CB-WP-0007 T01+T03: window the metric, budget it, cap meta at 25%

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>
This commit is contained in:
tegwick 2026-08-01 14:12:07 +02:00
parent 98e09c4394
commit b79ea9690d
11 changed files with 229 additions and 26 deletions

View file

@ -24,7 +24,7 @@ TOOLS := $(REPO)/tools
# Every cargo recipe runs at the repo root; the shell does not persist cd.
IN_REPO := cd $(REPO) &&
.PHONY: check test sim bench bench-test coverage dep-weight cost cost-test cost-pin cost-budget cost-mix loop-lint self-tests env-test task-done status facts-check facts-gen mutation-check size-metrics runtime-metrics build-time am6 replay-test loc all
.PHONY: check test sim bench bench-test coverage dep-weight cost cost-test cost-pin cost-budget shape-budget cost-mix loop-lint self-tests env-test task-done status facts-check facts-gen mutation-check size-metrics runtime-metrics build-time am6 replay-test loc all
## fmt + clippy (deny warnings) + HashMap deny-lint
check:
@ -140,6 +140,13 @@ task-done:
@test -n "$(T)" || { echo "usage: make task-done T=CB-WP-0004-T02" >&2; exit 2; }
$(PY) $(TOOLS)/task-done.py $(T) $(ARGS)
# CB-WP-0007 T03 / InnerLoop v1.5: session shape for the window since the
# last commit. Deliberately NOT in `make all` — failing the build on
# context would block committing, and committing is the natural point to
# compact. A gate that blocks the remedy is a trap.
shape-budget: cost-test
$(PY) $(TOOLS)/cb-cost.py --shape-budget
# CB-01/CB-02: live spend since the last commit.
cost-budget: cost-test
$(PY) $(TOOLS)/cb-cost.py --budget

View file

@ -1,6 +1,31 @@
# The Inner Loop — Assimilate and Surpass
Status: **v1.4** — corrected from CB-WP-0005 (assertion coverage) on
Status: **v1.5** — corrected from CB-WP-0007 (session shape) on
2026-08-01. Change from v1.4: a **soft 25% meta budget** (below), after a
spend review found the project 38% product / 62% loop-meta with cost per
response degraded 2.9x from its best window.
> **Meta budget — soft, 25% of spend per pass.** Work on the loop's own
> instruments and process is capped at a quarter of a pass. Measured by
> `make status` from each workplan's `kind:` frontmatter
> (`product` | `meta` | `mixed`).
>
> **Soft on purpose.** A task already started may be finished — stopping
> mid-task to satisfy a ratio wastes the work and leaves the tree in a
> worse state than either finishing or never starting. What the budget
> forbids is *opening* new meta work above the line.
>
> A pass that exceeds it **says so in its evidence file and names the
> product work displaced**. That is the whole enforcement: this is a
> reporting budget, not a gate, for the same reason the session-shape
> budget is (CB-RES-0005 §4) — it constrains judgment, not artifacts.
>
> *(v1.5, from the CB-WP-0007 spend review: CB-WP-0005 and CB-WP-0006 cost
> ~$74, 31% of all spend, for zero measured efficiency gain. Their return
> was correctness of claims, which is real and is not optimization. The
> budget exists so that distinction has to be made out loud.)*
v1.4 — corrected from CB-WP-0005 (assertion coverage) on
2026-07-31. Change from v1.3: where a claim rests on numbers, the
adversarial reviewer must read the assertion behind each quoted number and
**mutate it** — re-running the command that prints a number is not

View file

@ -396,7 +396,8 @@ def collect(slug, pin_ref=None, since_ref=None):
for r in responses:
r["cost"] = price_of(prices, r["model"], r["toks"], r["timestamp"])
attribute(responses, commit_index(pin))
commits = commit_index(pin)
attribute(responses, commits)
by_task = collections.defaultdict(float)
# T02: per-task token detail, so a task close can report *measured*
@ -454,9 +455,28 @@ def collect(slug, pin_ref=None, since_ref=None):
mech = [r for r in responses
if set(r.get("categories") or []) & MECHANICAL]
# The window the SH-* targets compare against: responses since the
# last commit when unpinned, or the whole pinned range when pinned
# (a pin is already a window someone chose deliberately).
if pin or since:
window_responses, window_label = responses, "the pinned/--since range"
else:
last = commits[-1][0] if commits else ""
window_responses = [r for r in responses if last and r["timestamp"] > last]
window_label = "since the last commit"
sub = sum(r["cost"] or 0 for r in responses if r["subagent"])
return {
"session_shape": session_shape(responses),
# CB-WP-0007 T01. SH-1/SH-2 as a cumulative mean over every
# response ever recorded cannot detect a worsening trend: the
# history outvotes the present. CB-WP-0006 ran at 503,464 mean
# context and the cumulative figure still printed 206,952.
# Both are reported; only the window is the metric.
"session_shape_window": session_shape(window_responses)
if window_responses else None,
"window_responses": len(window_responses),
"window_label": window_label,
"tool_mix": {
"turns": dict(mix_turns),
"cost": dict(mix_cost),
@ -550,16 +570,33 @@ def render(rep, by_task=False, composition=False):
f"{mix['mechanical_turns']:>5} turns ${mix['mechanical_cost']:>8,.2f}"
f" = {100*mix['mechanical_cost']/(rep['total'] or 1):.0f}% of pass")
def _shape_lines(sh, indent=" "):
print(f"{indent}SH-1 mean context {sh['SH-1_mean_context']:>12,.0f} tok "
f"[{'ok ' if sh['SH-1_mean_context']<=200_000 else 'FAIL'} target 200,000]")
print(f"{indent}SH-2 p90 context {sh['SH-2_p90_context']:>12,.0f} tok "
f"[{'ok ' if sh['SH-2_p90_context']<=300_000 else 'FAIL'} target 300,000]")
print(f"{indent}SH-3 batching rate {100*sh['SH-3_batching_rate']:>11.1f}% "
f"[{'ok ' if sh['SH-3_batching_rate']>=0.20 else 'FAIL'} target 20.0%]")
print(f"{indent} {sh['tool_calls']} tool calls in "
f"{sh['responses_with_tools']} responses; "
f"{sh['calls_in_batched_turns']} in batched turns")
win = rep.get("session_shape_window")
print(f"\n session shape (specs/SessionShape.md) — window: "
f"{rep.get('window_label', '?')}")
if win:
print(f" THE METRIC — {rep['window_responses']} response(s)")
_shape_lines(win)
else:
print(" THE METRIC — no responses in the window "
"(nothing since the last commit)")
sh = rep["session_shape"]
print("\n session shape (specs/SessionShape.md)")
print(f" SH-1 mean context {sh['SH-1_mean_context']:>12,.0f} tok "
f"[{'ok ' if sh['SH-1_mean_context']<=200_000 else 'FAIL'} target 200,000]")
print(f" SH-2 p90 context {sh['SH-2_p90_context']:>12,.0f} tok "
f"[{'ok ' if sh['SH-2_p90_context']<=300_000 else 'FAIL'} target 300,000]")
print(f" SH-3 batching rate {100*sh['SH-3_batching_rate']:>11.1f}% "
f"[{'ok ' if sh['SH-3_batching_rate']>=0.20 else 'FAIL'} target 20.0%]")
print(f" {sh['tool_calls']} tool calls in {sh['responses_with_tools']} responses; "
f"{sh['calls_in_batched_turns']} in batched turns")
print(f"\n history, {rep['responses']} responses — context only, NOT the "
f"metric.")
print(" A cumulative mean cannot detect a worsening trend; the history")
print(" outvotes the present (CB-RES-0005 §1).")
_shape_lines(sh)
if rep["unpriced"]:
print(f"\n UNPRICED ({len(rep['unpriced'])} responses, model not in sheet):")
@ -690,6 +727,63 @@ def self_test():
return 0 if ok else 1
# CB-WP-0007 T03 / CB-RES-0005 D2. Soft = the SessionShape target; hard is
# 1.5x, set here BEFORE the next measurement per InnerLoop §Step 4.
# Reported, never 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. A gate that blocks the remedy is a trap.
SHAPE_SOFT = {"SH-1": 200_000, "SH-2": 300_000}
SHAPE_HARD = {"SH-1": 300_000, "SH-2": 450_000}
def shape_budget(slug):
"""SH-* for the window since the last commit, with soft/hard verdicts."""
try:
rep = collect(slug, None)
except Abort as e:
print(f"ABORT — {e}", file=sys.stderr)
return 1
win = rep.get("session_shape_window")
head = subprocess.run(
["git", "-C", REPO, "log", "-1", "--format=%h %s"],
capture_output=True, text=True, check=True).stdout.strip()
print("session-shape budget — the window since the last commit")
print(f" last commit {head}")
# Positive control: a budget that reports ok because it measured
# nothing is the harness-does-nothing shape, in a reporting tool.
if not win or rep["window_responses"] == 0:
print(" no responses since the last commit — nothing to report",
file=sys.stderr)
return 1
print(f" window {rep['window_responses']} response(s)")
breach = 0
for key, label, val in (
("SH-1", "mean context", win["SH-1_mean_context"]),
("SH-2", "p90 context ", win["SH-2_p90_context"]),
):
soft, hard = SHAPE_SOFT[key], SHAPE_HARD[key]
mark = "ok " if val <= soft else ("SOFT" if val <= hard else "HARD")
print(f" {key} {label} {val:>10,.0f} tok [{mark}] "
f"soft {soft:,} / hard {hard:,}")
breach = max(breach, 0 if val <= soft else (1 if val <= hard else 2))
rate = win["SH-3_batching_rate"]
print(f" SH-3 batching {100*rate:>9.1f}% "
f"[{'ok ' if rate >= 0.20 else 'SOFT'}] floor 20.0%")
if breach >= 2:
print("\n HARD — compact before continuing. Context this size costs "
"~10x per turn\n against a compacted session "
"(specs/SessionShape.md §2).", file=sys.stderr)
return 1
if breach == 1:
print("\n SOFT — over target. Compaction is the remedy and it is free.")
else:
print("\n within budget")
return 0
def budget(slug, soft, hard):
"""Live cost budget (specs/CostAccounting.md §7).
@ -734,6 +828,8 @@ def main():
ap.add_argument("--composition", action="store_true")
ap.add_argument("--session-shape", action="store_true",
help="SH-1..SH-3 (always shown in the default report)")
ap.add_argument("--shape-budget", action="store_true",
help="SH-1/SH-2/SH-3 for the window since the last commit")
ap.add_argument("--budget", action="store_true",
help="CB-01/CB-02: spend since the last commit, live")
ap.add_argument("--soft", type=float, default=10.00)
@ -745,6 +841,8 @@ def main():
if args.self_test:
return self_test()
if args.shape_budget:
return shape_budget(args.slug)
if args.budget:
return budget(args.slug, args.soft, args.hard)

View file

@ -48,6 +48,7 @@ def parse_workplan(path):
m = re.search(rf'^{key}:\s*"?(.*?)"?\s*$', text, re.M)
return m.group(1) if m else None
kind = fm("kind") or "untagged"
tasks = []
for m in re.finditer(r"```task\n(.*?)```", text, re.S):
body = m.group(1)
@ -57,7 +58,7 @@ def parse_workplan(path):
return mm.group(1) if mm else None
tasks.append((f("id"), f("status"), f("priority")))
return fm("id"), fm("title"), fm("status"), tasks
return fm("id"), fm("title"), fm("status"), tasks, kind
def workplans():
@ -69,7 +70,7 @@ def workplans():
def next_task(plans):
"""First todo task of the first non-done workplan, with its heading."""
for wid, _title, status, tasks in plans:
for wid, _title, status, tasks, _kind in plans:
if status == "done":
continue
for tid, tstatus, prio in tasks:
@ -130,7 +131,7 @@ def report():
+ (f" ({len(dirty)} uncommitted)" if dirty else " (clean)"))
print("\n workplans")
for wid, title, status, tasks in plans:
for wid, title, status, tasks, _kind in plans:
done = sum(1 for _, s, _ in tasks if s == "done")
if status == "done" and done == len(tasks):
print(f" {wid} {status:<12} {done}/{len(tasks)}")
@ -161,6 +162,34 @@ def report():
except Exception as e:
print(f" UNAVAILABLE — {e}")
# InnerLoop v1.5 — soft 25% meta budget. Reported, never gated.
try:
_, _, mod = cost_lines()
rep = mod.collect("-home-worsch-clay-borg", None)
detail = rep["by_task_detail"]
by_kind = {"product": 0.0, "meta": 0.0, "mixed": 0.0}
attributed = 0.0
for wid, _t, _s, tasks, kind in plans:
for tid, _ts, _p in tasks:
d = detail.get(tid)
if d and kind in by_kind:
by_kind[kind] += d["cost"]
attributed += d["cost"]
if attributed > 0:
# `mixed` splits evenly; stated rather than hidden.
meta = by_kind["meta"] + by_kind["mixed"] / 2
share = 100 * meta / attributed
mark = "ok " if share <= 25 else "OVER"
print(f"\n meta budget [{mark}] {share:.0f}% of "
f"${attributed:,.2f} attributed (soft 25%, InnerLoop v1.5)")
print(f" product ${by_kind['product']:,.2f} "
f"meta ${by_kind['meta']:,.2f} "
f"mixed ${by_kind['mixed']:,.2f} (split 50/50)")
print(" NOTE: attributed tasks only — early workplans used bare")
print(" T## ids that collide across passes and are excluded.")
except Exception as e:
print(f"\n meta budget UNAVAILABLE — {e}")
print("\n fast gates")
code, out = run_tool("loop-lint.py")
last = out.splitlines()[-1] if out else "?"
@ -209,18 +238,18 @@ def self_test():
fh.write(sample)
tmp = fh.name
try:
wid, title, status, tasks = parse_workplan(tmp)
wid, title, status, tasks, _kind = parse_workplan(tmp)
check("workplan frontmatter parses", (wid, title, status)
== ("CB-WP-0009", "Sample", "in_progress"), f"{wid} {status}")
check("both task blocks parse with status and priority",
tasks == [("CB-WP-0009-T01", "done", "high"),
("CB-WP-0009-T02", "todo", "medium")])
check("next task is the first todo, not the first task",
next_task([(wid, title, status, tasks)])
next_task([(wid, title, status, tasks, "meta")])
== ("CB-WP-0009", "CB-WP-0009-T02", "medium"))
check("a fully-done workplan yields no next task",
next_task([("X", "t", "done",
[("X-T01", "done", "high")])]) is None)
[("X-T01", "done", "high")], "meta")]) is None)
finally:
os.unlink(tmp)
@ -232,9 +261,14 @@ def self_test():
sum(len(t) for *_, t in plans) >= 20,
f"{sum(len(t) for *_, t in plans)} task(s)")
check("every real workplan has an id and a status",
all(w and s for w, _, s, _ in plans))
all(w and s for w, _, s, _, _ in plans))
# InnerLoop v1.5: the meta budget needs every workplan classified, or
# the ratio is computed over an unknown denominator.
check("every workplan declares kind: product|meta|mixed",
all(k in ("product", "meta", "mixed") for *_, k in plans),
", ".join(f"{w}={k}" for w, *_, k in plans))
check("every real task has an id and a status",
all(i and s for *_, ts in plans for i, s, _ in ts))
all(i and s for _, _, _, ts, _ in plans for i, s, _ in ts))
# A heading must be one line. The first version used a dot-all match
# and returned the prose of the preceding task as the heading.

View file

@ -1,5 +1,6 @@
---
id: CB-WP-0001
kind: product
title: "Establish the assimilate-and-surpass inner loop via the GROUND game kernel"
status: done
state_hub_workstream_id: "a1b434dc-b1c6-46b5-bbd9-80a4e6b7620f"

View file

@ -1,5 +1,6 @@
---
id: CB-WP-0002
kind: meta
title: "Make agentic cost measurable, so D2 claims are falsifiable"
status: done
state_hub_workstream_id: "b7c22f69-fbe9-48df-9619-007db79ae338"

View file

@ -1,5 +1,6 @@
---
id: CB-WP-0003
kind: meta
title: "Harden the inner loop: executable rules, session economics, dead policy"
status: done
state_hub_workstream_id: "39d61dc0-870d-45c1-a595-bcf91f289dce"

View file

@ -1,5 +1,6 @@
---
id: CB-WP-0004
kind: meta
title: "Move mechanical turns off the token budget, and prove it worked"
status: done
state_hub_workstream_id: "6880ac78-d817-41b9-b267-f12ff9deea28"

View file

@ -1,5 +1,6 @@
---
id: CB-WP-0005
kind: meta
title: "Make the instruments count assertions, then fix what they expose"
status: done
state_hub_workstream_id: "0b95a1e3-7780-43d0-81e9-072ef7978734"

View file

@ -1,5 +1,6 @@
---
id: CB-WP-0006
kind: mixed
title: "Instrument the acceptance table, then implement what it exposes"
status: done
state_hub_workstream_id: "8a6327cc-fd5c-4e2c-a29b-b437c27d1e71"

View file

@ -1,7 +1,8 @@
---
id: CB-WP-0007
kind: meta
title: "Make session shape measurable in the window that matters, then enforce it"
status: proposed
status: in_progress
state_hub_workstream_id: "bee19b76-fb60-4dcf-94d0-9242b43e0e42"
---
@ -33,6 +34,24 @@ Per InnerLoop §Step 4, no target moves in the commit that measures it —
and CB-RES-0005 D3 states that up front, because all three targets are
currently unmet by wide margins and the temptation is obvious.
## Scope cut, 2026-08-01 (maintainer decision)
A spend review before starting found the project **38% product / 62%
loop-meta**, with cost per response degraded 2.9× from its best window and
INTENT stage 0 still missing a CLI player and bots. CB-WP-0005 and
CB-WP-0006 cost **~$74 — 31% of all spend — for zero measured efficiency
gain** (their return was correctness of claims, which is real but is not
optimization).
So this workplan is **cut to T01 and T03**, the two tasks that attack the
2.9× regression directly. **T02 and T04 are cancelled**, not deferred —
SH-4 trend reporting and the batching trial are more instrument work, and
the instrument-failure taxonomy is not converging. T06 becomes the
project-level retrospective the maintainer asked for.
**A soft 25% meta budget is established here** (T03), because the review's
finding needs a standing number, not a memory.
## Phase A — measure the right window
## Task: window the session-shape metrics
@ -58,11 +77,16 @@ target compares against.
**Refuted if** it lands within 20% of the cumulative number, in which case
the aggregation was not the problem and this pass should re-plan.
## Task: SH-4 — a windowed trend the instrument can see
## Task: SH-4 — a windowed trend the instrument can see (CANCELLED)
> **Cancelled unstarted 2026-08-01.** More instrument work, against a
> review finding that instrument work has stopped paying. The ceiling T03
> adds catches a bad pass; a trend line would catch a slow slide, which is
> a real gap — recorded as open, not built.
```task
id: CB-WP-0007-T02
status: todo
status: cancel
priority: medium
state_hub_task_id: "1d5a32ef-4086-4021-b874-24ed2160c4dd"
```
@ -119,11 +143,17 @@ Carries `--self-test`. Its positive control is the one this project keeps
needing: a budget that reports `ok` because it measured nothing must
abort instead.
## Task: batch deliberately, and report what the rate reaches
## Task: batch deliberately, and report what the rate reaches (CANCELLED)
> **Cancelled unstarted 2026-08-01.** `SessionShape.md` §4 already puts
> the ceiling at **$24 on a $93 pass** — the cheapest of the three
> metrics to move and the least valuable. Spending a task on it while
> stage 0 lacks a CLI player is the misallocation the review found.
> SH-3 remains measured, unmet at 0.0%, and unfalsified.
```task
id: CB-WP-0007-T04
status: todo
status: cancel
priority: medium
state_hub_task_id: "c284db6f-2e16-40fb-be8b-83ae5391c058"
```
@ -154,7 +184,10 @@ priority: high
state_hub_task_id: "2f78b272-e1f9-4530-bc8c-3a9d82833546"
```
Commit `evidence/CB-EV-0006-session-shape.md`. Four tests, all reported:
Commit `evidence/CB-EV-0006-session-shape.md`. **Reduced with the scope
cut** — tests 1, 2 and 4 remain; test 3 (SH-3) is cancelled with T04.
Four tests, all reported:
1. **Does the windowed metric differ from cumulative?** Against T01's
prediction; refuted if within 20%.