How it works
The page runs in the order the work does. Where the questions come from, what happens to a document when you upload it, where your own notes go, and what becomes of a sentence the platform drafted before anyone read it.
- SoftwareAnswered
What is tested, and is that enforced before a merge?
test coverageautomated testingCI gateregression suiteChecked against the repository extract
- People & OrganizationAnswered
Is knowledge concentrated in one person, and who is the second owner?
key person risksingle point of failuresuccessionknowledge transferChecked against the commit authorship
- SecurityNothing found
How often is access reviewed, and who signs it off?
access reviewjoiners movers leaversprivileged accessentitlement reviewWhat the silence implies
Access governance is undocumented. Standing privilege accumulates when nobody reviews it.
Added to the request list under Evidence gaps:Access review procedure and the last two completed review records
- Architecture & DesignAnswered
What still runs on hardware the company does not operate?
on-premiseshosteddata centermigration planChecked against the deployment configuration in the extract
Six review areas carry a probe kit like this today: Software, Security, Architecture & Design, Performance & Scalability, People & Organization and Engineering Process. The other six run from their question library and rating contract. Every probe runs and is recorded, including the ones that come back empty.
Four probes from one run. The Security probe found nothing, so the figure shows the implication it states and the request it added to the list.
- Software
- Security
- Architecture & Design
- Performance & Scalability
- AI/ML
- Open Source
- Engineering Process
- People & Organization
- Product Management
- Suitability for Target Market
- Technical Operations
- Support & Services
Before anything is read
The questions were written before your documents arrived.
Six of the twelve review areas carry a full probe kit today; the other six run from their question library and their rating contract. The probes were written by people who took these assessments to completion and saw which omissions mattered, and they are built in, so every one of them runs on every project.
Each probe has four parts.
- 1
The question the area has to answer
Written for a person to read. Ask a general-purpose tool to assess engineering process and it paraphrases the phrase back at the documents a few times and reports the overlap as coverage.
- 2
The words to look for it under
The memorandum calls it key person risk; the board pack calls it a single point of failure; the engineering wiki is less polite about it, and all three are the same finding. The probe carries every phrasing we have met, so a risk the target described in its own words is found.
- 3
What checks it in the code
Contributor counts, dependency licenses and file health can be turned into checked facts from the repository extract, and the probe names the check where one exists.
- 4
What its silence means
The implication, the rating that absence alone justifies, and the item to add to the request list. A probe that finds nothing writes that ask itself, under Evidence gaps.
What took the transactions to write
Searching a folder is the easy half. The list of what to ask, in what order, and what to do when the room goes quiet was assembled across completed transactions and argued over line by line. A firm adds its own layers on top: analysis rules, section prompts, a request template, line items. The shipped probes stay underneath.
Reference knowledge
The Brain keeps a record of what it consulted.
Alongside the probes sits a reference library: the practitioners' working notes on what good looks like in each area, versioned like code. We call it the Brain. The analysis consults it the way an expert draws on experience, in context and on the record.
Activated by the section's questions
The library is never poured in wholesale. An AI/ML section calls up different norms than a security review, and the analysis can consult an asset by name while it reasons.
Stamped into the section's record
Every consultation lands in the section's record with the asset's id and version. When the library improves, you can see which reports predate the improvement.
Kept out of the citations
A finding about the target may cite only the target's own material. Reference knowledge shapes what to check and how to read the answer. Each asset it renders carries a stamp, and a check run before every release reads every stored section for that stamp.
Section
Model governance and evaluation practice
Calibrated by
reference library, never citedai-adoption-normsv12What a gap between claimed and shipped AI features usually looks like.
model-eval-red-flagsv7The checks a mature team runs before a model ships.
Grounded in
project evidence, cited in the report- Platform Architecture Pack v4page 14
Model scoring runs nightly; evaluation before release is a manual spot check by the data science lead.
- Technical deep-dive, 12 March00:41:12
We do not have a formal evaluation suite yet. The team reviews sample outputs by hand each sprint.
Both records stay with the section. The assets are stamped by id and version, so a later reader can see which expertise calibrated the judgment, and a release check reads every stored section to confirm that none of it reached the report.
One section's record: the reference assets consulted, with id and version, above the passages it cited, each with its exact place. The report can draw only on the second list.
The pipeline
From the request list to the issued report.
Step 1
Collect
It starts as a request list
A new project opens on the standard request template: 85 items in 16 sections. Your team adds to it, and every probe that finds nothing adds its own ask under Evidence gaps, naming the section that asked. Files arrive against the item that asked for them, from your team or from the target, whose people are invited as Contributors: they work the request list and see none of your analysis. A request the target cannot satisfy is closed as Nothing to provide, with a reason, a name and the date, and stops counting as outstanding.
Step 2
Ingest
Read in, with the layout kept
PDF, Word, Excel, PowerPoint, CSV, text, Markdown and images, up to 100 MB a file. A table keeps the note above it and a figure keeps its caption. Every passage keeps its exact place: the page, the line, the heading, or the moment in the meeting. A file the platform could not read says so, right on the file.
Step 3
Describe
Named the way you will search for it
Former names, subsidiaries and alternate spellings are recorded in the project brief, so a search for the target finds every mention of it. Meetings arrive as transcripts, uploaded or pasted with the date and the participants. The analyst attributes the speakers, and that attribution travels with every quotation.
Step 4
Align
Claims read against the evidence
The assertions in the management material are extracted and read against the project documents. Each lands as supported, partly supported, contradicted or without evidence, carrying the passages behind that reading. A contradiction is shown with both sides cited.
Step 5
Compose
Blocks into a report
Sections are drafted from the aligned evidence and edited as blocks, each carrying its review stage and its audience. The report is issued as a numbered version in the editions you choose, shared by invitation with named readers, and can go out for a review round.
The parts
What ships, in numbers.
- 12
review areas
named in the strip above
- 6
probe kits
areas with a full kit today
- 85
request items
in 16 sections, on day one
- 26
block types
in the report composer
- 4
reading depths
Executive through Everything
- 18
code figures
in the Software section, once signed
Documents
Open the passage it cites.
Every figure on this site says exactly where its evidence lives. This is the screen that opens it: a search over the whole project, the viewer it lands in, and the state of each file beside it.
- Platform Architecture Pack v43 Infrastructure › 3.2 Batch processingp. 23CurrentGo to source
Settlement batch (SETL-02) executes nightly against the customer-hosted AS/400 instance. Migration to the managed service is scheduled but not yet started.
- Runbook 20235 Overnight jobsp. 4SupersededGo to source
The settlement batch is operated by the customer platform team and monitored by pager.
Arden Systems
Platform
- Single multi-tenant service across three regions
- The platform is fully cloud-based, with no remaining on-premises dependency.
- Nightly settlement completes before 04:00 in every region
CL-014Contradicted
Architecture Pack v4, p. 23: the settlement batch still runs on customer-hosted hardware.
Index state
- Management presentation.pptxIndexed
32 slides, 118 passages. Shown as the rendered PDF.
- Headcount by function FY25.xlsxDescribing figures
40 passages searchable already; the charts are still being described.
- Incident log 2024.csvPartly readable
38% of what was extracted is too thin to use.
- Board minutes Q4 (scan).pdfNo readable text
A scan with no text layer. A text-based copy would fix it.
Also answers
Platform Architecture Pack v4.pdf was filed against one request and reads like it answers two more.
Disaster recovery arrangements and the date of the last test
Hosting providers and regions in use
Versions
Current infrastructure diagram
- v3LivePriya Nair (Arden), 14 AprilRemoves the decommissioned Frankfurt site
- v2SupersededT. Adeyemi, 6 AprilRollback from v1
- v1SupersededPriya Nair (Arden), 2 AprilAs provided
v2 to v3
RemovedSecondary site: Frankfurt (decommission planned Q2)
AddedSettlement batch remains on customer-hosted hardware pending migration
sha256 9f3c…e21
One viewer for every format. A file the index could not read says so on its chip, and a chip never reads Indexed over zero passages.
A search for the settlement batch, the slide it landed on with the contradicted claim pinned to it, and one document's three versions with the difference between two.
One viewer for every format
PDF, Word, Excel, PowerPoint, CSV, text, Markdown and images open in the same viewer. An Office deck or document is shown as the PDF rendered at indexing, so slide 11 looks like slide 11, and search hits step through the file with a page or slide counter.
A chip that says what the index could read
A scanned PDF with no text layer indexes without an error and can never be found by a search, so its chip reads No readable text. A thin scan reads Partly readable, a file mid-way reads Indexing, and an indexed one carries its passage count.
One search over text, tables and figures
Cmd or Ctrl+K opens a palette over the whole project. Results are grouped by document under the heading they sit beneath, marked Current or Superseded when a newer version has arrived, and Go to source lands in the viewer at the passage.
Versions, with the difference shown
When the target replaces a file, every version stays, with who filed it and why. A rollback says it is a rollback. Two versions compare at passage level, and a download returns the bytes of the version you asked for, with a fingerprint short enough to read aloud.
The file suggests where else it belongs
Once indexed, a document is read against the open requests, and the ones it appears to answer are proposed as links. A person accepts or declines each one; a declined suggestion is not raised again.
Web pages, only when you name them
An analyst can add a named web page to the project as evidence. The analysis never searches the open web on its own; public package registries are consulted for the code analysis and nothing else.
Citation
What a citation is here.
Citation is most of the product, so it is worth being exact. A finding names the document and the exact place inside it, and that place takes the shape the format allows.
A page, for anything paged
A PDF passage carries its page, so the source opens at the paragraph. The Office formats are rendered to PDF at indexing and cite the same way, by page or by slide.
Text carries a line range or a heading
Plain text, Markdown and pasted material cite the line range or the section heading the passage sits under.
A transcript cites the moment
A meeting arrives as a transcript, uploaded or pasted, and its evidence is anchored to the timestamp it was said at. The speaker is whoever the analyst attributed the line to, and the transcript is named as the source, so you can tell a claim that rests on something said from one that rests on something written.
Code cites files and a date
A software finding names the files it was found in and the date the extract was taken. Nothing about the code is presented as live.
Claim
“The platform is fully cloud-based, with no remaining on-premises dependency.”
Management presentation, slide 11
The architecture pack still shows a nightly settlement job on customer-hosted hardware, and in the 12 March transcript the CTO describes it as unchanged.
Sources (2)
- Platform Architecture Pack v4page 23Document
Settlement batch (SETL-02) executes nightly against the customer-hosted AS/400 instance. Migration to the managed service is scheduled but not yet started.
- Technical deep-dive, 12 March00:41:18Transcript
It's the one thing still sitting on their own tin. We've talked about moving it for two years now.
A contradicted claim. Slide 11 says fully cloud-based; page 23 of the architecture pack and a transcript at 00:41:18 say otherwise, and both are attached.
Ask it
Ask this project a question.
The questions that come up the night before a committee are small ones: has the SOC 2 Type II report been issued, where does the settlement batch run. The panel answers from the project documents and nothing else, cites the passage, lists the searches it ran and puts a confidence figure on the answer.
Every sentence traces to a passage
The answer cites its sources by number, and a citation opens the document at the page. The Brain shapes how the question is read and is never cited as evidence.
When the documents are silent
A question the documents do not answer gets one fixed sentence, what would settle it, and the request to make.
Does any production component run on customer-operated hardware?
Yes. The nightly settlement batch (SETL-02) runs against a customer-hosted instance, and migration to the managed service had not started at the date of the architecture pack [1]. In the 12 March session the CTO described it as unchanged [2].
Sources (2)
- [1]Platform Architecture Pack v4 p.23“Settlement batch (SETL-02) executes nightly against the customer-hosted AS/400 instance. Migration to the managed service is scheduled but not yet started.”
- [2]Technical deep-dive, 12 March 00:41:18“It's the one thing still sitting on their own tin. We've talked about moving it for two years now.”
Has a SOC 2 Type II report been issued?
The project documents do not answer this question.
The one related file is a readiness assessment describing an observation period that had not closed at its date. The report itself, with its bridge letter, would settle this.
Request to make: Has a SOC 2 Type II report been issued?
Sources (1)
- [1]SOC 2 readiness assessment p.2“The Type II observation period commenced in January and is expected to conclude in the third quarter.”
Every sentence in an answer traces to a passage in the project. A question the documents do not answer gets the sentence above and the request, never a guess.
Two questions. The first is answered from page 23 and a transcript; the second is not answered, and the panel says so and names the request.
Analyst files
Where your own notes go.
A platform that holds only what the target gave you and the report you wrote loses everything worked out in between: the management call, the penetration test you commissioned, the note that a mis-transcribed word is the name of an acquired subsidiary. Percomb files those beside the target's documents, as evidence, under your team's name.
Team notes
An observation, an issue or a question, anchored to the passage that prompted it, and yours the moment you write it.
Other materials
A report you commissioned, a customer reference call, a benchmark. Filed as evidence and analyzed with the rest of the project.
Meeting transcripts
Pasted or uploaded with the meeting date and the participants. Claims are mined from them and cited to the timestamp.
Filed as yours
Each is filed under your team’s name and marked Internal, kept apart from what the target provided and analyzed alongside it.
Proposals, from a person or an agent
A finding, a rating, a note or a question can be proposed against a section by a colleague, by the platform, or by an agent you have given a key. Each is recorded with its evidence and its author and stays pending until a person accepts, adjusts or rejects it. A judgment a person has accepted stands over the platform's proposal. Where two sources disagree, both are shown with their citations; nothing picks a side for you.
The review stage
A draft stays a draft until somebody has read it.
The rule lives in the data model rather than in a prompt: a drafted block carries the Draft label until a person accepts it or rewrites it, and the report says which above every block.
Draft
The platform proposes a finding from retrieved passages. It is stored as a proposal, with its evidence and its author, and the block above it reads Draft.
Reviewed
A person reads it and accepts it as written.
Revised
A person changes it. The original is kept, so a revised finding remembers what it was revised from.
Analyst
Written by a person from the first word.
The label sits above every block at the same weight as the block type. One report reads at four depths, Executive, Management, Technical and Everything, from one version history, and there are 26 block types to compose it from.
Proceed, subject to three conditions.
The platform is fit for the plan and the engineering practice holds up. Two exposures are worth pricing in: a settlement job that is not where management says it is, and one engineer holding the service that moves money.
What it costs
$240k
first twelve months, across four lines
- Migrate the settlement batch$85k
- Second owner and runbook for settlement$70k
- Complete the attestation observation period$55k
- Patch three advisories$30k
Coverage
11 of 12 areas
74 indexed documents, 3 meeting transcripts, one repository extract
Technical Operations is Not Assessed. The incident log was asked for eleven days ago and has not arrived.
Worst findings
- Critical
Key person risk on the settlement service
People & Organization
- Critical
Compliance attestation not held as claimed
Security
- Warning
Settlement batch runs on customer-operated hardware
Architecture & Design
The document
- Executive Summary18Warning
- Software9Warning
- Security11Critical
- Architecture & Design10Warning
- Performance & Scalability6Good
- AI/ML3Info
- Open Source8Warning
- Engineering Process7Good
- People & Organization8Warning
- Product Management5Good
- Suitability for Target Market5Good
- Technical Operations2Not Assessed
- Support & Services4Good
Thirteen sections, 96 blocks. Each row opens the document at that section.
Issue 3 went out as two editions from the same blocks: a Board pack for the committee, which stops at executive tier, and a Full report for the deal team.
The Executive Summary of a fictional report, read at four depths. The counter says how many of its blocks the current depth shows.
2 people outside your team can still open a superseded issue. R. Okafor and D. Ferreira lose access to Issue 2. Issue 3 is unaffected.
- Working draftDraft
Two sections ahead of Issue 3. Six drafts to review before it can be issued.
- Issue 3Current18 AprilFull reportBoard pack
Security moved from Warning to Critical. Four elements rewritten, two added, one removed.
Added 2 · Rewritten 4 · Removed 1 · Reordered 0What changedStop sharingWithdraw
- R. OkaforLender’s adviseraccepted
- L. BrandtCo-investorinvited
- S. WhitcombeArden CTOaccepted
- Issue 2Superseded4 AprilFull report
Architecture section added. People & Organization moved from Good to Warning.
- R. OkaforLender’s adviseraccepted
- D. FerreiraCo-investoraccepted
- M. SatoBoard observerexpired
2 people can still open this issue.
- Issue 1Withdrawn21 March
Reason: Headcount table restated after the org chart arrived. Kept in the history with its reason and its recipient list.
Review round 2 on Issue 3
Arden management · 2 threads · 1 openQuestion on Finding · Settlement batch runs on customer-hosted hardware
S. Whitcombe (Arden): Migration to the hosted environment completed 3 April. Runbook v7 attached.
Open
Suggestion on Recommendation
R. Okafor: Name the second owner of the settlement service as a closing condition.
Resolved. Block now Revised.
Outside readers
Single-use invitation link, valid 14 days
L. Brandt
Confidentiality agreement v3 and non-reliance notice v2, accepted 19 April
R. Okafor
Confidentiality agreement v3. Reliance letter executed 12 April, recorded against this share.
- Every page watermarked with the reader’s name
- Print and save-as locked on every report surface
- A reliance letter is recorded as an executed instrument against the share it covers.
Nothing here is deleted. A withdrawn issue stays in the history with its reason, and a bookmarked link to it says so instead of showing findings the team has retracted.
The versions page of a fictional project. Two people still hold a superseded issue, and one action closes it. Below, a review round with its threads, and the gate an outside reader passes.
After it is issued
Who can still open the version you replaced?
Counsel asks this a month later. The answer is on the page, with the button that closes it.
Versions accumulate
Publishing takes a frozen snapshot; nothing points at a live draft. Issues are numbered, any two compare — from a one-line summary down to the words that moved — and an issue can be restored, or withdrawn with the reason recorded.
Editions at publish
The same draft can be issued as a Full report, a Board pack, a Management report or a Technical report. A Board pack carries nothing below the executive tier, so its reader cannot page into the appendix. A reading depth hides blocks from the person reading. An edition takes them out of the file that goes to them.
A review round
An issued release goes to the target's management, a co-investor or your client for markup. Comments anchor to the block they dispute, the round is numbered, and it closes with a verdict: approved, or changes requested. Reviewers hold no seat.
Sharing under agreement
An outside reader gets a single-use invitation link and passes a confidentiality and non-reliance gate before the report opens. Every page is watermarked with their name and print is locked. Where reliance is granted, the reliance letter is recorded as an executed instrument against that share.
Source code
The code is read at arm's length.
We never hold a credential to your source host. Your engineers run the extractor inside your own network and upload what it produces, and a public repository is connected by pasting its URL and cloned without credentials. Every figure states the date its extract was taken.
Sources
3 codebases in scope
Index
2,318 commits, 412 components read
Review
Complete with gaps: 2 of 3 codebases reviewed
Enrich
Registry and advisory lookups done
Decide
2 conclusions, 1 request, signature in place
Cross-check
History read against 74 documents
Report
18 figures built from the signed review
Codebase
4contributors
3 still active; 1 key person. The memorandum says fourteen engineers.
Components
412components
Ships 187. Builds and tests only 219. Undetermined 6.
Findings
4open
1 High advisory with a fix available; 2 license conclusions to decide.
Posture
41days to fix, median
No update mechanism declared. 14 of 412 components past end of life.
Assumptions this run made
These are decisions the analysis made so it could proceed. Each can be wrong. Change one and the review re-evaluates.
- firm
Distribution model: hosted service
Deployment runbook, page 3
- likely
Copyleft stance: strong reciprocal in shipped code is a finding
Organization default
- weak
Vendored tree origin: a fork of an upstream 2.x release
Header comments only. Change it here.
Needs you
- Decide
Approve 2 license conclusions
A second person approves each before it is written.
- Send
Send 1 disclosure request to the target
The question, the evidence held, the answer format expected.
- Sign
Accept the delta against the signed review
Run 4 finished after the signature; nothing signed was replaced.
Three acts stay yours whatever the switch says: sending a request to the target, signing the review, and approving a license conclusion.
Components
Ships 187Builds and tests only 219Undetermined 6pkg:npm/ledgerline@4.18.2
- Catalog
- Ships
- License:
- Permissive
- Advisory:
- High advisory open, fix available
- Freshness:
- 2 stable releases behind
- Modified:
- —
pkg:pypi/settle-core@2.1.4
- Catalog
- Ships
- License:
- Weak reciprocal
- Advisory:
- None open
- Freshness:
- Current
- Modified:
- Yes: 3 files changed from upstream
pkg:golang/queue-fabric@1.9.0
- Catalog
- Undetermined
- License:
- Unrecognized
- Advisory:
- Medium advisory open, no fix yet
- Freshness:
- No release in 1,140 days
- Modified:
- —
pkg:npm/spec-runner@9.0.1
- Catalog
- Builds and tests only
- License:
- Permissive
- Advisory:
- None open
- Freshness:
- Current
- Modified:
- —
License conclusion
pkg:pypi/settle-core@2.1.4
- Declared
- MIT
- In the tree
- One file carries a GPL-2.0 header
- Registry
- MIT
- Proposal
- MIT, likely
One header disagrees with the declaration and the registry. A person decides.
Approved by J. Lindqvist. Written when a second person approves it.
Disclosure request to the target
Request 2 of 3
Is settle-core distributed to customers, or does it only run the hosted service?
Sent 10 April
CompletedAnswered by Priya Nair (Arden), 16 April: hosted only. The finding re-evaluated.
Sign-off
Signed by T. Adeyemi on 18 April.
Pins runs 1 to 3 and policy v4. Run 4 finished since and opens a delta below; the signed review is unchanged.
Delta from run 4: 2 conclusions, 1 request. Nothing signed has changed.
The report Software section is built from the three pinned runs: 18 figures, each naming the rows it was computed from.
The Brief for a fictional codebase after a signed review, with one later run opening a delta against it. Every figure in the frame is illustrative.
The Brief
The Software tab opens on one screen with seven stops: Sources, Index, Review, Enrich, Decide, Cross-check, Report. Each says where it stands and what remains. Four readings sit one click below: Codebase, Components, Findings, Posture.
Decide for me
Turn it on and the analysis takes every decision it is allowed to, writing each one on the Brief as an assumption graded firm, likely or weak, with where to change it. Turn it off and it asks you instead. Three acts stay a person's either way: sending a request to the target, signing the review, and approving a license conclusion, which takes two people.
The inventory
One row per component across every repository, each identified by its exact package and version. What ships, what only builds and tests, and what is undetermined stay on screen at all times, with the license reading, the advisory severity, whether a fix exists, and whether the target modified its copy. The Posture reading adds the target's own time to fix and, once consent and enough peers exist, a position against comparable companies.
Signed, then in the report
The review closes when a named person signs it, which freezes exactly what was analyzed and under which policy. An unsigned review produces no report figures. A signed one becomes eighteen figures in the report's Software section, each naming the rows it was computed from, and the Cross-check stop reads the history against what the documents claim: four active contributors, fourteen in the memorandum.
For advisory firms
Your method sits on top of ours.
If your firm already runs a diligence practice, Percomb is machinery for it. The shipped probes and review areas stay in place; the layers a firm's method lives in are yours to set.
Analysis rules
Tone, business context, and the critical keywords the analysis always probes, applied to every project your firm runs.
Section prompts
How each of the thirteen report sections is written for your practice. Edit the prompt; the citation requirement and the review stage stay.
Request template
Every new project starts from your firm’s list. Add items, reword them, regroup them. Reset to default puts the standard list back.
Line items
Add your own line items to any review area. The areas themselves are added or retired for you.
Agreements in your counsel’s words
Confidentiality and non-reliance text as versioned templates, in clickwrap or acknowledge mode, with per-project overrides. Whoever accepted version 2 is recorded as accepting version 2.
Your own agents
An analyst mints a project-scoped agent key: read, write or research. Findings the agent submits arrive as proposals attributed to the person and the agent, pending until a person accepts them.
Branding is applied for you and grows with the plan: your logo and colors, then full firm branding, then white label.
The analyst's seat
A seat for the close read.
An analyst reads for what does not add up, and that reading has usually survived as a margin note, if it survived at all. The Percomb desktop appis a seat for it: a note pinned to the region of the page that prompted it, filed into the project as yours.
Anchored where the thought happened
A note, an issue or a question is pinned to the exact region of the page (why has this customer not deployed for ten months?) and lands in the project as an analyst note beside the evidence it is about.
Coverage you can answer for
Reviewed, skimmed or flagged, tracked per document. When the team asks what today turned up, the answer is on the record: nine of ten read, four flags, one major.
The notes from the plane
A senior reviewer writes a page of notes at best, and those lines are often the most valuable thing in the project. So they are filed as analyst notes, in the project, next to the document that prompted them.
The demo runs the whole chain.
A fictional deal, taken from the request list to an issued report, including the question the evidence cannot answer.