Scope cut first, on the maintainer's decision after a spend review: the
project is 38% product / 62% loop-meta, cost per response is 2.9x worse
than its best window, and INTENT stage 0 still lacks a CLI player and
bots. CB-WP-0005 and CB-WP-0006 cost ~$74 — 31% of all spend — for zero
measured efficiency gain. T02 and T04 are cancelled unstarted.
T01: SH-1/SH-2/SH-3 now report over the window since the last commit, and
the cumulative figure is retained but labelled "history, NOT the metric".
The prediction held decisively — window 655,744 mean context against
cumulative 255,307, a 2.6x gap against a 20% refutation threshold. A
cumulative mean over 1,094 responses cannot detect a worsening trend
because the history outvotes the present.
T03: `make shape-budget`, modelled on CB-01/CB-02. Soft thresholds are the
existing SessionShape targets; hard is 1.5x, set before the next
measurement per §Step 4. Deliberately not in `make all` — failing the
build on context would block committing, and committing is what closes the
attribution window and is the natural point to compact, so a gate that
blocks the remedy is a trap. It fires HARD on its first run: 656,574
against a 300,000 ceiling.
InnerLoop v1.5 establishes the soft 25% meta budget. Workplans declare
kind: product|meta|mixed and `make status` reports the share; mixed splits
50/50 and says so. Soft on purpose — a task already started may be
finished, because stopping mid-task to satisfy a ratio wastes the work.
What it forbids is opening new meta work above the line. A pass that
exceeds it must say so in its evidence and name the product work
displaced.
First reading: 68% OVER, of $74.22 attributed. Product reads $0.00 because
the only product workplan, CB-WP-0001, predates qualified task ids and its
bare T## labels collide across passes — stated in the output rather than
papered over.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The retrospective question was whether the mechanism set is complete or
each pass still finds a new class. This pass produced both an eighth
instance AND a fourth class, so the answer is the uncomfortable one.
Ledger: 10 instances, 4 classes, across 3 workplans. Every pass has
produced at least one class the previous pass had not seen.
HDN harness-does-nothing 5 executable assertions
TA trusted arithmetic 3 re-derivation
SSB same-sample blind spot 1 assertions over ALL the data
DFD duplicated-fact drift 2 NEW -- reading a copy against source
DFD is genuinely distinct: no positive control catches it, because both
copies are internally consistent, and re-derivation does not either,
because the copy faithfully reproduces what it was copied from. Found
when an inlined price sheet went stale within an hour of T11 changing
the real one.
So v1.2 stops trying to enumerate classes in advance. Every error in
three passes was corrected in-session for under ~1% of the pass, so the
stated design goal is now cheap CORRECTION: keep raw data so numbers are
re-derivable, keep artifacts small and committed so a wrong number is
one grep from everywhere quoting it, give every number a command.
Plus the one rule the new class earns: single source of fact.
The original hypothesis is revised rather than confirmed. "A rule that
cannot be executed is not a rule" is wrong -- the two most valuable
corrections in the project came from a decorative rule that cannot be
automated (re-derive inherited numbers). An executable rule fires
reliably and catches one class; a decorative one fires unreliably and
can catch any class, including unnamed ones. Keep both.
Gates this pass: loop-lint caught 3 real violations on first run, then
failed on its own author within the hour when a T07 edit pushed
InnerLoop.md to 407 lines against its own 400 limit.
CB-WP-0003 complete: 11 of 11 tasks done.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Five scenarios encoded U-item defaults with no owner and no review date,
so they could shape the kernel indefinitely while looking handled.
Each now carries provisional_owner and provisional_raised, and the
runtime's ScenarioFile learned both fields (deny_unknown_fields meant
adding them to YAML alone would have failed every scenario -- the parser
had to agree).
make coverage reports every provisional item with its owner and age in
days, warns on any with no owner, and warns past 30 days. It WARNS
rather than breaking the build, on purpose: the ruling is ground-game's
to make and the kernel cannot make it for them. What the loop can
enforce is that evidence files list them, which is now stated in
GroundRules.
rule-coverage --self-test gained an assertion that every provisional
item has both fields, so the next one added without them fails a
command rather than passing quietly.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
make coverage now reports a second number: how many claimed rules are
also NAMED in the aggregate source. 58/58 tag coverage was weaker
evidence than it read as, and this says how much weaker.
Measured: 49 of 58. Nine rules are claimed by a scenario and appear
nowhere in games/ground/src/lib.rs --
GR-D07 GR-F02 GR-L03 GR-O03 GR-P01 GR-P02 GR-P03 GR-P04 GR-T01.
The gate REPORTS rather than fails, on purpose. Closing the gap by
adding those IDs to comments would satisfy the check without
establishing that any of the nine is implemented -- the overclaim
InnerLoop implementation rule 2 exists to prevent, and one CB-WP-0001
already committed once. Each needs its implementation confirmed before
it is tagged; promoting M-D1-LNK to a failing gate is correct after
that, not before.
Also added: a phantom check that fails when a rule id appears in code
that the spec does not define (currently zero).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The review step targeted the survey. Every serious error in this project
has been in measurement or build configuration, so the loop was
adversarially reviewing the artifact cheapest to fix and leaving
unreviewed the one where errors occur.
Step 2 now routes by risk: when the claim rests on numbers, the reviewer
gets the harness and the evidence file too, and must reproduce the
number independently rather than read about it.
The addition that matters more, because it was learned the hard way: a
reviewer re-derives the author's claims and therefore inherits the
author's SAMPLING. CB-WP-0002's dedup invariant was checked twice --
survey 206/206 groups, then the reviewer independently -- and both used
the main transcript. It is false in the 8-response subagent tree neither
looked at. Two independent verifications, one shared blind spot.
Rule: the reviewer re-derives on a different sample than the author
used, and where only one sample exists, says so rather than reporting
a clean verify.
Also recorded: what review demonstrably DOES do. $0.66 and $1.11 across
two passes, ~1% of each, both finding approval-blocking defects. Cost is
not a reason to skip it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A blanket "no retargeting in the measuring commit" rule would have been
wrong. CB-WP-0002 moved AC-1 three times in exactly that shape and every
move was correct -- each time a new instrument disproved the old figure.
Four legitimate corrections would have been forbidden to catch one bad
retarget.
The test is mechanical rather than a statement of intent:
correction the target moves and the implementation does not; legal in
the same commit provided the instrument's output is there
retarget the same commit changes both the target and the code the
target measures; requires an ADR stating why the new target
binds on future work
Applied retroactively: AM-4a/AM-4b are UNRATIFIED. They were set after
seeing the measurement, in the commit that produced it, with the
implementation changing too -- a retarget by this test. make dep-weight
is currently enforcing a target no reviewed decision stands behind.
Recorded as an open item; ratifying or changing them is a maintainer
decision, not an implementer's.
Also: specs/InnerLoop.md split into InnerLoop.md (process) and
InnerLoopReference.md (rubric, template, rules, definition of done).
Not a stylistic choice -- `make loop-lint` failed on the commit that
pushed the file to 407 lines against its own ~400 limit. The gate added
this morning to make that rule executable caught its own author within
the hour, which is the cheapest possible demonstration that it works.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The mechanism never fired. Two rolls across two workplans (9, then 2)
against ~0.2 expected firings at d10. At ~2 tier decisions per workplan
it would take roughly twenty workplans to see four overrides -- a rate
that prevents its own evaluation, which T06 named as the one option to
reject.
Raised to d4 for the next 12 tier declarations, then kept, reverted to
d10, or deleted on the evidence. Expected ~3 firings.
Cost stated up front rather than discovered: a chaos-L override on
S-shaped work buys a full survey + review + ADR, and the comparable
measured figure is CB-WP-0001 T03 at $9.91. The window should cost
$20-30. Cheaper than carrying an unevaluated ritual indefinitely.
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>
The task's original premise was wrong and is recorded rather than
deleted. It was written to prescribe one task per session; measurement
says the variable is context, not turn count.
Measured:
- compaction cut context 27x (542,991 -> 19,974) and cost/turn 3.1x;
the 202 turns after C1 cost less than half the 136 before it
- a turn costs $0.010 at 20k context and $0.270 at 540k
- break-even for a compaction is 2-11 turns, so: compact whenever
context exceeds ~300k and work remains
- a fresh session is NOT free -- cold start floors at ~51k and must
then re-read the artifacts a compact summary already holds (~66k).
Prefer compaction to continue work; prefer a fresh session when the
task changes, because then prior context is pure overhead.
cb-cost now emits SH-1..SH-3 so the targets come from the instrument
rather than from analysis, per InnerLoop v1.1. All three are UNMET
(mean context 232,982 vs 200,000; p90 492,042 vs 300,000; batching 7.8%
vs 20%) and are reported unmet rather than retargeted -- retargeting in
the commit that first measures is precisely what T07 exists to prevent.
Eighth error instance found while writing this: CB-WP-0001's claim that
"0 of 330 tool calls were batched" is wrong. 330 was the count of
single-call responses, not the total; 31 responses batched, covering 76
calls. It was carried into this workplan unverified. Trusted-arithmetic
class -- the one the T01 audit flagged as having no executable defence,
confirming that finding within hours of making it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Both items were completed as part of the T01 audit: the two older
reporting tools gained --self-test entry points, and loop-lint's
self-test check makes the contract mechanical for any new tool.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
41 rules classified executable / checkable / decorative, each tagged with
the failure class it catches. Counts: 11 executable, 22 checkable, 4
decorative (one of them dead policy).
Audit: history/260731-inner-loop-rule-audit.md
New tools/loop-lint.py makes 7 rules executable (tier declared, chaos
roll recorded, tier-L review trail, unmeasured-in-evidence, whole-file
loadability, reporting tools expose --self-test). It found three real
violations on its first run, none previously visible:
- specs/ArchitectureBlueprint.md was 543 lines against a ~400 limit
the loop has stated since v0.2 and never measured. Split at its own
section boundaries into Blueprint (1-8) + Runtime (9-15).
- tools/dep-weight.py and tools/rule-coverage.py had positive-control
logic and no --self-test, so nothing verified the control worked.
Adding rule-coverage's self-test exposed a latent instance of the exact
class this workplan is about: if the spec regex stopped matching, rules
was empty, missing was empty, and the tool exited 0 reporting "0/0" --
a silent pass, in the tool that reports our headline AM-1 number. Both
tools now assert they found something before reporting.
Two demotions applied in the spec rather than left implicit: "structured
over prose" is marked guidance (nothing can check it), and the 8k/10k
token budget is struck through and marked DEAD POLICY pointing at T05.
The audit's uncomfortable finding: rule 13 (re-derive inherited numbers)
has no mechanical form, is deliberately left decorative, and caught the
LARGEST error in CB-WP-0002. That is a counter-example to this
workplan's own hypothesis. "A rule that cannot be executed is not a
rule" is wrong as stated; the defensible version is that such a rule
cannot be relied on to fire, so it must not be the only defence for a
class that matters.
Class coverage: harness-does-nothing has five executable rules;
trusted-arithmetic has ZERO and produced the largest single error.
make loop-lint and make self-tests wired into `make all` and CI.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The framing hypothesis -- a rule that cannot be executed is not a rule
-- survives but was too coarse. CB-WP-0002 corrected one number four
times and each correction came from a DIFFERENT mechanism, none finding
more than one. Two limits are now measured rather than assumed:
- a positive control cannot catch trusted arithmetic (both errors in
$248.46 were sums over data that really existed)
- adversarial review cannot catch a same-sample blind spot: survey and
reviewer both verified dedup on the main transcript, and it is false
in the subagent tree neither examined
Task-level revisions:
T01 also classify each rule by WHICH failure class it catches; seven
instances across three classes are now on record
T02 scope reduced -- the --self-test contract, cb-cost's five
assertions, make cost-test in CI are delivered; remaining work is
rule-coverage/dep-weight plus a check that new tools carry one
T03 now backed by two measurements ($0.66, $1.11 at ~1% of a pass),
and must also state what review CANNOT do
T04 premise overturned -- sessions are bounded-quadratic, not
quadratic. /compact cut context 27x and cost/turn 3.1x. Measure
compaction against a fresh session before prescribing either
T05 unblocked, with calibration data ($1.29-$21.02/task) and the
constraint that a budget binds only on the 67.5% attribution
reaches, only after the closing commit
T07 sharpened -- CB-WP-0002 retargeted three times, correctly, so a
blanket prohibition is wrong; separate instrument-disproved-target
from implementation-missed-target
T10 its prediction is already falsified; the successor question is
whether class-by-class hardening converges at all
T11 NEW -- the price sheet cannot express a dated rate and becomes
wrong on 2026-08-31
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Written back by fix-consistency: workstream and task UUIDs for the ten
CB-WP-0003 tasks, plus the regenerated WORK-RECORDS.md.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Covers what the immediate CI fixes did not. Framing finding: writing a
rule into InnerLoop did not prevent the next instance of the failure it
was written for; making it a CI step caught one on its first run. Four
instances of the harness-does-nothing class, zero prevented by prose.
Working hypothesis: a loop rule that cannot be executed is not a rule.
Each item is made executable or deleted, default delete.
Phase A makes rules executable: audit every InnerLoop rule as
executable/checkable/decorative, extend the positive-control gate beyond
benchmarks, and point adversarial review at measurement rather than only
research — all four real errors were in measurement and build config,
which the review never reads.
Phase B is session economics, from measurement: 592 turns, mean context
245k tokens, $0.245/turn in cache reads, zero batched tool calls, cache
reads 58% of cost against output at 12%. Cost is turns x context and
context grows with turns, so a long session is quadratic. Also replaces
the dead 8k/10k token budget, which was never enforced and which T08
exceeded by orders of magnitude with no signal.
Phase C removes or fixes the rest: resolve the chaos roll (never fired,
and at 1-in-10 will stay unevaluated), require retargeting to be an ADR
rather than something the implementer does in the commit that measures
the metric, give provisional items an owner and expiry, and strengthen
the coverage gate beyond tag-counting.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
