Paste an operations manual, SOP, or workflow — one procedure per line — and the one-sentence goal it serves. Get three readings from one input: a resilience reading (can the manual keep the goal on track when a key person is out, a vendor fails, or an emergency hits — does it absorb, stay anchored, and recover), the gaps (a risk named with no stated control, single points of failure with no documented backup), and a coherence score (is each procedure a clear, testable action, or a vague label). Same engines as Gap Check and Framework Audit, pointed at a new domain. In the atlas grammar: the manual is the claim; its named backups and controls are the witness; brittleness is the difference. Heuristic, deterministic, entirely on your device — nothing you paste here ever leaves your browser.
+
Look for handover gaps
+
What happens when the usual person is away? Answer five short questions about a task and its backup. You will get questions to follow up, not a safety score.
-
+
+
Start with one task
+
+
Leave anything you do not know empty. Your answers stay on this page; this tool does not save or send them.
+
+
+
+
+
+
+
Naming someone does not confirm their permission or access. Leave this empty if you do not know.
+
+
+
+
+
Even a yes answer is not verified here. A real attempt and its result need to be checked separately.
The original Continuity Audit looks for word patterns in a longer document. Paste one procedure per line to see its gap findings. Optional numeric diagnostics are text-based heuristics; they do not measure operational resilience or verify actual arrangements.
+
-
+
-
-
+
+
@@ -70,6 +122,7 @@
Delta Atlas - Continuity Audit
Honest limit: this reads words and patterns, not meaning or your operation's actual arrangements. It flags likely gaps to point you where to look; it does not deliver a verdict, and it cannot see anything you haven't written down. The judgment stays yours. Atlas does not submit the text you enter to a model or API. A private overlay is optional and data-only; invalid or unsupported files are rejected.
+
@@ -302,31 +355,29 @@
Delta Atlas - Continuity Audit
const parts=raw.map(function(t,i){return SCORER.scorePart(t,goalTokens,raw,i);});
const roll=SCORER.rollup(parts);
function col(v){return v>=68?'var(--green)':v>=45?'var(--accent)':v>=28?'var(--amber)':'var(--red)';}
- const vtxt = roll.overall>=75?['Solid manual','var(--green)'] : roll.overall>=55?['Mostly solid, a few soft spots','var(--accent)'] : roll.overall>=38?['Half-formed, needs work','var(--amber)'] : ['Mostly vague','var(--red)'];
-
let h='';
- // Resilience headline — frames the whole audit: does the goal stay resilient throughout?
+ // Preserve legacy numeric outputs, but present findings before optional diagnostics.
const R = resilienceReading(A, roll, parts);
- const rv = R.overall>=72?['Reads resilient — verify against real events','var(--green)'] : R.overall>=52?['Holds, with soft spots','var(--accent)'] : R.overall>=34?['Brittle in places','var(--amber)'] : ['Not yet resilient','var(--red)'];
- h+='
Resilience of the goal
';
- h+='
Can this manual keep the operation on its stated goal when disrupted — a key person out, a vendor failing, an emergency? '+R.overall+' · '+rv[0]+'
';
- [['Withstand','absorbs the hit — named risks have a control',R.withstand],['Anchor','stays true to the goal under stress',R.anchor],['Recover','resets to normal after a disruption',R.recover]].forEach(function(row){
- h+='
'+row[0]+' '+row[1]+''+row[2]+'
';
- });
- h+='
Resilience here follows the Resilience Ledger: absorb load, stay anchored to a fixed reference, recover to baseline — the withstand/recover pattern from ecological resilience (Holling 1973) and cyber-resiliency engineering (MITRE 2011; NIST SP 800-160 v2). A whole-manual reading, not a per-line score, and it credits real structure (controls that defend named risks, recovery procedures) — not reassuring vocabulary. Honest limit: it still reads words, so a determined author can inflate a middling read by naming controls that aren\'t really there. A clean read points you to verify; it does not certify.
';
-
- h+='
Gaps & coverage
';
- h+='
'+A.M.length+' governance terms found · '+A.areas+'/'+KB.clusters.length+' areas touched · '+A.defended+'/'+A.risks.length+' risks defended
';
- if(!A.findings.length){ h+='
No gaps found
Every risk this scan recognizes has a stated control alongside it.
'; }
+ h+='
Gaps to check in the text
';
+ h+='
These are word-pattern findings. Confirm whether each arrangement is current, relevant, authorized and usable. Arrangements are not verified by this scan.
';
+ if(!A.findings.length){ h+='
No pattern-based gaps detected
This scan did not match a gap rule. It has not checked whether a backup exists, has access, or can do the task.
'+roll.overall+' — '+vtxt[0]+' · '+roll.clearN+' of '+parts.length+' procedures are clear and testable'+(roll.vagueN?'; '+roll.vagueN+' need work.':'.')+'
All values below are 0–100 textual heuristics, not probabilities, verified controls or measurements of operational resilience. Merely listing control words can inflate them. The existing calculations are retained for comparison with earlier runs.
';
+ h+='
Whole-text pattern scores
Combined text-pattern score: '+R.overall+'/100.
';
+ [['Control-word coverage','recognized risks and control terms occur in the same text',R.withstand],['Goal-word alignment','goal overlap on lines matching action and measure patterns',R.anchor],['Recovery-word coverage','recognized recovery terms, reduced by some matched gaps',R.recover]].forEach(function(row){
+ h+='
'+row[0]+' '+row[1]+''+row[2]+'
';
+ });
+ h+='
'+A.M.length+' vocabulary terms matched · '+A.areas+'/'+KB.clusters.length+' areas matched · '+A.defended+'/'+A.risks.length+' recognized risks have a related control term in the text. Co-occurrence does not establish an actual defense.
';
+ h+='
Procedure-wording diagnostics
Wording score: '+roll.overall+'/100 · '+roll.clearN+' of '+parts.length+' lines reach the existing wording threshold. This does not verify that a procedure is clear, testable or workable.
';
SCORER.DIMS.forEach(function(d){ var v=roll.dimAvg[d.k];
- h+='
'+d.label+' '+d.help+''+v+'
';
+ var label=d.k==='continuity'?'Backup-word pattern':d.label, help=d.k==='continuity'?'backup wording near a single-person phrase':d.help;
+ h+='
Paste an AI policy, project plan, or vendor claim. It reads the governance terms you used and flags what is missing: risks named with no control, autonomy with no oversight, and terms that are still thin. It checks structure, not correctness, and it runs entirely on your device. In the atlas grammar: your document is the claim; its own stated controls are the witness; the findings are the difference.
+
Delta Atlas · Gap Check
+
Review the wording of an AI plan
+
Find possible gaps using fixed risk and safeguard patterns. This check can miss everyday wording and cannot verify that controls work. Your text stays on this device.
-
+
+
+
Describe what the AI can do, what could go wrong, and who can approve or stop its actions.
+
-
-
-
-
+
+
+
-
Paste something on the left and press Check for gaps. Or press Load example to see how it works.
+
Review results
Try the example to see a plan and its findings, or add your text and select Check for gaps.
+
diff --git a/Delta-Atlas-Tracer-Calibration.md b/Delta-Atlas-Tracer-Calibration.md
index 85cdeb6..d5d7234 100644
--- a/Delta-Atlas-Tracer-Calibration.md
+++ b/Delta-Atlas-Tracer-Calibration.md
@@ -151,3 +151,15 @@ The shell brand is now a native Home anchor, enhanced with the existing one-entr
Twenty-one local text-document links now use a reader with persistent Home/Library links and an original download. Its 120-path public manifest restricts file selection; textContent rendering keeps document markup inert, with a 1 MiB preview bound. Raw source URLs and machine-readable bytes remain available. Seventeen external-source links now open separately, preserving the Atlas page. This does not add site controls to third-party pages or change a browser's own raw-file viewer.
Local screenshots at 390 x 844 verified a long Markdown document, JSON and an invalid-path state. Home remained at an 8px top edge after approximately 24,118px of document scrolling and returned to the front page with one click. Standalone Ask/Explore at 390px, Verify/Field/404 at 320px, and desktop embedded sample/Tracer return were inspected. A mobile sidebar-height override was corrected during review. No horizontal document overflow was observed in these checks. These are browser viewport observations, not physical iOS/Android installed-app certification. Cache v93 includes the navigation and reader assets plus directly linked documents; arbitrary manifest-listed files are not all pre-cached. No tool scoring engine, dataset acceptance, artwork or motion policy is changed.
+
+## 2026-09-05: Beginner entry and honest text-tool results
+
+The owner requested simpler descriptions and a useful starting point without needing a policy document or action log. A live utility review found that an ordinary customer-email AI idea matched no Gap Check vocabulary; Continuity mixed a correct single-keyholder flag with an 80-point backup score for waiting for the absent owner; Tracer suppressed an explicit bypass after a preceding blocked-check clause. These observations remain counterexamples to broad utility and operational-assurance claims.
+
+The homepage now leads with a five-question work-handover checklist and describes AI-plan wording review and technical action-log review in plain English. Each tool has an example above its inputs. The checklist reports missing answers and follow-up questions, with arrangements always not verified; it uses literal text rendering, no model calls, no submission and no persistence. The older Continuity text audit and its unchanged numeric calculations remain under Advanced, with findings first and numeric text-pattern diagnostics collapsed. Naming controls still does not establish that they work.
+
+Gap Check retains its detector and data. Its no-findings output no longer invents oversight, unmatched ordinary wording offers an example, and recognized terms and safeguards are labeled as wording matches. Suggested definitions use native buttons; primary inputs and result updates have labels, focus and status. The zero-term result no longer hides a separately recognized red flag.
+
+Tracer pattern version 0.4.1 retains explicit local bypass wording when a preceding clause says a check held or an authority refused a change. The recorded T3/T4 counterexamples now flag a bypass rather than returning clean held/amended results. Positive cases for denied bypasses, valid stops and active/passive external amendments remain checked. Output is phrased as text clues and does not verify actions, permission or safety. Complex negation, quotation, hypothetical wording and multiple actions per text unit remain limits.
+
+The focused corpora cover 24 Tracer, 19 Gap Check and 26 Continuity cases, including eight handover checks. Browser review covers desktop and mobile-width first screens, examples, follow-up results, keyboard definition toggles and Home returns. Automated and agent reviews are not independent human usability testing; beginner comprehension, representative accuracy and productivity gains remain unestablished. Cache v94 adds the handover script. Routes, original data, research history, garden assets and motion behavior remain available. The new proposal is queued for cold read; implementation approval is not independent acceptance of the research.
diff --git a/Delta-Atlas-Tracer.html b/Delta-Atlas-Tracer.html
index 40ff888..fa4c818 100644
--- a/Delta-Atlas-Tracer.html
+++ b/Delta-Atlas-Tracer.html
@@ -12,9 +12,10 @@
textarea{flex:1;min-height:300px;resize:vertical;border:1px solid var(--line);border-radius:10px;padding:12px 14px;font-size:13px;font-family:ui-monospace,Menlo,Consolas,monospace;line-height:1.5;background:var(--panel);color:var(--txt);}
textarea:focus{outline:none;border-color:var(--accent);}
.row{display:flex;gap:8px;margin-top:10px;flex-wrap:wrap;}
- .btn{cursor:pointer;border:1px solid var(--line);background:var(--panel);color:var(--txt);font-size:13px;padding:9px 15px;border-radius:9px;}
+ .btn{cursor:pointer;border:1px solid var(--line);background:var(--panel);color:var(--txt);font-size:13px;padding:9px 15px;border-radius:9px;min-height:44px;}
+ .btn:focus-visible,textarea:focus-visible,summary:focus-visible{outline:3px solid var(--accent);outline-offset:3px;}
.btn.pri{background:var(--accent);color:#1e2b17;border-color:var(--accent);font-weight:600;} .btn:hover{border-color:var(--accent);}
- main{flex:1;padding:16px 22px 80px;overflow:auto;min-width:0;}
+ main{flex:1;padding:16px 22px 80px;overflow:auto;min-width:0;scroll-margin-top:calc(var(--atlas-return-height,0px) + 12px);}
.placeholder{color:var(--dim);text-align:center;margin-top:34px;}
.head{display:flex;gap:16px;flex-wrap:wrap;margin-bottom:14px;}
.metric{background:var(--panel);border:1px solid var(--line);border-radius:10px;padding:10px 15px;min-width:130px;}
@@ -47,23 +48,21 @@
-
Delta Atlas - Priority Tracer
-
Paste an agent's reasoning trace or execution log. It looks for ordering drift: the moment the mission climbs over the reference, where a goal is used to override, skip, or reinterpret a rule that should have held. It also marks where the reference held. It reads the grammar of the trace, not its meaning, and runs entirely on your device. In the atlas grammar: the trace is the claim; the rules it started with are the witness; drift is the difference.
+
Review an AI action log
+
Delta Atlas · Priority Tracer. Paste a record of what an AI assistant did, one action per line. This tool highlights wording about skipping rules, stopping for approval, or changing a limit. It checks the text on your device; it does not verify what happened.
-
-
-
- Log format (optional, improves precision): one step per line; declare binding rules early as
- Assert Baseline Invariant [ID]: "the rule";
- state objectives as Set Objective [Name]: ....
- The tracer still reads messy free text - the template only removes segmentation ambiguity. It is a
- convention, not a schema: a format requirement strict enough to be load-bearing would also be a
- recipe for writing traces that pass.
-
+
Try an example—no log needed.
+
+
+
Include the relevant rule and the recorded actions or tool calls. Use one action per line. You do not need the AI's private reasoning.
+
+
+ Optional format for declared limits
To compare later wording with a specific limit, start a line with Assert Baseline Invariant: "the rule". This adds a text-matching clue; it does not prove that later actions obeyed the rule.
+
-
Paste a trace and press Trace it, or load an example.
+
Paste an action log and press Review log, or try an example.