any directory with no
prefix.
make env-test is the positive control, and is in `make all`: every tool
runs from / with PATH=/usr/bin:/bin. Without it this fix could regress
silently and invalidate T05's measurement — the whole point of the
control loop.
dep-weight's "cargo not on PATH" error is kept rather than deleted. It
should now be unreachable, and --self-test asserts cargo_bin() resolves
unaided; a control that never fires is cheaper than a regression.
loop-lint failed on repo.py on its first run — a reporting tool with a
positive control but no --self-test entry point. Second time the gate
has caught work from its own pass within the hour.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Review of where token-priced turns did work a deterministic tool could
do. Method: classify every turn in both transcripts by the tool calls it
made. The classifier is committed in tools/cb-cost.py and emitted by
`make cost-mix`, so the baseline is reproducible and the same command
can later falsify the predictions.
mech environment setup 84 turns $15.33
mech ad-hoc text patching 75 turns $13.86
git 37 turns $13.85
mech hub task status 25 turns $ 7.46
mech orientation / inspect 49 turns $ 6.87
hub other 32 turns $ 6.22
mech ad-hoc transcript 39 turns $ 4.56
mech workplan status edit 21 turns $ 4.06
MECHANICAL (dedup) 290 turns $51.26 = 38% of pass
Largest category is `cd` and `export PATH` -- pure friction, and
dep-weight.py already patched it at the leaf, which is evidence it was
noticed and fixed in the wrong place. Second is heredocs string-patching
markdown, which is also the mechanism behind duplicated-fact drift, the
error class InnerLoop v1.2 names and cannot gate.
Explicitly NOT automated: git (37 turns, $13.85) is mostly commit
message authorship -- the highest-output turns in the corpus and the
project's reasoning record. Automating it would save money and destroy
what makes corrections cheap.
CB-WP-0004 implements five candidates and predicts $33-41 recovery
(25-30%), below the 38% measured share on purpose: some inspection and
patching is genuinely exploratory.
The control loop is the deliverable, not a formality. T05 tests three
things and must report all: did mechanical turns disappear, did they
RELOCATE into prose, and did quality hold. If mechanical turns fall and
prose rises by as much, the saving is zero and that is the result to
publish.
Also a self-indictment worth recording: every hub update_task_status in
this project carried hand-typed token estimates, in a repo whose central
finding is that estimated token counts are worthless. T02 fixes it by
reading measured values from cb-cost.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The price sheet had two defects of one shape -- a schema that could not
hold the fact it needed, the same criticism the cost survey levelled at
the State Hub.
CA-16 time-boxed rates are DATA. Sonnet's intro price lived in a
`# intro ...` comment and was invisible to the collector that
reads the file. Now promo_input/promo_output/promo_until,
applied per response at its own timestamp.
CA-17 the 90-day staleness rule was prose in MetricsAndScenarios 1a
that every M-D2-CST verdict silently inherited. Now `recorded`
+ `max_age_days` in the sheet, and a stale sheet ABORTS.
Both are exercised by make cost-test: the promo rate must apply before
2026-08-31 and lapse after, and a 102-day-old sheet must trip.
Applying CA-16 moved AC-1 from $93.32 to $93.15 -- the $0.17 CB-EV-0002
predicted, now collected rather than noted. That is a legitimate
retarget under T07's distinction: the instrument disproved the target,
and its output is in this commit. The number has now been stated five
times ($248.46, $92.21, $92.87, $93.32, $93.15), each correction from a
different mechanism.
Evidence tables regenerated from the tool rather than hand-patched,
per CA-15 -- which is the rule that exists because hand-typed tables
were the only thing the adversarial review found wrong.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The 8k/10k per-task token budget was never referenced or enforced and
T08 blew past it silently. Replaced with a budget that can actually
fire.
The design constraint is the interesting part: per-task cost needs the
commit that CLOSES the task, so a per-task budget is unavoidably
retrospective -- it can only report a breach after the money is spent,
which is the dead-policy failure again. What IS observable mid-task is
spend since the last commit, because the transcript is append-live. So
the budget binds on the open remainder.
CB-01 budget = USD since the last commit, via `make cost-budget`
CB-02 soft $10.00 (state progress, decide), hard $22.00 (stop)
Calibrated on the 32 non-empty commit intervals of CB-WP-0001: p50
$1.40, p90 $9.36, max $10.80. Soft sits just below the observed maximum
-- it would have fired exactly once on the calibration pass. Hard is ~2x
the observed max, a value never reached in 32 intervals, so reaching it
means the session is doing something the data has no example of.
Both thresholds are set ABOVE every observed value, so they bind on
future work rather than ratifying present work -- the distinction T07
is about.
Stated limit: it is a command, not a daemon. An agent that never runs it
gets no signal, which is the dead-policy failure one level up. Mitigated
only by being free to run and on the one command surface.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
AC-1 $93.32 ($92.21 main + $1.11 subagent), residual $0.000000. No
unmeasured rows. Composition: 88.0% cache, 11.9% output, 249:1 context
re-read to text written.
The question that could not be answered before: what /compact costs.
The transcript records it directly — 542,991 -> 19,974 tokens, a 27x
reduction. Cost per turn fell from $0.457 (136 turns before) to $0.149
(202 turns after). The 202 post-compact turns cost less than half the
136 before them.
This qualifies CB-WP-0003 T04: sessions are bounded-quadratic, not
quadratic. Cost grows with context between compactions and resets at
each one, so the failure mode is a long UNCOMPACTED session. Whether a
fresh session beats a compaction is now measurable, and T04 should
measure it before prescribing one task per session.
Stale $0.66/$92.87 targets updated across survey and workplan.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
