Published numbers

A number a reader cannot reproduce is not evidence. This table is the whole reader-facing surface's quantitative content, each figure against the exact command that produces it.

Scope, stated so a green result cannot be over-read. The audit enforces registration on README.md, MCP_LISTINGS.md and index.html: every numeric token a reader sees on those three must be a row below or a declared non-claim, and python claims_audit.py --numbers fails otherwise. CHANGELOG.md and docs/ are not token-enforced — they are covered only by the weaker check that every artifact path they name exists. Their numbers are not audited here, and reading this page as "every number in the project is backed" would be exactly the over-read it exists to prevent.

The ratio

• 350 numeric tokens are published across the 6 enforced files: README.md, docs/DEEP_DIVE.md, MCP_LISTINGS.md, index.html, compare.html, claude-code.html.
• 190 of those are quantitative claims, in 103 registry rows below.
• 78 rows (78/103) are reproducible by a command committed to this repository (REPRODUCIBLE needs nothing but this checkout; REPRODUCIBLE-WITH-DEPS needs a service or dataset we cannot redistribute, named in the command column).
• The remaining 25 are PENDING-HARNESS, EXTERNAL or WITHDRAWN.
• The other 160 tokens are declared non-claims — citation years, article numbers, ordinals, ports, example literals — each with a reason and an exact expected count, so adding one silently is not possible either.

Counts by status:

• REPRODUCIBLE — 26
• REPRODUCIBLE-WITH-DEPS — 52
• PENDING-HARNESS — 2
• EXTERNAL — 21
• WITHDRAWN — 2

The table

#	file	figure(s)	claim	status	command that reproduces it
1	MCP_LISTINGS.md	30 26 2026	WITHDRAWN: the previous '30 tools' figure, kept as the record of the correction	WITHDRAWN	python claims_audit.py --numbers
2	MCP_LISTINGS.md	68	The MCP server exposes 68 tools	REPRODUCIBLE	python claims_audit.py --numbers
3	MCP_LISTINGS.md	68	The enumerated tool list matches the server	REPRODUCIBLE	python claims_audit.py --numbers
4	README.md	10 13 3	Framework adapters: 10 of 13 verified against current upstream, 3 recorded broken	REPRODUCIBLE	python claims_audit.py --numbers
5	README.md	0	The control: with the guard off we score zero, so the number is the mechanism	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
6	README.md	0 86.7 13.3 95 3.3 26.7	Graphiti keeps the correction 86.7% of the time; resurrection 13.3%, 95% CI [3.3, 26.7]	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
7	README.md	2.0.11 53.3 46.7 95 30.0 63.3	mem0 2.0.11 keeps the correction 53.3%; resurrection 46.7%, 95% CI [30.0, 63.3]	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
8	README.md	30 2.0.11 2026 2.0.18	Sample size per system, and the exact competitor version measured	REPRODUCIBLE	curl -s https://pypi.org/pypi/mem0ai/json
9	README.md	100 0	inspeximus keeps a corrected fact 100% of the time; it never resurrects the old value	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
10	README.md	30	Sample size: 30 trials per system	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
11	README.md	13.3	The Graphiti row's raw resurrection decomposed: four pre-echo extraction misses	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
12	README.md	0 26	Graphiti's bi-temporal invalidation held 26/26 corrections that were extracted pre-echo; its 13.3% raw resurrection is four extraction misses, not echo failures	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
13	README.md	0	On echo-attributable resurrection specifically, Graphiti scores 0% -- the separator is whether the supersession link is recorded at write time, not which vendor recorded it	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
14	README.md	68	The MCP server exposes 68 tools	REPRODUCIBLE	python claims_audit.py --numbers
15	README.md	0	Mutation gate: zero seeded defects survived	REPRODUCIBLE	python tools/mutation_check_parallel.py
16	README.md	98.3 0.01	Our own production store: source populated vs actually re-checkable	REPRODUCIBLE-WITH-DEPS	curl -sO https://raw.githubusercontent.com/DanceNitra/agora/main/research/probes/can_we_reconcile_our_own_index.py && python can_we_reconcile_our_own_index.py
17	README.md	2,600 175	Suite size, and the mutation gate that makes it evidence: 175 seeded, 175 killed	REPRODUCIBLE	python tools/mutation_check_parallel.py
18	README.md	0	Zero required dependencies -- every requirement in the wheel is an optional extra	REPRODUCIBLE	curl -s https://pypi.org/pypi/inspeximus/json
19	claude-code.html	68	The MCP server exposes 68 tools	REPRODUCIBLE	python claims_audit.py --numbers
20	compare.html	0	The control: with our own guard off, we keep the correction 0% of the time	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
21	compare.html	100	The objection names our own headline number	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
22	compare.html	0 86.7 13.3 95 3.3 26.7	Graphiti keeps the correction 86.7%; resurrection 13.3%, 95% CI [3.3, 26.7]	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
23	compare.html	100	The control restated: guard off resurrects 100% of the time	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
24	compare.html	53.3 46.7 95 30.0 63.3	mem0 2.0.11 keeps the correction 53.3%; resurrection 46.7%, 95% CI [30.0, 63.3]	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
25	compare.html	30	Trials per system	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
26	compare.html	100 0	inspeximus keeps the correction 100% of the time and resurrects the old value 0%	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
27	compare.html	30	Sample size, stated in the page's own structured data so the two cannot drift	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
28	compare.html	2026	The competitor version and month measured, in the structured data	REPRODUCIBLE	curl -s https://pypi.org/pypi/mem0ai/json
29	compare.html	2026	The month the competitor figure was measured, said in prose next to the claim	REPRODUCIBLE	curl -s https://pypi.org/pypi/mem0ai/json
30	compare.html	30 0	Why intervals are shown; and our 0% with the guard on	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
31	docs/DEEP_DIVE.md	13 0 5	The example claims_audit run: 13 checks pass, 5 are not testable from this package	REPRODUCIBLE	python claims_audit.py --local
32	docs/DEEP_DIVE.md	8	The bedrock synthesis was checked from ~8 directions	EXTERNAL	—
33	docs/DEEP_DIVE.md	15 18 60 2 9 1 0 8 4	regex_extractor chain binding on benchmarks/chain_binding/ (15 chains, 18 unrelated pairs, 60 prose sentences): chains collapsing to one record 2/15 -> 9/15; false binds on unrelated pairs 1/18 -> 0/18; non-declarative prose keyed 8/60 -> 4/60	REPRODUCIBLE	python benchmarks/chain_binding/probe.py
34	docs/DEEP_DIVE.md	0.36	Per-memory outcome attribution reaches only ~0.36 power at n-of-1	EXTERNAL	—
35	docs/DEEP_DIVE.md	66.9 71.2	mem0 and Zep's self-reported LLM-judged QA scores	EXTERNAL	—
36	docs/DEEP_DIVE.md	4	...and by ~4x at one-eighth budget	EXTERNAL	—
37	docs/DEEP_DIVE.md	1.8	Value-ranked consolidation beats FIFO by ~1.8x at half budget	EXTERNAL	—
38	docs/DEEP_DIVE.md	0.17 1.00	A soft delete leaves the value recoverable in 5 of 6 stores (0.17); a wired hard delete scores 1.00	REPRODUCIBLE	python probes/forget_verification_bench.py
39	docs/DEEP_DIVE.md	1.00 0.17	Same six-store fan-out measurement, restated in the four-operations table	REPRODUCIBLE	python probes/forget_verification_bench.py
40	docs/DEEP_DIVE.md	0.0000	Run-to-run determinism at a fixed instant: arm (a) divergence 0.0000 on every corpus	REPRODUCIBLE-WITH-DEPS	python probes/reinforce_accuracy_ablation.py
41	docs/DEEP_DIVE.md	20	Pruning hub notes lifts lexical recall ~20% on a link-spammed store only	EXTERNAL	—
42	docs/DEEP_DIVE.md	0.00 0.05	In-repo cross-system echo cell: resurrection rate inspeximus 0.00, mem0 0.05, Graphiti 0.00	REPRODUCIBLE-WITH-DEPS	python probes/integrity_bench_echo.py --systems inspeximus
43	docs/DEEP_DIVE.md	5 0.94 0.25	Lexical recall@5 decays 0.94 -> 0.25 as the store grows	EXTERNAL	—
44	docs/DEEP_DIVE.md	2026 0.78 0.65	The superseded pair, quoted inside the note that discharges its caveat	REPRODUCIBLE-WITH-DEPS	python benchmarks/locomo/run.py --subset full --retrieval-only
45	docs/DEEP_DIVE.md	0.83 0.70	The copy-paste command with its expected output inline	REPRODUCIBLE-WITH-DEPS	python benchmarks/locomo/run.py --subset full --retrieval-only
46	docs/DEEP_DIVE.md	0.19 0.29	A withdrawn 0.19->0.29 delta, cited as an example of a confound we found and corrected	EXTERNAL	—
47	docs/DEEP_DIVE.md	1536	The LOCOMO question denominator behind the retrieval pair	REPRODUCIBLE-WITH-DEPS	python benchmarks/locomo/run.py --subset full --retrieval-only
48	docs/DEEP_DIVE.md	25 0.83 0.70	LOCOMO retrieval-recall@25 = 0.83 (any evidence turn) / 0.70 (all), n=1536, reinforce=False	REPRODUCIBLE-WITH-DEPS	python benchmarks/locomo/run.py --subset full --retrieval-only
49	docs/DEEP_DIVE.md	1536,	The LoCoMo config size behind recall_any@1	PENDING-HARNESS	python probes/retrieval_recall_locomo.py --k 25
50	docs/DEEP_DIVE.md	0.7839 0.6484 0.783 0.648 1536	The OLD published pair reproduces exactly at its own operating point (reinforce=True)	REPRODUCIBLE-WITH-DEPS	python benchmarks/locomo/run.py --subset full --retrieval-only
51	docs/DEEP_DIVE.md	68	The MCP server exposes 68 tools	REPRODUCIBLE	python -c "import re,pathlib;print(len(re.findall(chr(64)+chr(109)+chr(99)+chr(112)+chr(46)+'tool', pathlib.Path('inspeximus/mcp_server.py').read_text(encoding='utf-8'))))"
52	docs/DEEP_DIVE.md	0.592 0.544 2	MemOps answer accuracy: keep-all 0.592, mem0 0.544; ~2% of mem0 extractions failed to parse	EXTERNAL	—
53	docs/DEEP_DIVE.md	0.593	MemOps answer accuracy: inspeximus 0.593	EXTERNAL	—
54	docs/DEEP_DIVE.md	519 917 606 24	mem0's default pipeline spends 519-917 s (median 606) of LLM extraction per MemOps scenario	EXTERNAL	—
55	docs/DEEP_DIVE.md	2	~2% of mem0's MemOps extraction calls failed to parse	EXTERNAL	—
56	docs/DEEP_DIVE.md	24 50	MemOps: 24 long-context scenarios, ~50 sessions each	EXTERNAL	—
57	docs/DEEP_DIVE.md	42	Operating-point trap: a cosine top-1 store scores 42%	REPRODUCIBLE-WITH-DEPS	python probes/operating_point_memory.py
58	docs/DEEP_DIVE.md	100	The layered store scores 100% across all three operating points	REPRODUCIBLE-WITH-DEPS	python probes/operating_point_memory.py
59	docs/DEEP_DIVE.md	0 8	...and 0/8 on poison	REPRODUCIBLE-WITH-DEPS	python probes/operating_point_memory.py
60	docs/DEEP_DIVE.md	0 8 67	...0/8 on updated facts; a recency store scores 67%	REPRODUCIBLE-WITH-DEPS	python probes/operating_point_memory.py
61	docs/DEEP_DIVE.md	5 0.86 0.20	On paraphrase queries semantic recall@5 is 0.86 vs 0.20 lexical	EXTERNAL	—
62	docs/DEEP_DIVE.md	0.00 0.57 1.00	RAMR ECHO-RESISTANCE: keyed-without-guard 0.00, add-based 0.57, echo_guard 1.00	EXTERNAL	—
63	docs/DEEP_DIVE.md	0.397	recall_any@1 = 0.397 with nomic task prefixes on one LoCoMo config	PENDING-HARNESS	python probes/retrieval_recall_locomo.py --k 1
64	docs/DEEP_DIVE.md	30 2.8	At a 30% keep-budget, access-decay retains 2.8% of high-value/low-frequency memories	EXTERNAL	—
65	docs/DEEP_DIVE.md	3 2.2 7	~3x more value kept, persisting at ~2.2x even at a 7% budget	EXTERNAL	—
66	docs/DEEP_DIVE.md	20 100 64	...20% of total value, vs 100% and 64% for the value-aware blend	EXTERNAL	—
67	docs/DEEP_DIVE.md	0.65 2.6	Semantic recall@5 holds ~0.65 at full scale, ~2.6x lexical	EXTERNAL	—
68	docs/DEEP_DIVE.md	0.2213	NEGATIVE CONTROL: with the salience bar removed, rejection collapses to 0.2213	REPRODUCIBLE	python probes/session_digest_multisession.py
69	docs/DEEP_DIVE.md	7	close_session costs 7 ms on the 2,606-record fixture	REPRODUCIBLE	python probes/session_digest_multisession.py
70	docs/DEEP_DIVE.md	2,606 1.000	SessionEnd digest -> SessionStart injection, 8-session / 2,606-record fixture: injection recall 1.000 of a session's conclusions reach the next session	REPRODUCIBLE	python probes/session_digest_multisession.py
71	docs/DEEP_DIVE.md	1.0000	Below-threshold rejection 1.0000 on the same fixture	REPRODUCIBLE	python probes/session_digest_multisession.py
72	docs/DEEP_DIVE.md	0 8	WITHDRAWN: 'severe-test 8/8' -- the probe reports 0/24 and nothing here produces an 8/8	WITHDRAWN	python probes/supersession_replication.py
73	docs/DEEP_DIVE.md	0.61	A cosine classifier separating a contradiction from a rephrase scores AUROC ~0.61	REPRODUCIBLE-WITH-DEPS	python probes/supersession_replication.py
74	docs/DEEP_DIVE.md	0.613 41.7 0.0	The 2026-08-01 re-run of that probe, quoted with its date	REPRODUCIBLE-WITH-DEPS	python probes/supersession_replication.py
75	docs/DEEP_DIVE.md	42	A similarity-based store serves the stale value ~42% of the time	REPRODUCIBLE-WITH-DEPS	python probes/supersession_replication.py
76	docs/DEEP_DIVE.md	0	The deterministic SRO key drives the stale-value rate to 0%	REPRODUCIBLE-WITH-DEPS	python probes/supersession_replication.py
77	docs/DEEP_DIVE.md	0.9 10	Content-declared corroboration falls to a sybil at ~0.9 attack-success across 10 models	REPRODUCIBLE-WITH-DEPS	python probes/memory_defense_layer_probe.py
78	docs/DEEP_DIVE.md	0 80	GAP CONTROL: 0 of 80 top-1 answers move when the two reads are not separated at all	REPRODUCIBLE-WITH-DEPS	python probes/recall_over_a_time_gap.py
79	docs/DEEP_DIVE.md	80	The time-gap measurement runs on four LOCOMO conversations, 80 questions sampled from each	REPRODUCIBLE-WITH-DEPS	python probes/recall_over_a_time_gap.py
80	docs/DEEP_DIVE.md	64 83 320	Reading the same untouched store twice ~2s apart moves 64-83 of 320 top-1 answers (four LOCOMO conversations, 80 questions each, reinforce=False)	REPRODUCIBLE-WITH-DEPS	python probes/recall_over_a_time_gap.py
81	docs/DEEP_DIVE.md	0.0094	Across five randomised insertion orders the hit@1 change over the gap runs +0.0094 to -0.0219	REPRODUCIBLE-WITH-DEPS	python probes/recall_over_a_time_gap.py
82	docs/DEEP_DIVE.md	16 80	The effect saturates: 16 of 80 move at a two-second gap and the same count at ten	REPRODUCIBLE-WITH-DEPS	python probes/recall_over_a_time_gap.py
83	docs/DEEP_DIVE.md	-0.0219 2 5 -0.0062	The hit@1 change is negative in only 2 of 5 randomised insert orders; natural conversation order alone reads -0.0062	REPRODUCIBLE-WITH-DEPS	python probes/recall_over_a_time_gap.py
84	docs/DEEP_DIVE.md	0.847	Every top-1 answer that moved across the gap moved between records reported at the same score, e.g. 0.847 against 0.847	REPRODUCIBLE-WITH-DEPS	python probes/recall_over_a_time_gap.py
85	docs/DEEP_DIVE.md	100 6	100% of the moved answers stayed inside a displayed tie, across all six insert orders	REPRODUCIBLE-WITH-DEPS	python probes/recall_over_a_time_gap.py
86	docs/DEEP_DIVE.md	10,000	Contradiction detection runs in production over the ~10,000-note vault	EXTERNAL	—
87	docs/DEEP_DIVE.md	10,000	inspeximus has run daily over a ~10,000-note vault	EXTERNAL	—
88	index.html	9 0	Homepage counter: 9 framework adapters	REPRODUCIBLE	python -c "import pathlib;print(sorted(p.stem for p in pathlib.Path('inspeximus/integrations').glob('*.py')))"
89	index.html	0.00	Benchmark bar: Graphiti 0.00	REPRODUCIBLE-WITH-DEPS	python probes/integrity_bench_revert.py --systems inspeximus,graphiti --n 20
90	index.html	0.75	Benchmark bar: inspeximus 0.75	REPRODUCIBLE-WITH-DEPS	python probes/integrity_bench_revert.py --systems inspeximus --n 20
91	index.html	0.20	Benchmark bar: mem0 0.20	REPRODUCIBLE-WITH-DEPS	python probes/integrity_bench_revert.py --systems inspeximus,mem0 --n 20
92	index.html	100	The control: with our guard off we resurrect every time, so the number is the mechanism	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
93	index.html	30	Sample size for the native-config echo run	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
94	index.html	0 13.3 46.7	Corrected-fact resurrection per system on their native configs	REPRODUCIBLE-WITH-DEPS	python ramr_echo_resistance_backends.py # RAMR repo
95	index.html	10 13 3	10 of 13 framework adapters verified against current upstream; 3 recorded broken	REPRODUCIBLE	python tools/integration_conformance.py
96	index.html	68 0	Homepage counter: 68 MCP tools	REPRODUCIBLE	python claims_audit.py --numbers
97	index.html	68	Homepage heading: 68 MCP tools	REPRODUCIBLE	python claims_audit.py --numbers
98	index.html	2.0.11 2026	The exact competitor version and date measured, stated rather than implied as current	REPRODUCIBLE	curl -s https://pypi.org/pypi/mem0ai/json
99	index.html	0.75 0.20 0.00 20 95	Cross-system revert success over n=20: inspeximus 0.75, mem0 0.20, Graphiti 0.00	REPRODUCIBLE-WITH-DEPS	python probes/integrity_bench_revert.py --systems inspeximus --n 20
100	index.html	0.75 0.20 20 0	Homepage counter restating the revert cell	REPRODUCIBLE-WITH-DEPS	python probes/integrity_bench_revert.py --systems inspeximus --n 20
101	index.html	98.3	Our own store: fraction of records carrying a source field	REPRODUCIBLE-WITH-DEPS	curl -sO https://raw.githubusercontent.com/DanceNitra/agora/main/research/probes/can_we_reconcile_our_own_index.py && python can_we_reconcile_our_own_index.py
102	index.html	0.01	Our own store: fraction whose source actually resolves	REPRODUCIBLE-WITH-DEPS	curl -sO https://raw.githubusercontent.com/DanceNitra/agora/main/research/probes/can_we_reconcile_our_own_index.py && python can_we_reconcile_our_own_index.py
103	index.html	0	Homepage counter: 0 runtime dependencies	REPRODUCIBLE	python claims_audit.py --local

Notes

• mcp-tool-count — Published as 30 until 2026-08-01 -- 26 short -- while the homepage said 15 in one place and 56 in another. Three surfaces, one server, no error anywhere. Now read from the code.
• readme-adapter-conformance — Read from docs/integration_conformance.json by _live_consistency(), which now checks BOTH index.html and README.md -- a second copy of a number is a second place for it to go stale. The other 12 in this file is the EU AI Act article number and stays a declared non-claim; the counts moved 9/12 -> 10/13 when the llm-errata adapter landed, and this line is why the drift surfaced instead of shipping; COUNT-DRIFT caught the collision the moment this line was added, which is the whole point.
• readme-echo-graphiti — Measured on the vendor's own native config (Neo4j + OpenAI), n=30.
• readme-echo-mem0 — Version-stamped on purpose: mem0 is on 2.0.18 as of 2026-08-11 and we have NOT re-run it.
• readme-mcp-tool-count — Checked against the live @mcp.tool() count by _live_consistency(), not read from here.
• readme-own-source-coverage — Published as our own failure, not a product claim. 210,499 records across ten stores.
• cc-tool-count — Checked against the live @mcp.tool() count by _live_consistency(), not read from here.
• cmp-graphiti — The bare 0 on this line is the MAJOR VERSION in "Graphiti 0.x", not a measurement. It is listed rather than excused, because declaring "0" a non-claim file-wide would also excuse the two real zeros this page publishes.
• cmp-mem0 — Version-stamped deliberately: mem0 is on 2.0.18 and we have NOT re-run it.
• readme-audit-summary — Self-referential, so it is checked against len(CHECKS) and len(NOT_TESTABLE_HERE) rather than trusted. The block used to name inspeximus-1.24.1 while the package was at 1.89.0; the version line was dropped rather than pinned, because it would go stale on every release.
• readme-bedrock-directions — A count of the analytical directions taken, not a measurement. Left in because the sentence labels itself 'a synthesis over those cases, not a proof'.
• readme-chain-binding — The 'before' column is measured against git show main:inspeximus/core.py on the same fixture, not quoted from elsewhere. The false-bind row is the control: a keyer that binds everything scores a perfect 15/15 while tripping all 18 negative pairs, which is why the bind rate alone is not evidence.
• readme-competitor-judges — Other projects' published numbers, cited as not comparable across harnesses -- which is the point the sentence makes.
• readme-erasure-fanout-table — Replaced a 'measured 15/15 on a verified-forgetting severe-test' for which no artifact in this repository produces a 15/15 of anything. The bench that DOES exist scores 0.17 / 1.00 over six stores, so the sentence now cites the number the committed code prints.
• readme-integrity-echo-cell — The inspeximus column runs locally and free; the mem0/Graphiti columns need OPENAI_API_KEY and a live neo4j, which is why this is WITH-DEPS rather than REPRODUCIBLE.
• readme-lexical-decay — Agora Lab b4c260, cited in place. No probe in this repository reproduces it.
• readme-locomo-confound — Kept deliberately: it is a retraction, not a claim. Removing it would erase the correction.
• readme-locomo-headline — This row was PENDING-HARNESS for the whole of this audit, and it is the reason that status exists. The harness landed as benchmarks/locomo/ and the row moved WITHOUT the number having been re-asserted in the meantime. Verified here against the committed result benchmarks/locomo/results/full_retrieval.json rather than against the prose: recall_any 0.8262 / recall_all 0.6986 on the published 1536-question denominator, pinned at k=25, mode=hybrid, prefer=speaker, reinforce=false. WITH-DEPS because LOCOMO is not ours to redistribute -- the command needs locomo10.json downloaded (sha256 pinned in config.json), though the committed result is readable without it.
• readme-locomo-old-pair — The pair this audit opened on. It was never wrong -- it was measured with recall()'s reinforce=True default, which mutates value/last_access, so each benchmark query was answered by a store the previous queries had modified and the score depended on question order. Pinning reinforce=False makes the run deterministic and scores 4-5 points HIGHER. A number that moves when you fix the instrument is exactly what an unreproducible number hides.
• readme-mcp-tools — Checked against the live @mcp.tool() count by _live_consistency(), not by reading it here.
• readme-memops-cost — Agora MemOps harness; needs mem0 + an LLM budget, so it cannot ship here.
• readme-memops-scenarios — Harness lives in the Agora repo (agora_output/lab/memops), already linked in place.
• readme-operating-cosine — Needs a local nomic-embed-text (Ollama).
• readme-paraphrase — Agora Lab 3501f1.
• readme-ramr-echo — From RAMR, a separate repository. The README presented these as if produced here; it now says where they come from AND points at this repo's own echo cell, which measures a different quantity and does NOT flatter us.
• readme-recall-any1 — Same dataset blocker as the headline pair; already flagged in place as not reproducible here.
• readme-retention-cold — Agora Lab 19d802.
• readme-semantic-hold — Agora Lab b4c260.
• readme-session-digest-control — Registered deliberately rather than dropped: without it a rejection of 1.0000 cannot be told apart from a fixture that contained nothing to reject.
• readme-supersession-auroc — Re-run 2026-08-01: AUROC 0.613. Needs a local nomic-embed-text (Ollama) and numpy.
• readme-supersession-stale — Re-run 2026-08-01: 41.7%.
• readme-sybil-attack — The harness is committed; reproducing the number needs ten models and a judge, which no checkout can ship.
• readme-time-gap-cause — Registered deliberately rather than dropped: without a zero-gap arm, 'the ranking depends on when you ask' is only an observation about two reads and names no cause.
• readme-time-gap-spread-low — The natural-order figure is registered beside the spread on purpose: alone it reproduces to four decimals every run and reads as a systematic loss, which is the fixture (LOCOMO gold turns skew late, so gold records are newer) and not a property of the store.
• readme-vault-contradictions — Same private deployment as the hero line.
• readme-vault-hero — Our own private Obsidian vault. There is no command; the text now says so instead of implying the reader could check it.
• site-adapters — Was 6 while the README said nine and the package ships nine agent-framework adapters (autogen, crewai, google_adk, haystack, langchain, langgraph, llamaindex, openai_agents, pydantic_ai).
• site-echo-row — Replaced a STALE caveat claiming all three tie on this cell; the later run separates them.
• site-integration-conformance — Read from the committed ledger docs/integration_conformance.json by _live_consistency(), not typed. The page previously said 'Drop-in for' all nine frameworks with no qualifier at all, while crewai 1.15.6, openai-agents 0.18.3 and langgraph-checkpointer 1.2.9 were recorded broken -- an unqualified capability claim contradicted by a JSON file in the same repo.
• site-mcp-tools-counter — Was 15. The counter renders data-count, so the figure a reader sees lives in an attribute -- which is why the scanner hoists data-count out of the tag before stripping tags.
• site-revert-bench — The inspeximus column runs locally; mem0 needs OPENAI_API_KEY and Graphiti a live neo4j. Methodology and CIs: probes/INTEGRITY_BENCHMARK.md.
• site-zero-deps — This is the c_zero_deps check, which reads installed METADATA or, failing that, the declared pyproject dependencies -- and hard-fails if it can read neither.

Known unenforced numbers

These are outside the 6 token-enforced files, so the guard above does not cover them. They are listed because "absent from the table" and "not a problem" are different statements, and here only the first one is true. None may be promoted onto the reader-facing surface while it still says PENDING-HARNESS.

where	figure	status	why it is here
inspeximus/core.py — recall_iterative docstring	0.057 -> 0.186 (3.3x)	PENDING-HARNESS	Quoted with no scope. That ratio is n=70 over the THREE HARDEST LoCoMo conversations; across the full benchmark it is 0.145 -> 0.297 (2.05x), n=276, all ten conversations. A flattering subset ratio published without the subset is the exact defect this audit exists to find, and neither figure reproduces from a clean checkout (same LOCOMO dataset blocker as readme-locomo-headline). Scope it to the subset AND give the full-benchmark pair, or drop it, once a harness lands.
docs / docstrings — the recall tie policy	n/a (a stated policy, not a figure)	PENDING-HARNESS	'equal relevance => newest first' is measured to FAIL in hybrid and auto modes: RRF gives equivalent records distinct fused scores, so the tie-break never fires and top-1 is the OLDEST. Anywhere the policy is asserted unconditionally it needs scoping to the modes where it holds.
bench/ — MemoryAgentBench Conflict Resolution	any CR score	PENDING-HARNESS	Our own gate killed the 'supersession wins CR' headline: a naive keep-all store ties us (97 vs 96, and 85 vs 87 on the faithful re-run), and a 6k-vs-32k context discrepancy in bench/README.md is still open. No CR number may go onto the reader-facing surface until that is resolved.
index.html / MCP_LISTINGS.md / README.md — the MCP tool count	60	REPRODUCIBLE	No longer hypothetical. Between this audit and its rebase the server grew from 56 to 58 tools; a sibling corrected ONE of the four places that publish the count (the homepage heading) and left the homepage counter, both MCP_LISTINGS figures and the README at 56. This audit named all four on its first run after the rebase. Checked against the live @mcp.tool() count, never typed.
bench/README.md — its own committed JSON	9 of 12 cells	PENDING-HARNESS	Reported to disagree with the JSON it is generated from. Outside this audit's token-enforced scope (README.md, MCP_LISTINGS.md, index.html), and left to the unit that owns the reconciliation rather than guessed at from here.
docs / homepage — LongMemEval end-to-end	0.45, vs a 0.50 oracle ceiling and a 0.05 no-memory floor, n=20	PENDING-HARNESS	Landed as a pilot. No figure from it has reached README.md, MCP_LISTINGS.md or index.html, and none should until it carries its scope -- n=20, a pilot, and a band check that exits 5 by design.

What was withdrawn, and why that is the point

Removing a number we cannot back is a win. Every WITHDRAWN row above is a figure this audit deleted from the reader-facing surface rather than dress up: either no artifact in this repository produces it, or the artifact it named does not exist. Showing the gap is what makes the rest worth reading.