What this page is & Why it exists
What this page is. An index, not a measurement. It introduces no new figures. Every claim Skill Heaven / Hell has made in public during Arc I is listed here with the record it traces to — a committed
hh-ledger/v1line, a committed census field, a committedharness-probe/v1run record — or it is marked ‡ = declared uncommitted workstation context, or it is named as not probed and carries no number at all.Why it exists. B4 (
founder/RATIFICATION.md): "The ledger is always on, the claim-discipline table binds all public copy, and no claim ships ahead of its benchmark. A 'will not work' ledger is as first-class as a 'will work' one — verified negative findings are recorded with the same rigor." B4 says public copy is bound to the ledger. Until this page existed, that binding was asserted; a reader had no route from a sentence on the site to the JSON line under it. This is that route.This page is machine-gated. It sits in
content/reports/hh-benchmark/, whichscripts/hell-heaven-bench/check-claims.tsscans by derived directory listing — so it is checked on every run, and so is the next report anyone drops beside it. Every number below either equals a committed value or carries ‡. Run it yourself:npx tsx scripts/hell-heaven-bench/check-claims.ts npx tsx scripts/hell-heaven-bench/ledger.ts validate
Methodology & Status Vocabulary3 Evidence Classes6 Status Definitions
The three evidence classes (and which ones the gate can read)
| Class | Where it lives | Gate reads it? |
|---|---|---|
hh-ledger/v1 records | scripts/hell-heaven-bench/data/ledger.jsonl | yes |
| Two-part-dose census | content/reports/hh-benchmark/data/r0-census.json | yes (per-record doses + per-surface standing sums) |
harness-probe/v1 run records | scripts/hell-heaven-bench/harness-probes/runs/ | no — committed, reproducible, but structurally invisible to the gate (its findings are listing shapes, not token counts). Named as the next natural extension, not a silent gap. |
A claim backed by the third class is committed and reproducible but not machine-bound — this page cites its file path so a reader can open the bytes.
Status vocabulary
- RECORD — traces to a committed ledger line or census field; the gate binds it.
- RUN RECORD — traces to a committed
harness-probe/v1file; cite the path. - ‡ UNCOMMITTED — real workstation measurement whose artifact was never committed. Honest, corroborating, never load-bearing.
- NOT PROBED — no measurement exists. No number appears, anywhere, in any form.
- SOFTENED — published copy overstated; the fix is the copy, not the number.
- OUT OF SCOPE — a different program's claim, stated rather than silently omitted.
Inventory View:
29 claims total (13 RECORD · 9 ‡ UNCOMMITTED · 7 OTHER)
Section A — The Live Site4 RECORD1 SOFTENED1 OUT OF SCOPE
A — the live site
| # | Claim as a reader sees it | Where | Record | Status |
|---|---|---|---|---|
| A1 | top-five loadout reduced the standing dose from 9,453 to 249 tokens — −97.4% | content/blog/claude-5-system-prompt-shrink/post.md | census h1Restatement.publishedStandingTokens.{allGraphNodeListings, top5Evidenced} | RECORD |
| A2 | "evicts every installed skill … admits back only the grilling-native ones" | app/page.tsx | contradicted. The KC4 clean-room probe shows curated's listing is ["heaven-set:kc4-curated-marker","doctor"], not empty; P8 (RATIFICATION.md) rules doctor irreducible while the door is open and rules absolute zero internal-only, never public copy. "grilling-native" rides gate (e), still unverified, on the authority of D13, retired 2026-07-24; shipped P6 admits the user's own named skills. | SOFTENED — copy fix owed |
| A3 | "one step below vanilla" / "cleaner than vanilla" | app/page.tsx | the two floors are ledger records 11 and 12 (below). The native pole it is measured against is not a record. | RECORD (floors) + ‡ (native) |
| A4 | Context Diet: 49,687 → 29,040 chars, 41.6%, 124/124 rules | same blog post | own committed artifacts under content/reports/context-diet-lab-001/ | OUT OF SCOPE — different program |
| A5 | /research/hh-benchmark method page | methodology.md | in the derived scan set; passes the gate | RECORD |
| A6 | /about Skill Heaven card | app/about/page.tsx | no figures; "the dose measured rather than guessed" is a method statement | clean |
Section B — gaia-research Markdown3 RECORD3 ‡ UNCOMMITTED1 GATE HOLE1 OUT OF SCOPE
B — gaia-research markdown
| # | Claim | Where | Record | Status |
|---|---|---|---|---|
| B1 | floor 30,661 · curated 31,624 · +963 · standing 227 · sha256:14c4642… | m2-live-demo.md | ledger records 9 and 10 (hh-m2-smoke, claude 2.1.216). The +963 is blessed as a signed delta between two committed perTurn values — it needs no record of its own. | RECORD |
| B2 | the same pair, matrix M2 rows | harness-capability-matrix.md | the same two records | RECORD |
| B3 | native 46,849 ‡ · delta −16,188 ‡ · invocation 5,917 ‡ | m2-live-demo.md | emitted to gitignored scripts/.hh-demo/; never committed | ‡ UNCOMMITTED — the reference use of the sigil |
| B4 | gate (a) ladder: ≈25.2k · ≈17.0k · ≈19.2–19.6k figures across the three rungs | harness-capability-matrix.md | workstation smoke from gate-a-resume-recomposition.sh, never committed — and structurally ungated: normalizeNum accepts plain integers only, so every k-suffixed magnitude is skipped, not checked. The matrix carries 18 such figures. Declared in the gate's own header, dated, unfixed on purpose (closing it is a scope call, not a quiet edit). | ‡ UNCOMMITTED + declared gate hole |
| B5 | gate (c) self-dose ≈57 ‡ tok · door command 24 ‡ tok | matrix; claude-heaven/README.md | chars4 over draft copy, never committed; to be re-priced at WS4. NB 57 ‡ happens to exist as a committed magnitude elsewhere, so the gate would wave it through — a live instance of the declared magnitude-existence, not record-binding limit. The sigil, not the gate, is what makes this honest. | ‡ UNCOMMITTED |
| B6 | census: the two-dose corpus totals and the 11.1× / 23.2× / 40.8× overstatement factors; graph-node listings 9,384 · named-skill listings 14,498 | r0-census.md | data/r0-census.json — contracts.records[] per-record doses and registryListings.*.standingTokens.sum. The corpus totals and distribution statistics are deliberately not restated here: the gate binds per-record doses and per-surface sums, not derived summary stats, so quoting them on an index page would assert a binding the gate does not provide. Read them on the census page, where they sit under their own table. | RECORD |
| B7 | this-repo sanity check: 6 contracts, standing 597 ‡ vs invocation 10,376 ‡ (13.8× ‡) | r0-census.md | --repo mode run; its JSON was never committed | ‡ UNCOMMITTED |
| B8 | oracle passages carrying +515, ["doctor"], skills=[], entry counts | founder/RATIFICATION.md | the oracle is deliberately outside every scan set. This page names where each of those figures' record lives instead; the oracle is not edited by KC8. | OUT OF SCOPE |
Section C — The skill-heaven Repo2 RECORD2 RUN RECORD3 ‡ UNCOMMITTED1 NOT PROBED
C — the skill-heaven repo
| # | Claim | Record | Status |
|---|---|---|---|
| C1 | F7 — the door costs +515 tok: product floor 20,176 vs benchmark floor 19,661, claude 2.1.216, probed 2026-07-24 (README.md, claude-heaven/README.md) | ledger records 11 and 12 (hh-f7-floor-split / door-cost-probe), backfilled 2026-07-30 from packages/core/src/compile.ts FLOOR_EVIDENCE (arithmetic pinned by packages/core/test/compile.test.ts). +515 is the signed delta between them. LOCKED by founder ruling 2026-07-30: cited as measured on 2.1.216 and deliberately NOT re-derived against 2.1.220 — the version drift is accepted, not an oversight. Both records disclose in notes that model was unrecorded and wallClockMs: 0 means unmeasured, not a 0 ms run (the schema has no null for it — a named gap, not a quiet normalization). | RECORD |
| C2 | F6 negative — /skill-heaven answers Unknown command at the doorless floor | record 11 carries objectiveEndpoint.pass: false. A verified negative finding, recorded with the same rigor as a positive, per B4. | RECORD (negative) |
| C3 | "−28.9% off native 28,379 ‡" (claude-heaven/README.md) | hh-ledger/v1 has no native arm, and arm: "heaven" with an empty loadout would misrepresent the vanilla pole. Following the m2-live-demo.md precedent, the native pole stays uncommitted rather than being forced into the wrong arm. The missing arm is a disclosed schema gap. | ‡ UNCOMMITTED (both the pole and the derived percentage) |
| C4 | KC4 clean room — curated listing ["heaven-set:kc4-curated-marker","doctor"], 2/2 byte-identical, claude 2.1.220, 2026-07-30 | the probe script is committed (packages/claude-heaven/scripts/probe-kc4-listing-residual.sh, on the KC4 lane branch — not yet on skill-heaven main, so this row deliberately carries no link); the observed output is committed nowhere, in this repo or that one. Script-reproducible, output not committed. Not a token count, so the gate is structurally blind to it either way. | NOT COMMITTED — declared |
| C5 | KC4 (superseded, 2026-07-29): "~64-entry" / "~68-entry" listings | approximations with no artifact. They stay approximate and never become anyone's numbers. | ‡ UNCOMMITTED |
| C6 | codex "74 → 73 entries, 2/2 byte-identical" | this does not match the committed record. codex-g1-2026-07-29.run.json reports 67 → 66 skills with input_tokens 18,986 → 18,925, 2/2 byte-identical, on codex-cli 0.145.0. The public 74 → 73 figure is a different, uncommitted probe. The committed run record is not presented as backing it. | RUN RECORD (67 → 66) · public 74 → 73 unbacked |
| C7 | pi: "0/15" leaks observed, 95% upper bound ≈21.8% | pi-race-and-argv-order-2026-07-29.run.json (n: 15, leaksObserved: "0/15", upperBound95pct: "21.8%", pi 0.82.1) | RUN RECORD |
| C8 | cursor | the binary probes are DEFERRED — no availability to run them. No measurement exists, so no figure appears here, and none is inferred from a doc-verified cell. A doc-verified capability cell is not a measurement and is never borrowed as one. | NOT PROBED |
Section D — KC9 Three-Minute Demo2 RECORD2 ‡ UNCOMMITTED3 CITED BY PATH
D — the KC9 three-minute demo
Live run, claude 2.1.220, sonnet at --effort low, 2026-07-30. One task
asked three ways with a byte-identical prompt and a single shared objective
endpoint — the loadout is the only variable, so "curated succeeds" is a result
rather than a definition. Writeup: kc9-three-minute-demo.md, in this
directory and therefore in the derived scan set.
| # | Claim | Where | Record | Status |
|---|---|---|---|---|
| D1 | floor 30,601 · curated 55,924 · standing 227 | kc9-three-minute-demo.md | ledger records 13 and 14 (hh-kc9-demo / side-stripe-review). The floor→curated +25,323 is blessed as a signed delta between two committed perTurn values and needs no record of its own. | RECORD |
| D2 | only the curated arm solves the task; both other arms fail the same endpoint | kc9-three-minute-demo.md | the same two records carry objectiveEndpoint.pass false / true against the identical regex. The floor's false is a verified negative, recorded with the same rigor as a positive, per B4. | RECORD (negative + positive) |
| D3 | native 46,490 ‡ and the derived native−floor 15,889 ‡ | kc9-three-minute-demo.md | emitted to gitignored scripts/.hh-demo/ and never appended — hh-ledger/v1 still has no native arm (C3's disclosed schema gap), and arm: heaven with an empty loadout would misrepresent the vanilla pole. | ‡ UNCOMMITTED |
| D4 | the demo's first run inverted: floor 64,658 ‡ against native 46,463 ‡ | kc9-three-minute-demo.md | never committed, and reported as a method finding rather than as evidence for anything: perTurn sums usage across a whole headless run, so differencing two of them prices turn count unless both runs took a comparable number of turns. Same trap the M3 paired run hit from the other direction. | ‡ UNCOMMITTED — method caveat |
| D5 | the replay page and the kc9-demo-transcript/v1 beats | kc9-demo-replay.html, kc9-demo-transcript.jsonl | committed and reproducible. The page carries no figure that is not read out of the transcript at render time, and it is not a markdown doc with numbers in prose, so the gate is structurally blind to it — same treatment as C1/C4's run-record citations. | cited by path |
| D6 | F7 and cursor in the KC9 artifacts | — | neither appears. F7 stays as C1 states it, locked and not re-derived — the demo never launches the product floor. Cursor stays as C8 states it, deferred with no figure anywhere. | unchanged by KC9 |
| D7 | the demo video served on the method page, and its spoken narration | kc9-demo.mp4, cut by record-kc9-video.mjs; words in kc9-script.mjs, audio manifest kc9-narration.json | a screen capture of the replay page's ?autoplay=1 mode, so it introduces no figure of its own: every number it shows or says is the one D1/D3 already place, interpolated from the transcript rather than typed into a line. The spoken and on-screen text are the same string, so the narration cannot assert something the caption does not. Both are binaries or code the gate cannot read, and both derive from artifacts it already covers — same cited by path treatment as D5. | cited by path |
What this Index Does & Does Not SettleSettlement Scope
What this index does and does not settle
- Does: give every public Arc I figure a named home — a ledger line, a census field, a run-record path, or an explicit ‡ / not-probed marker; put the page itself inside the machine gate rather than beside it; record the two places public copy is currently ahead of its evidence (A2's eviction wording, C6's 74 → 73) instead of quietly restating them.
- Does not: fix A2's copy (a visitor-visible change, owed a separate
human-gated pass), close the k-suffix gate hole, teach the gate to read
harness-probe/v1records, or re-derive F7. Each of those is named above with its reason, which is the point: a declared gap is first-class evidence; a silent one is a defect.