The claims ledger
Every factual assertion rendered on this site has a row here naming the file it was checked against and how strong that evidence is. Written by Shane Pilon, and generated from claims.md in the site repository at build time, so this page cannot drift from the file it publishes.
What the check does
A verifier resolves every cited source and fails if one is missing, if a row is malformed, if the numbering has a gap, or if a result tries to ship on an evidence status the ledger forbids, which is a single-seed result or a proxy metric. It then writes a lock file recording each row as worded together with a hash of the file it cites. The site build re-derives the rows and refuses to compile if a row has no lock entry or has drifted from the wording it was locked in. 68 rows are locked as of 2026-08-08.
What the check does not do
- It never opens a source to judge whether that source actually supports the claim as worded. It checks that the file is there and that its contents have not changed since the row was locked. A row pointed at a real file that says something else would pass.
- It never reads the rendered page. If copy on the site drifts away from the row that licenses it, the check still passes. That class of drift is caught by reading, not by the build.
- Most sources are private research files and are not published here. What is published is which file, of what kind, and how strong the evidence is called. The external rows are the exception: they cite external-sources.md, a file in this site’s repository, and every one of its entries is printed in full further down this page so you can check them without taking my word for anything.
Read the check itself, all 92 lines
This is scripts/verify-claims-lock.mjs in full, read off disk when this page was built. It runs in prebuild; if it exits non-zero this page does not deploy. It is the half of the gate that runs on a builder that has never seen the private research files, which is why it checks the lock rather than the sources: it proves that every row shipping today is a row that was checked against a real source, in the exact wording it was checked in.
#!/usr/bin/env node
// The half of the gate that runs everywhere, including on a Vercel builder that
// has never seen the palace.
//
// scripts/verify-claims.mjs resolves every cited source against the palace and
// writes claims.lock.json. That file cannot run on Vercel: the palace is a local
// Windows path. So the published sentence "the build fails if any of them loses
// that source" was false of the shipped build, because `build` was a bare
// `next build` with no hook, no vercel.json and no CI.
//
// This script closes that. It re-derives the active rows from claims.md and
// fails when:
// - claims.lock.json is missing,
// - an active row has no lock entry,
// - a row's wording, source path or status has drifted from what was locked,
// - the lock carries a row that claims.md no longer has.
//
// It does not, and cannot, prove that a palace file still exists on this
// machine. It proves that every row shipping today is a row that was checked
// against a real source, in the exact wording it was checked in. Re-run
// `npm run verify:claims` on the box with the palace to re-lock.
import { existsSync, readFileSync } from "node:fs";
import { join, resolve } from "node:path";
import { parseClaims, claimFingerprint } from "./claims-parse.mjs";
const LOCK_PATH = join(resolve(process.cwd()), "claims.lock.json");
if (!existsSync(LOCK_PATH)) {
console.error("FAIL: claims.lock.json is missing. Run `npm run verify:claims` with the palace mounted.");
process.exit(1);
}
let lock;
try {
lock = JSON.parse(readFileSync(LOCK_PATH, "utf8"));
} catch (err) {
console.error(`FAIL: claims.lock.json is not readable JSON — ${err.message}`);
process.exit(1);
}
const { rows, tableRowCount } = parseClaims("claims.md");
if (rows.length === 0) {
console.error("FAIL: no claim rows parsed from claims.md");
process.exit(1);
}
let bad = 0;
if (tableRowCount !== rows.length) {
console.error(
`FAIL: ${tableRowCount} table row(s) found in the active section but only ${rows.length} parsed — a row is malformed`
);
bad++;
}
const byNumber = new Map((lock.claims ?? []).map((c) => [String(c.n), c]));
for (const row of rows) {
const entry = byNumber.get(String(row.num));
if (!entry) {
console.error(
`FAIL claim ${row.num}: no lock entry. An unlocked claim does not ship; run \`npm run verify:claims\` with the palace mounted.`
);
bad++;
continue;
}
byNumber.delete(String(row.num));
if (entry.claim !== claimFingerprint(row)) {
console.error(
`FAIL claim ${row.num}: wording, source or status has changed since it was locked against ${entry.source}. Re-verify it against its source, then re-lock.`
);
bad++;
}
}
for (const [n, entry] of byNumber) {
console.error(
`FAIL: claims.lock.json still carries claim ${n} (${entry.source}) but claims.md no longer does — the lock is stale.`
);
bad++;
}
if (bad > 0) {
console.error(`\n${bad} problem(s). The build refuses an unlocked or drifted claim.`);
process.exit(1);
}
console.log(
`OK: ${rows.length} claims, every one locked to a verified source${
lock.generatedAt ? ` (locked ${lock.generatedAt})` : ""
}.`
);The external sources, in full
Three rows on this site cite something that is not mine: two reported external events, and one general property of the published literature. All three have public primary sources, and none of them is one of my own measurements.
E1 · June 2026
VentureBeat VB Pulse survey of 157 enterprises of 100 or more employees.
- roughly half of respondents deployed an AI agent or LLM feature that passed internal evaluations and still caused a customer-facing failure
- one in four experienced that failure more than once
- only 5% say they fully trust the automated evaluations informing release decisions
- 66% permit some production deployment without human review or are building toward it within 12 months
venturebeat.com, the primary source
n=157 is the eval-gap sub-survey of a five-survey VB Pulse programme totalling 573 respondents, so 157 is the number that travels with these figures. A circulating 29% figure on real-world alignment is not confirmed at the reachable primary source and is deliberately not used anywhere on this site.
E2 · Published 2026-02-23
OpenAI stops reporting SWE-bench Verified and recommends other model developers do the same.
- The benchmark is SWE-bench Verified, a human-validated 500-problem subset of SWE-bench released by OpenAI in August 2024.
- OpenAI audited 138 of those 500 problems: the ones its own o3 did not solve consistently across 64 independent runs. Each case was reviewed by at least six experienced software engineers.
- 59.4% of the 138 were found to have material issues in the test design and/or the problem description. The breakdown is 35.5% narrow test cases (tests enforcing specific implementation details, so functionally correct submissions fail) and 18.8% wide test cases (tests checking functionality the problem description never specified).
openai.com, the primary source
The denominator and the selection criterion travel with the percentage, because 59.4% of a failure-selected 138 is a different statement from 59.4% of the benchmark. This row used to cite two news aggregators and record the weaker figure at least 59.4%; the primary post states 59.4%, not a floor, and the aggregators are now listed as corroboration only.
E3 · Published literature, 2020 and 2025
Knowledge distillation: a student smaller than its teacher commonly trails it, a shortfall the literature calls the capacity gap.
- Knowledge distillation trains a smaller student model from a larger teacher, and the student commonly ends up performing below the teacher it was distilled from.
- The literature attributes that shortfall to the difference in capacity between the two models and names the phenomenon the capacity gap.
- A larger or stronger teacher does not necessarily produce a better student: student performance can drop when distilling from an oversized teacher.
arxiv.org/abs/2506.18244, the primary source
This is a general, well-established property of distillation, not a measurement of anything of mine, and nothing here has ever seen my student or my teacher. It is a frame for reading the direction of a result and may never be used to explain or excuse the 51.4 against 71.8 pair, nor to say anything about where that pair goes next. Students can sometimes match or beat their teachers, so the wording stays usually or commonly and never always. A second arXiv paper, Reducing the Teacher-Student Gap via Spherical Knowledge Distillation (2010.07485), states the same capacity-gap property and is recorded alongside this one. A third item, the OpenReview paper Can Students Outperform Teachers in Knowledge Distillation based Model Compression, is the standing statement of the counter-case; it is listed in the record but its page answered a bot-verification screen, which was not bypassed, so it was never read at source and nothing here rests on it.
The claims, as worded on the site
68 rows
01
A red-team pass caught six fake breakthroughs in one day
record · Maestro_Memory_Palace_July2026/WINS_DIGEST.md
02
One argument, strict=False, let the loader accept a checkpoint even where it did not match, and 112 trained keys were dropped in silence
record · Maestro_Memory_Palace_July2026/WINS_DIGEST.md
03
Every routing evaluation had been scoring a randomly routed model: the weights were trained, the routing they were supposed to use never loaded
record · Maestro_Memory_Palace_July2026/state/CONFIRMATIONS.md
04
The run was scoring preferred answers against rejected ones, and it learned to put its margin on the end-of-sequence token instead of on the answer. The headline number improved. The model did not
record · Maestro_Memory_Palace_July2026/state/HANDOFF_LATEST.md
05
Only the first-token preference telemetry and a per-source autopsy told the two apart
record · Maestro_Memory_Palace_July2026/state/HANDOFF_LATEST.md
06
In November 2025 I ran a source-grounded re-audit of a system I had built and was pitching: 16 questions, my own harness, 28.1% verified accuracy, 8 of the 16 scoring zero, and fabrications delivered with full confidence
record · Maestro_Memory_Palace_July2026/WINS_DIGEST.md
07
I wrote that no amount of re-running would fix it, and stopped trying to fix it from the outside
record · Maestro_Memory_Palace_July2026/WINS_DIGEST.md
08
Zero fabricated memories across roughly 70 banked training runs on my own memory benchmark; one later arm broke that zero on a second seed and is held provisional, not banked
measured · Maestro_Memory_Palace_July2026/WINS_DIGEST.md
09
Under my evaluation harness the recall path is hard-masked, so on that path unstored content is unavailable rather than discouraged: the mask reads a flag the harness arms using oracle knowledge of what was stored, and the deployed chat path never arms it, so fabrication on free-form input is unsolved
documented · Maestro_Memory_Palace_July2026/lessons/jun22_fp0_armed_by_harness_not_deploy.md
10
The prediction and the result that would refute it are both written down before the run starts
documented · Maestro_Memory_Palace_July2026/METHOD.md
11
Before a number counts, I check that the harness can return an answer I already know to be correct
documented · Maestro_Memory_Palace_July2026/METHOD.md
12
Every result gets a pass whose only job is to refute it. What survives gets recorded. What does not survive gets recorded too
documented · Maestro_Memory_Palace_July2026/METHOD.md
13
Masked means a configuration in which the evaluation harness arms the recall mask, which it can only do because it knows what was stored: the masked configuration scored 57.1, unmasked scored 57.1, identical seed-for-seed across two seeds; both numbers come from my own multi-turn memory benchmark, an internal measure with no external baseline; on that benchmark, making honesty structural cost nothing
two-seeded · Maestro_Memory_Palace_July2026/WINS_DIGEST.md
14
The instrument is a 21-question set, which is thin, and I have said so in my own notes
documented · Maestro_Memory_Palace_July2026/reference/shane_hiring_panel_report.md
15
Maestro overview: an explainer for the research that gives the method and the results and withholds the recipe, one self-contained HTML page on Vercel
documented · docs/portraits/DEPLOY.md
16
One is a portrait narrated across 62 audio segments, each anchored to the paragraph it describes
documented · docs/portraits/audio/manifest.json
17
The other pairs a walkthrough player to the page: every cue carries a text hint, so when the copy changes the player drops the highlight instead of highlighting the wrong line
documented · docs/portraits/DEPLOY.md
18
Static HTML plus audio, on Vercel. The portrait is synthesized narration; the walkthrough is my own recorded voice, transcribed and aligned to the page
documented · docs/portraits/DEPLOY.md
19
Three pricing and ROI calculators built for two operating businesses. Each is one self-contained HTML file with no build step
documented · docs/portraits/DEPLOY.md
20
Two of them autosave scenarios to the browser, and to Postgres when the tool is embedded in an admin dashboard, with the database assigning the owner so a saved scenario cannot be claimed by the wrong account
documented · docs/portraits/DEPLOY.md
21
The product catalog behind one of them was extracted from the supplier's PDF rather than retyped, and the two places where that PDF is genuinely broken are surfaced in the interface as warnings rather than guessed at
documented · docs/portraits/DEPLOY.md
22
The other five are live but unlisted, and three are internal tools for businesses I have a stake in. Available on request, except the one that exposes a supplier's cost sheet
documented · docs/portraits/DEPLOY.md
23
One 96GB GPU. Lit 2026-07-05 at 17:55:36 UTC on a budget of 8 billion tokens, finishing in August
record · Maestro_Memory_Palace_July2026/state/current.md
24
One evaluation instrument crashed along the way; the training loop did not
record · Maestro_Memory_Palace_July2026/state/current.md
25
The rig needed two servers only because of one subsystem. I cut that subsystem on June 27th, 2026, and the whole thing has run on one card since
record · Maestro_Memory_Palace_July2026/decisions/jun27_cut_spectral_decision.md
26
Startup checks that refuse to launch on bad disk or a busy GPU, rolling checkpoints, small evaluations that fire every few thousand steps, and an escalation path that raises a flag for review instead of killing the run
documented · Maestro_Memory_Palace_July2026/state/current.md
27
Read a 75,000-token document in four to five seconds, and 1.2 million tokens in about twenty minutes. Both figures are my own measurements on my own documents. Neither is a public benchmark
self-reported · Maestro_Memory_Palace_July2026/WINS_DIGEST.md
28
A hand-built log file, vision file and goal file, fed to a model through a context window a small fraction of today's size. There was no compaction and no long-context option to fall back on
self-reported · Maestro_Memory_Palace_July2026/history/pre_palace_and_paper_era.md
29
I did not write code until 2025
self-reported · Maestro_Memory_Palace_July2026/history/pre_palace_and_paper_era.md
30
Five months in, in August 2025, I filed a provisional patent with 17 claims; provisionals last twelve months, and I am not claiming anything is in force today
record · Maestro_Memory_Palace_July2026/history/pre_palace_and_paper_era.md
31
The work listed above was built since, on one GPU, with no employees and no co-founders. There is no firm behind this
documented · Maestro_Memory_Palace_July2026/reference/shane_hiring_panel_report.md
32
From Pennsylvania
record · Maestro_Memory_Palace_July2026/claude_memory_backup/shane-and-maestro-project.md
33
Maestro is a 2.5-billion-parameter language model built to remember
record · Maestro_Memory_Palace_July2026/state/current.md
34
The memory mechanism is trained into the model rather than bolted around a finished one at query time
documented · Maestro_Memory_Palace_July2026/WINS_DIGEST.md
35
What it remembers lives in a cache the model was trained to read, not in its weights
documented · Maestro_Memory_Palace_July2026/GOAL.md
36
It holds what you told it earlier in a conversation seven times in ten: multi-turn recall 70% on a 40-question set, with every control reading zero
measured · Maestro_Memory_Palace_July2026/state/current.md
37
Cross-session recall: 78.5% against 61.6% on shuffled queries, across 2,000 items. Two asterisks travel with it: a leak guard fired on the baseline arm, so the lift contrast is confounded; and the probe set uses world-answerable entities, which put a parametric floor of at least 2.5% under any cross-session number. I read this as indicative, not settled
measured · .superpowers/sdd/2026-07-31-pilonics-website/numbers-of-record.md
38
It deflects rather than invents when it does not have something
measured · Maestro_Memory_Palace_July2026/state/current.md
39
The exam below was scored on July 12th, 2026, at 26% of training, against bars I wrote down before the data existed. Every benchmark here is one I built and run myself. None of them are public leaderboards
record · Maestro_Memory_Palace_July2026/state/current.md
40
Fabrication on that exam: 0.0 on the multi-turn and cross-session gauges, and 0.4% cross-sibling on the control set, which my own threshold treats as sub-threshold rather than clean
measured · Maestro_Memory_Palace_July2026/LEDGER_POST_SPECTRAL.md
41
On 216 held-out reasoning questions scored on my own harness, the student reads 50.9 at 172k steps and 51.4 at 180k, plain forward with no cache, where the finished 4B teacher holds 71.8 both times; the teacher arm is bare-cloze prompted and likely understated, so the true gap is probably wider than 51.4 against 71.8 shows, not narrower, and the per-lane deltas are unquotable at n=24
measured · .superpowers/sdd/2026-07-31-pilonics-website/numbers-of-record.md
42
The student is 2.5B parameters, read mid-training at 180k steps; the teacher is a finished 4B model
record · .superpowers/sdd/2026-07-31-pilonics-website/numbers-of-record.md
43
The same 216 questions get re-scored at each checkpoint as training continues
record · .superpowers/sdd/2026-07-31-pilonics-website/numbers-of-record.md
44
I grade my own project by the same evidence tiers I would apply to yours. Every capability claim is tracked by that tier, and the record keeps the ones that had to be downgraded, not only the ones that held. Proxy metrics decide what I try next. They are not what I hand you as capability
record · Maestro_Memory_Palace_July2026/WINS_DIGEST.md
45
Architecture details stay unpublished. What is on offer is the method, and the results as they actually are
documented · docs/portraits/DEPLOY.md
46
Every harness I have torn down has been my own, in one problem class: memory recall and fact conditioning in a language model I trained myself. I have never audited a third party's evaluation
documented · Maestro_Memory_Palace_July2026/reference/capability_sweep_aug2026/sweep-evaluation.json
47
I have never worked inside RAGAS, promptfoo, DeepEval, LangSmith, Braintrust, OpenAI Evals or W&B. If your evals live in one of those, I will be reading it for the first time on your clock
documented · Maestro_Memory_Palace_July2026/reference/capability_sweep_aug2026/sweep-evaluation.json
48
This is single-machine work. No Spark, Ray, Beam, Dask, Airflow, dbt, Kafka or warehouse appears anywhere in what I have built. The largest artifacts are on the order of two gigabytes and tens of millions of tokens
documented · Maestro_Memory_Palace_July2026/reference/capability_sweep_aug2026/sweep-corpus.json
49
Deduplication is exact-match and cap-per-prototype, not fuzzy or near-duplicate matching at scale
documented · Maestro_Memory_Palace_July2026/reference/capability_sweep_aug2026/sweep-corpus.json
50
The next run after this one is already specified and priced at about forty dollars of compute, and it stays parked until the current one gives me a reason to spend it
documented · Maestro_Memory_Palace_July2026/WINS_DIGEST.md
51
For months, one capability measured zero, not low but zero, across every routing evaluation I ran
record · Maestro_Memory_Palace_July2026/WINS_DIGEST.md
52
The evaluation harness loaded checkpoints with strict=False, and it dropped 112 trained routing keys in silence
record · Maestro_Memory_Palace_July2026/WINS_DIGEST.md
53
Every one of those evaluations had been scoring a model whose routing was random: the weights were trained, and the routing they were supposed to use never loaded
record · Maestro_Memory_Palace_July2026/state/CONFIRMATIONS.md
54
The load returns the keys it could not place, and the load line prints how many were dropped against how many were loaded, which is where the bug was visible the whole time
documented · Maestro_Memory_Palace_July2026/state/CONFIRMATIONS.md
55
EXTERNAL (VentureBeat VB Pulse survey, June 2026, n=157 enterprises): in a June 2026 survey of 157 enterprises, roughly half had shipped an AI feature that passed internal evaluations and still failed in front of customers, one in four more than once, and only 5% fully trust their automated evaluations
record · external-sources.md
56
EXTERNAL (OpenAI, published 2026-02-23): in February 2026 OpenAI published that it had stopped reporting SWE-bench Verified and recommended other model developers stop too, after auditing 138 of the benchmark's 500 problems, the ones its own model had not solved consistently across 64 runs, and finding material issues in the test design or the problem description of 59.4% of them
record · external-sources.md
57
Two lines did it. One dropped the long examples. The other scored the loss on the response alone, so nothing ever penalised the competing fact already sitting in the context, and the model was never taught to use the one it had been handed. On 750 of 1,000 grounded answers it had been trained to make something up
record · Maestro_Memory_Palace_July2026/WINS_DIGEST.md
58
EXTERNAL (knowledge-distillation literature, arXiv 2506.18244 and 2010.07485): a distilled student smaller than its teacher usually trails it, and the literature attributes that to the capacity difference between the two and calls it the capacity gap
record · external-sources.md
59
In April 2026, roughly 1,150 steps into a training run of mine, a data-dependent Python branch inside a gradient-checkpointed forward took two different routes in a single step and stopped the run: 162 tensors saved during the original forward against 161 saved during the recomputation. Run-killing and non-destructive: the model state was safe, and what it took was the run rather than the weights
record · Maestro_Memory_Palace_July2026/lessons/checkpoint_forward_branch_bug.md
60
Gradient checkpointing discards the intermediate activations of a wrapped function and runs that function again during the backward pass to rebuild them, and the non-reentrant implementation, the one selected by use_reentrant=False, counts the tensors each pass saved and refuses to continue when the two counts disagree
documented · Maestro_Memory_Palace_July2026/lessons/checkpoint_forward_branch_bug.md
61
PyTorch publishes the list of operations that are normally nondeterministic, and what most of them have in common is atomic accumulation: the order the additions arrive in is not fixed from one launch to the next, and floating-point addition is not associative, so the same inputs can land on a value that differs in its last bits. torch.use_deterministic_algorithms(True) makes a listed operation either take a deterministic path or raise rather than run, it is a switch thrown for a whole run rather than for one line, and on CUDA torch.mm, torch.mv and torch.bmm raise under that switch unless CUBLAS_WORKSPACE_CONFIG is set
documented · Maestro_Memory_Palace_July2026/lessons/checkpoint_forward_branch_bug.md
62
A quotient that lands at 0.9999999 on the way forward and 1.0000001 when it is computed again clamps to 0.9999999 in one pass and to exactly 1.0 in the other, so a comparison against 1.0 is true in one pass and false in the other
documented · Maestro_Memory_Palace_July2026/lessons/checkpoint_forward_branch_bug.md
63
scale is computed inside torch.no_grad(), so the branch never entered the autograd graph; a no_grad block cannot change a gradient and can still change the shape of the recorded computation, which is the thing the non-reentrant checkpoint implementation counts, and that same boundary is why the telemetry read is safe once it is written to a plain Python attribute
record · Maestro_Memory_Palace_July2026/lessons/checkpoint_forward_branch_bug.md
64
The fix was to remove the conditional and multiply unconditionally, because the scale is already clamped with max=1.0, so an untouched position gets multiplied by 1.0 and the multiply is the identity there
record · Maestro_Memory_Palace_July2026/lessons/checkpoint_forward_branch_bug.md
65
The record I wrote for this bug is wrong about use_reentrant=False in two ways, and the note publishes the correction: the flag did not fail to prevent the bug, it is the reason there was a crash to read, because the saved-tensor count comparison exists only in the non-reentrant implementation and the same flip under use_reentrant=True raises nothing at all; and replaying the random number generator is preserve_rng_state, which defaults to on in both implementations, rather than anything this flag buys
record · Maestro_Memory_Palace_July2026/lessons/checkpoint_forward_branch_bug.md
66
if self.training: and if self.cap > 0: are decided by Python state, they are identical on both passes, and they were never part of this problem
documented · Maestro_Memory_Palace_July2026/lessons/checkpoint_forward_branch_bug.md
67
With nothing near the threshold every scale is 1.0 and the comparison is false in both passes, with everything past the threshold it is true in both, and the failure needs the regime in between; the crash came shortly after the counter tracking how much of the model was exercising the cap stepped up from a few to noticeably more
record · Maestro_Memory_Palace_July2026/lessons/checkpoint_forward_branch_bug.md
68
On my own model source the audit returned two hits worth reading: one live in a checkpointed path, which is the branch that had already crashed, and one in a code path that was switched off by default, so it was not executing
record · Maestro_Memory_Palace_July2026/lessons/checkpoint_forward_branch_bug.md
What was refused
The more useful half of the ledger. These are claims that were considered and are not made, with the reason. Several of them shipped on the previous version of this site and are recorded here as overstated by their own sources.
Multi-tenant memory isolation, per-tenant anything, tenancy of any kind
A full-tree grep for multi-tenant / tenancy / per-tenant returns zero files across the palace, and the infrastructure sweep states it outright: "Security, compliance, access control, or tenancy for GPU infrastructure. Nothing in the record." It has never appeared on this site and must never appear.
Any safety framing of "red-team": jailbreak testing, prompt-injection testing, toxicity or harmful-content evaluation
The evaluation sweep's thin-ice list: a grep for "jailbreak" and "prompt injection" across the whole palace returns nothing. "Caught six fake breakthroughs" (claim 1) means adversarial review of a mechanism and a metric, and must never be read as safety red-teaming.
Bias, fairness or demographic-disparity auditing
Zero hits in the record, no protected-attribute analysis, no disparate-impact testing. It is a standard expectation of "AI auditing" and the record does not support it at all.
PII redaction, de-identification, GDPR/CCPA/HIPAA posture
The scrubber matches eight credential formats and nothing else. Calling that "PII scrubbing" would misrepresent an API-key regex as a privacy control. This is also why the data boundary on the site refuses personal, customer and regulated data outright instead of promising to handle it.
Benchmark decontamination against public benchmarks, or any LoCoMo / LongMemEval / MemBench / LongBench result
Leakage work is against his own eval sets and his own training mix. No standard memory or long-context benchmark has ever been run. The circulating "~7x the published state of the art on LoCoMo" splices an internal metric against other people's numbers and his own audit flags it as an error.
"85-90% at 9-11K, ~7x published frozen-backbone adapters"
WINS_DIGEST §A flags the cross-scale comparison trap — the baseline was never rescored on the harsher scale. Needs same-scale verification before it can be published.
DISC_RANK "+1.3 FP0"
Single-seed s0; s1 shows 68.8% with an 11.1% FP control breach. Provisional in the record.
Any tri-state / SMB automation success story from the old site
Unsourced and apparently invented.
"fabrications: zero" and "the model cannot emit text it never stored"
Both shipped on the live site and both overstate their source. The measured quantity is fabricated memories on a memory benchmark, and the mask is on the recall path and armed by the harness, not by the shipped model. Replaced by claims 8 and 9.
"every routing evaluation had been testing an untrained model"
Shipped on the live site and wrong by one word: the weights were trained, the routing never loaded. Replaced by claim 3.
"client", "in front of a client", "I hold one engagement at a time" as a report of past demand
All presuppose paid engagements. None are documented anywhere in the palace, and the hiring-panel report says the opposite in plain words. The one-engagement-at-a-time line ships only as a forward-looking constraint of the method, and says so.
"I work alone" (unqualified)
Rests on an absence rather than a source, and is checkable against a record dense with review passes and dispatched work. Replaced by "no employees and no co-founders" (claim 31).
"Of roughly 18 capability claims, ~5-10% are behaviorally proven and ~55-60% proxy-only"
A proxy-share figure quoted without its instrument. Replaced by claim 44, which states the practice instead.
Any statement that the NDA or the engagement letter is attorney-reviewed, legally vetted, standard, enforceable, or drafted by counsel, and any naming of a law firm
The engagement kit is a drafted, unreviewed document set. What ships is the process, which is true and forward-looking: what is signed, before what, and what it binds. The documents' legal status is not asserted in either direction.
Any URL other than https://maestro-overview.vercel.app
The other five deployments are live but unlisted, and one of them exposes a supplier's cost sheet.
This is the same discipline an engagement runs on. See what you can buy, or send me the number.
