Wells Fargo
What happened
Why it matters
What TILOS examines
Evidence limitations
Sourcing and verification status:
THE INTEGRITY LAYER / COMPUTATIONAL LAB
TILOS is an open-source values optimization engine for governance. This first demonstration build runs part of its computation on three historical cases drawn from public records, one cycle per year, in your browser.
You will see evidence become signals, signals measured against each organization's own earlier record, and each year placed in a review band, with every input labelled by how well it is sourced.
Public records become signals: phrase counts in annual reports, complaint and incident counts, and a tribunal record. Every signal is tagged REAL or PLACEHOLDER, year by year.
Each signal is measured against its own baseline years, mapped onto a value axis from the Values Registry, and combined into an Integrity Delta for each value family.
Each year is placed in one of three review bands. In this build TILOS classifies only. Proposing and testing responses belongs to the full architecture and is not yet part of this demonstration.
01 · Starting point
Public companies leave a long, dated trail. Annual reports state what they value and which risks they face. Regulators, courts and journalists record what happened. Because these records are public, anyone can check the inputs, rerun the computation and challenge the result.
That makes them a fair first test for a governance engine. The evidence is open, the timelines are long, and the outcomes are already known, so we can see whether the computation responds where the record says it should, and where it does not.
Public records also have limits. They capture what was disclosed, reported or litigated, not everything that happened inside an organization. TILOS treats them as evidence about behaviour, never as a verdict on intent or character.
Every input on this page carries one of three evidence labels:
02 · Choose a case
Each case was chosen to test a different part of the computation: a long pattern and its remediation, discrete incidents against a steady commitment, and an AI system acting for an organization.
These cases were selected to test how the computation behaves under different evidence structures. They are not a representative sample of any industry, of these companies' conduct overall, or of how often problems occur.
03 · The computation
Each year runs through six stages. As the demonstration plays, each stage lights in turn and shows its result for Wells Fargo in that year.
Read this year's declared and operational signals.
Compare each count with its own baseline years and scale it to −1…+1.
Combine signals into value families and place them on registry axes.
Measure each family's distance from its baseline.
Turn distance into severity and place the year in a review band.
Add the year's scalar error to the trajectory.
z = (count − baseline mean) / baseline standard deviation, using the baseline years (2012 to 2015; Air Canada: 2021). Then value = tanh(0.5 × z), with the sign reversed for risk-disclosure signals. When a baseline never varies, the standard deviation is set to 1 count, so small changes can saturate at −1 or +1.delta = 0 − observed. The target of 0 means "matches its own baseline". This measures deviation from each family's own history, not yet the gap between declared and operational state, and not yet distance from values set by a governance authority.severity = max(0, delta − 0.15), so only movement toward the counter-value pole beyond a tolerance of 0.15 counts. Tier 3 if any family's severity is 0.4 or more; Tier 2 if any severity is above 0; otherwise Tier 1. The architecture's Tier 1, 2 and 3 describe how much autonomy a corrective action may have. Here the same labels are used as severity bands only, and no action is taken.E(t) is the plain average of the family deltas for the year, plotted as the trajectory. Families better than baseline offset families worse than baseline, so E(t) should be read alongside the individual families.Parameters (0.5, 0.15, 0.4 and the baseline years) are documented design defaults, not validated values. The computation is deterministic: the same inputs always produce the same outputs.
04 · Run the demonstration
Press Play to run every year in order, or Step to move one year at a time. Switching cases resets the run.
Each wheel is one behavioural axis from the Values Registry. Bright wheels are driven by signals in this build; hover one to see its evidence status. Dark wheels have no signal wired in yet. This build speaks to Foundation, Drive and Voice so far.
What is this? A wheel's colour and pointer show where that family's observed value falls on its 11-point axis this year, relative to its own baseline. A wheel driven by placeholder data shows how the method behaves, not a finding about the company.
What is this? The scalar error E(t) for each year so far. Above the centre line, families on average sit further toward their counter-value pole than in the baseline years. Dot colour shows the tier. Dashed lines mark documented events.
Review band from this build's severity thresholds. TILOS classifies here; it takes no governance action.
Run the demonstration to see a plain-language reading of each year's result.
05 · Capability map
The TILOS reference architecture describes a full governance loop. This demonstration build implements a small, honest slice of it. Each part of the architecture is marked with what is actually on this page.
Working on this page
Every cycle is computed on this page from embedded data. The same inputs always give the same outputs, and nothing is fetched.
Each input carries a REAL or PLACEHOLDER tag, so every result can be traced to the quality of what drove it.
Six of 35 behavioural axes are driven by signals, using the registry's own 11-point labels and colours.
Declared and operational signals are read, but currently averaged per family and compared with the family's own baseline rather than with each other.
Reference architecture v1.1 modules
Sets the values every module reads. Here the reference point is each signal's own history, not configured values.
Tiers decisions by reversibility, reach, horizon and risk. This page's bands are severity bins, not decision classes.
Provenance is labelled per signal; confidence is not yet computed or carried into the results.
Actors, incentives and influence. Would help explain which pressures produced a movement.
One fixed tolerance window of 0.15. Real constraints would widen or narrow it, telling pressure apart from drift.
Signals map to six registry axes, as one blended value per family rather than separate declared and operational states.
A weighted average and a signed delta. How values reinforce or strain one another is not yet computed.
Severity and two thresholds. Persistence tests and the distinction between drift and healthy adaptation are not yet on the page.
Memory of unresolved issues carried from one cycle into the next.
Pattern knowledge that widens how a result is interpreted for human decision-makers.
Structured scenarios that surface the values actually guiding decisions. The wheels on this page are a visualization, not M8.
Proposes and evaluates corrective responses under human authorization. This page classifies; it proposes and tests nothing.
Learning across domains from resolved patterns. Requires real deployment history.
Module names follow the published TILOS Public Reference Architecture v1.1. A separate computational core implementing more of these modules is in development and is not connected to this page.
06 · Two horizons
How can our decisions and actions better reflect what we have chosen to value?
Evaluate behaviour against declared values, detect drift over time, and inform responses that bring action back toward what was chosen. This page is an early, partial demonstration of the evaluation step.
PARTIALLY REPRESENTEDWhat becomes possible when we change what we optimize for?
Compare what different value configurations could produce before committing to one. This is a research direction subject to validation, and nothing on this page models it yet.
FUTURE RESEARCH07 · Pilots
Historical records show whether the computation responds to what is already known. The next step is to test it where governance actually happens: with organizations that choose to examine their own declared values against their own operational evidence, under their own authority.
For organizations that want to see how their stated commitments compare with what their operations show, over time.
For teams deploying AI systems that act for them, where what a system does must match what its operator has committed to.
Every pilot is designed around four safeguards:
Reference
All 35 behavioural axes across the seven value clusters currently in scope. Each cell shows the axis's positive-pole colour, its two-letter symbol, and its full name. Cells marked LIVE are the six axes actually driving the Living Organization visualization above from real or placeholder Wells Fargo data; the rest are defined in the registry but have no signal wired into this build yet.
Reference only: this table does not compute anything itself. Colours shown are each axis's value-pole (+1.0) colour from the Chakra Spectrum Registry's Master Registry tab; each axis is actually an 11-point spectrum running from its counter-value pole through to this colour, not a single fixed state.
08 · Technical notes
Each signal's raw yearly count is converted to a z-score against its baseline years (2012 to 2015 for Wells Fargo and Nucor; 2021 for Air Canada), then passed through tanh(k·z) with k = 0.5 onto the −1 to +1 scale the Integrity Delta uses.
Phrases are tagged aspirational (more mentions mean more of the declared value, such as "vision and values") or risk disclosure (more mentions mean an admitted problem, such as "sales practices" appearing in risk factors from 2017). Complaint and incident signals always run in the risk-disclosure direction. "Code of ethics" is down-weighted to 0.3 in the Truth family because of confirmed boilerplate and exhibit reuse.
This is a documented design default, not a validated instrument.
| Parameter | Value and status |
|---|---|
| k | 0.5. Steepness of the tanh scaling. Design default. |
| Tolerance window | 0.15. Movement within this distance of baseline is not counted as severity. Design default. |
| Band 3 threshold | Severity of 0.4 or more in any family. Design default. |
| Target | 0, meaning "matches its own baseline". |
| Baseline years | 2012 to 2015; Air Canada 2021. |
Declared-state phrase counts from annual 10-K filings (EDGAR) are real and verified for 2012 to 2022. The 2023 to 2025 extension, and one candidate phrase ("resolution language", tracking wording such as "termination of all outstanding consent orders" from the March 2026 shareholder letter), are unverified placeholders pending a live EDGAR check.
Operational-state complaint volumes (CFPB) are placeholders for the entire range, pending a data pull currently blocked by the CFPB site's anti-automation protections.
The external events are real and sourced. Both signals are illustrative placeholders: the declared safety series has not yet been checked against Nucor's 10-K filings, and the incident series is shaped around the two sourced fatalities (2016 and 2023) rather than drawn from OSHA establishment-level data.
Air Canada is a single, documented tribunal matter rather than a multi-year corporate pattern. Its two signals are yes-or-no states taken from the published decision: whether the published bereavement policy was accurate (it was, throughout), and whether a documented, unresolved chatbot misstatement existed that year.
Because these are binary states passing through a pipeline designed for counts, with a one-year baseline, the size of each reading reflects the method's defaults more than the event. The coding of 2023 and 2024 is awaiting confirmation against the record.
A technical audit of this build identified the following. Any change to the computation will be published as a new version, with this original computation kept available for comparison.
Results on this page are computations over public records and placeholders. They are not findings of misconduct, cause, intent or integrity about any organization or person, and they should not be used to make decisions about any of the companies shown.
The Integrity Layer acknowledges that its work is carried out on the unceded traditional territories of the Sḵwx̱wú7mesh (Squamish), xʷməθkʷəy̓əm (Musqueam), and səlilwətaɬ (Tsleil-Waututh) Nations.