Both are read from a document you submit, both are delivered to this domain,
and both run on the same engine. What differs is what they read, what they emit, and how much of
each has been measured. They are set side by side so the difference is visible rather than
asserted — including where one column is fuller than the other.
There is a second reason they sit together. Both already emit the same
class of structured output — a closed classification, an evidence trail, a pay figure with
its basis stated — and that shared shape is what a future comparison layer would read. Role ×
Profile, the planned layer that reads a person's structured profile directly against a role's
structured demand, does not exist yet. When it does, it will not re-read either document; it will
read the two structured outputs this page already documents. Nothing below should be taken as
evidence that layer is built — the open backlog on each side, and the "On the roadmap" section
further down, are what is actually true today.
Role Intel
€7
Measured 1 Oct 2026
A structured read of a role — what a job description is actually asking
for, and how to read it.
Closed vocabularies
Same test as the column beside this one: for each value list, does the
prompt lock it or merely list it. Role Intel answers worse than Profile Intel on
the first vocabulary, and that is the most important line on this page.
Role archetypes 7, defined — not runtime-validated
Builder · Operator · Specialist · Connector · Fixer · Translator · Anchor
Until 10 September 2026 the prompt carried the seven bare words and
no definitions; the definitions existed only as tooltip strings in the renderer, written after
the fact, and the model never saw them. That is no longer true. Each of the seven now carries
a written definition in the prompt itself, together with the evidence necessary for it to apply at
all, the evidence sufficient to assign it decisively, and named patterns that resemble it and are
not it. Job titles, seniority words, industry, culture language and tool lists are excluded as
evidence by rule. A role carries a dominant type and may carry a secondary.
The vocabulary is still not runtime-validated. The reviewer is
instructed to fail an off-vocabulary value as severe, and a legacy eighth value is coerced before
storage — but neither is a deterministic check in code, which is why this heading says
defined rather than closed. Whether the definitions make the assignment reproduce
is a separate question, measured separately, and it has not been measured yet.
Seniority ladder 5 producible — not validated
Senior IC · Manager · Senior Manager · Director · VP/Executive
These five are the rungs the pay engine can actually price and label.
Three problems sit behind that sentence. The free-text matcher accepts seventeen tokens and
maps them onto those five. ic_junior is priced but has no band label, so a
junior role is captioned with the senior band's name, and it is absent from the ladder, so it can
never be scope-adjusted. And ic_mid, c_level and head_of are
referenced by three consumers but producible by none.
Occupational family · geography closed
Families resolve through an alias map to a canonical set. Geography
resolves to one of ten bands: four Romanian cities plus RO_other, and
EU_west, EU_cee, EU_south, EU_nordic_ch,
EU_unspecified. Pay is held per family × geography × rung.
Pay confidence closed
Calibrated or Estimated, with
Disclosed where the employer published a band. Any inferred seniority forces
Estimated.
This page claimed that on 1 September and the claim was not fully true until
7 September. Seniority is resolved four ways. The 1 September change caught two of them. It
missed the third: where a job title carries an ambiguous token — lead,
manager, principal, staff — the engine deliberately declines to
trust the title and takes the level from a model-generated signal instead, and those briefs kept
printing as Calibrated. The label now follows the actual basis in all four cases,
shipped 7 September 2026. More published figures read Estimated than did before,
and that is the correct direction. Briefs delivered between 1 and 7 September that carried an
ambiguous title may show Calibrated where they should read Estimated;
the figure itself is unchanged, only its confidence label. If you hold one and want it re-issued,
ask.
Two vocabularies are not closed, and one of them is not a vocabulary at all.
gap_type — the label naming what a role is missing — has no fixed
list: two runs of the same input produced eight distinct labels with one overlap. And
role_seniority_chip has no runtime validator anywhere in the pipeline; the
sister product has been observed emitting C-level, which is off-vocabulary, and it
reached a pay band regardless. Both are open. Recorded 2 September 2026.
What is comparable today
So the honest claim is that two Role Intel briefs are comparable on
money and on the classifications that set it, and not on the single word that names the role's
shape. That word is the most quotable thing on the brief and it is the least reproducible.
What is confirmed, and what is inferred
Everything comes from the job description submitted. Nothing is looked
up: the posting is not fetched, the employer is not researched, no listing is retrieved.
Confirmed from the document
- Stated title, location, work model
- Stated responsibilities and requirements
- A published salary range, where the JD gives one
- Stated team size, reporting line, budget
- Named tools, stacks and certifications
Inferred by the system
- Role archetype and secondary archetype
- True seniority and scope versus the title
- JD-inflation signals and severity
- Occupational family and geography band
- Pay figures, which follow from those two
- Transferability and talent-pool depth
These are readings of a posting,
not facts about the employer.
What the system does not do
- It does not fetch the job posting. A brief that claims it retrieved the JD is blocked before
delivery — that is a named gate rule.
- It does not evaluate a person, and it never sees one. A Role Intel brief is a read of a
document about a job.
- It does not tell you whether to apply. It has no verdict field of that kind.
- It does not price above a figure the employer published. The internal table may never outbid
a published ceiling — also a gate rule.
The delivery gate
Deterministic code inside the renderer, not a model. Nineteen
numbered rules, DG-01 to DG-19, gate version 1.8. Each
finding carries a rule id, a severity, a message and a machine-readable evidence object. Any
finding at BLOCK stops delivery; WARN findings are recorded and the brief
ships.
Why 1.8, briefly: 1.7 could reject its own correct output. One
rule demanded a verdict — under-priced, over-priced, or aligned — on every brief, including ones
priced against a job description that published its own salary grid, where no verdict was ever
computable. 1.8 tells the two cases apart before asking for a verdict. The incident that surfaced
it is below, dated.
Updated 5 October 2026. The paragraph and table above describe gate
1.8. The live gate is 1.14 and has twenty-one rules. DG-20
(gate 1.12, 4 Oct) blocks delivery when a posting that is not employment renders a salary figure.
DG-21 (gate 1.13, 5 Oct) warns, and never blocks, when the brief's written commentary
contradicts the computed comparison with a published pay range. Gate 1.14 (5 Oct) fixed
DG-17: it searched for a piece of stylesheet text rather than a rendered pay figure,
and in one tested case blocked a correct brief. Versions 1.9 to 1.11 are recorded in the
renderer's own change notes and are not itemised here yet.
The gate reports everything it runs — on briefs that ship. On blocked briefs it reported
almost nothing, until today. A delivered brief carries
delivery_gate_findings: the full serialised list of every finding, with ids,
severities and evidence, plus a pass flag and block and warn counts. But a blocked brief
throws, and the throw replaces the output — so the findings were lost precisely when they
mattered. On 1 September a real brief was blocked and the stored reason was 77 characters of
the tail of one sentence, with every rule id cut off; identifying the rule took a
reconstruction from the payload. Two fixes have shipped: the message is now a single line, and the
rule ids are appended at the end, where the truncation keeps them. Recorded 2 September
2026.
Engagement type, published-range comparison, ESCO
Three additions from 4 and 5 October 2026. Each is code around the
model's reading, not a new model judgement.
Engagement type 6, closed · since 4 Oct 2026
Employment · Contract / fractional · Board / advisory seat · Fellowship / programme · Investment call · Volunteer role
Every posting is classified before it is priced. Only employment is
priced against employment salaries. Anything else gets an economics panel and no salary figure,
with a fixed line saying why: it is not employment, ReLoop.me does not compare it with employment
salaries, and calibrated market data for that engagement type is not available yet. Gate rule
DG-20 blocks delivery if a figure renders anyway, and these postings are kept out of
the pay index. On the stored test set, 16 of 16 employment briefs came out unchanged and 44 of 44
non-employment briefs carried no figure. Profile Intel describes engagement with different
words today (salaried, fractional, founder with equity); the two lists are not yet
reconciled.
Comparison with a published range since 5 Oct 2026
When a posting publishes a pay range, the brief states where that range
sits against ReLoop.me's own estimate for the scope the posting describes: below, at the
low end, within or above. The comparison is made at the band edges, in gross
annual terms, by code. A typical line: “The published range sits below ReLoop.me's
estimate for this scope (…): even its ceiling (€XK) is under the band's floor
(€YK).” It compares an estimate with a published figure; it is not a market fact
and not an apply-or-pass verdict. No line is shown when no range is published, when only a floor
is published below the band's top, when the location cannot be anchored, or when pay is
equity-led. DG-21 flags a brief whose written commentary contradicts the computed
line.
The title now caps the model's seniority reading (salary
3.65, 5 Oct). A title containing “manager” is priced no higher than
Manager, and one containing “lead” or “leader” no higher than Senior
Manager, when the level comes from the model rather than the title. The cap only ever lowers a
level, and the brief says so in one line. Where the posting names no team, the brief prices the
work described rather than the title.
Found and fixed the same day, 5 October 2026. A live brief told its reader
a published range sat below the estimate, while the brief's own commentary called it
competitive. Three faults stacked: the model read the role a level above its title; a detector
trusted a category label written by the calibrator model as if the employer had written it; and
“product launch” in a consumer-goods advert rerouted the role to product marketing.
All three were fixed in salary 3.65. Re-run on the same advert, the brief now reads
within, matching the prediction written down before the fix. On roughly 40 recent briefs,
7 had been priced from a model reading above a plain title; about one recent brief in six would
now price lower.
Occupation code ESCO · exact matches only · since 5 Oct 2026
A brief can carry the role's occupation in ESCO, the European
Commission's classification of skills, competences and occupations (v1.2.1), linked to its ESCO
page. It is shown only when the job title matches a mapping a person has checked and marked an
exact match; closest-family matches are stored and not shown. The code is read from that
table with no model call, so the same title always gets the same code. Most titles have no
checked mapping yet, so most briefs carry no code. Requirement terms are looked up and
stored, not shown. How the mapping is built and kept current is under the common engine.
Determinism — what we measured
Same intent as the column beside this one, same answer in shape, but
the split falls in a different place: Role Intel's money reproduces and Role Intel's headline
word does not.
Identical input · pay path
Two runs of the same job description on the same day reproduced the pay
figures exactly. A third run, captured separately as a fixture on live code, reproduced an
earlier brief's figures to the euro. A fourth case was checked by recomputing the arithmetic
by hand from the published band: likely offer at the lower third of a
€89,700–€152,300 grid gives €110,566.7, rounded to €110,550, and
the ask at the grid maximum gives €152,300. Both matched the delivered figures
digit for digit.
The pay engine is deterministic code and behaves
like it. Where pay has moved between runs, the cause has been a classification changing
upstream, not the table.
Identical input · the archetype
Sixteen runs of byte-identical input returned Builder six times,
Operator six times, Fixer four times — near-uniform across three of the seven values.
Temperature was tested at three settings including 0; pinning it made the spread worse,
not deterministic.
Separately, six real job descriptions were each submitted twice by
accident, hours apart, byte-identical. The dominant archetype reproduced in 3 of 5 comparable
pairs, the secondary in 1 of 6. In one case a single submission was delivered twice from
literally the same bytes and came back Operator/Translator once and
Builder/Connector the other.
The cause was known and it was not mysterious: the seven
archetypes were handed to the model as a bare list with no definitions. That was the cheapest
experiment on the backlog, and on 10 September 2026 it was run — the seven now carry written
definitions, necessary and sufficient evidence tests, named false positives, and a stated order of
resolution rather than a best-fit judgement.
Those figures have now been replaced by repeat
runs. The paragraph that used to sit here promised numbers rather than more description.
They are in the card below.
Identical input, four runs · 23 Sep 2026
A real, already-delivered job description. Reproduced offline: no live
form submission. The live Request Builder node code was re-fetched from the workflow, and the two
clauses deployed that morning were located in the captured system prompt to confirm the prompt
under test is today's. The real user message was pulled from execution history straight to disk
and never hand-retyped. Four independent, blind completions against that prompt and that
message.
Seven fields of twelve reproduced across all four runs. Five
moved.
The archetype pair is stable as a set and unstable as an order.
All four runs returned the same two types — Translator and Operator — and all four
disagreed about which of them the role is accountable for, two each way, with the secondary the
exact mirror of the dominant every time. That is a sharper result than the sixteen-run spread
above: the rubric appears to have settled which types are in play without settling
which one is the answer. The step that assigns accountability is where the remaining
instability sits.
The pay engine is still deterministic. The reading handed to it is
not. Three runs read the role as Senior Manager and one as Director, with nothing else about
the document changed. On the live table those two bands, for this occupational family and this
region, are €3,650 and €4,700 net per month — a 29% swing in the delivered
figure, roughly €73,000 against €94,000 gross a year, on identical input. The
arithmetic below the classification reproduces to the euro; the classification above it does not.
The brief actually delivered from this document carried the Director reading — the
one that came up once in four.
Caveats, both limiting. One
document, four runs, is not a rate — it is one document on one day, and the two Profile
Intel studies on this page show that which field is unstable depends on the document. And the
harness is not the production caller: the live Role Intel request sets no temperature
parameter at all, taking the API default, and these four runs were independent completions at
whatever sampling the harness applied. Directionally informative about reproducibility, not a
claim of identical decoding.
What the 10 September change altered
The change also alters some outputs: where a
posting’s accountability and its dominant method point to different types, the dominant type
may now differ from what the same posting returned before 10 September. The direction is intended;
the size of the shift is unmeasured. A further clause narrowing one of the seven types shipped on
23 September, hours before the runs above; it is present in the prompt they were run against, and
is itself unverified until a widened set is graded.
A question for whoever maintains the Profile Intel column. The changelog below records an
earlier sixteen-input reproducibility study as unrecorded because its date and fields cannot be
sourced. The sixteen-run study above is a Role Intel archetype study. These may be the
same study, filed against the wrong product. We are flagging the possibility rather than merging
them, because guessing would be exactly the failure the log exists to prevent.
Repeat runs since 23 September · added 5 Oct 2026
1 October, thirty runs of one job description on production
settings (temperature 0.3), set beside two general models given a simple prompt
(gpt-5-mini, gemini-3-flash-preview, thirty runs each). Role type
Builder 30 of 30. Seniority Senior Manager 22, Senior 8, so the same posting can be priced on two
rungs. Capability names were close to unique: 124 distinct names in 140 slots. Capability themes
overlapped from run to run at 0.66, against 0.73 and 0.76 for the general models. So Role
Intel is not more repeatable than a general model on capability wording, and this page does not
claim it is.
1 October, title test, eight runs per arm. The same posting body
under two titles, “Head of Growth” and “Growth Marketing Specialist”. Role
Intel read the seniority as Mid-Senior 8 of 8 under both titles; both general models moved with
the title. What differs is judgement, not repeatability: it reads the scope the posting
describes, not the title.
Since then: on 2 October the generation call was
set explicitly to temperature 0. On 5 October the title cap above came in: for the posting in the
four-run card, titled “AI Practice Lead”, a Director reading is now priced as Senior
Manager, so the 29% swing shown there can no longer reach the delivered figure for titles of that
kind. Neither change has been re-measured with repeat runs yet. One job description per test is
not a rate.
What we claim, precisely
- The generation call runs at temperature 0.3, set explicitly at the request, and the
reviewer at 0.3 as well.
- Seniority band, family, geography and every pay figure reproduced exactly across the
identical-input cases above. Evidence from a handful of documents, not a guarantee.
- The role archetype does not reproduce, at 3 of 5 and 1 of 6, on samples that are small
but consistent, and on a mechanism we can name.
- Whether the generation call ever ran without an explicit temperature is not established.
The retained workflow history reaches back only to 1 September 2026 — four versions, all from
that day — so the question cannot be answered from it. It is not answered by assumption
either.
Narrowed 5 October 2026. The first item above was not reliably true: on
23 September the live request carried no temperature (see the caveat in the four-run card), and
which setting applied between then and 2 October is not settled. Since 2 October the generation
call runs at 0, read back from the live request; the calibrator and the reviewer stay at
0.3. The second item held on the cases it lists and not on the later runs. The narrower claim:
the pay figure reproduces exactly whenever the seniority, family and geography readings do; the
seniority reading itself does not always reproduce.
EU AI Act posture
Not analysed, and not inherited. The Profile Intel analysis turns
on a brief being commissioned by the person it describes. A Role Intel brief describes a
job, is commissioned by a candidate reading that job, and evaluates no natural person at
any point — a structurally different question that has not been worked through. The mechanism
set out under the common engine below is stated there because it is the same whoever reads it;
where the line falls for Role Intel specifically is open.
Submitted material and retention
Not established. Whether Role Intel makes the same retention
promise as Profile Intel, and whether a mechanism stands behind it, has not been audited. Stating
it before checking is how the Profile Intel promise ran ahead of its mechanism, and the same
mistake is not worth repeating on this side of the page.
Delivery identity
A Role Intel brief declares product = "Role Intel", and that
label appears in the brief's provenance paragraph and in its structured page data. Whether the two
products are visually distinguishable to a reader who lands on one has not been audited.
Open backlog
- Role archetype does not reproduce. 2 Sep 2026. Half closed 10 Sep 2026 —
the seven values are no longer undefined in the prompt. Repeat runs, 23 Sep 2026: four
blind runs of one real posting returned the same two candidate types every time and split
two–two on which of them is dominant, the secondary mirroring the dominant each time.
The rubric settled which types are in play; it did not settle which one is the answer.
The item stays open, now with a number and a narrower target: the instability is in the
accountability step, not in the vocabulary.
- Gap type has no closed vocabulary — 8 labels across 2 runs. 2 Sep 2026.
- No runtime validator on the seniority chip, in either product. 2 Sep 2026.
- Thirteen non-monotonic salary rows: Senior Manager prices above Director in two
families, so a scope adjustment upward lowers the band while the brief says it was
adjusted for scope. 2 Sep 2026.
ic_junior has no band label and cannot be scope-adjusted; a junior role is
priced correctly and captioned with the senior band's name. 2 Sep 2026.- Three phantom rungs are referenced by consumers and producible by none. 2 Sep 2026.
- Pay figures are not stored, so the database cannot answer what a given brief told a
customer about money. 2 Sep 2026.
- Romanian briefs leak English interface strings — 59 sites inventoried, 46 of them
reaching the page. The locale itself resolves correctly; these are untranslated constants.
2 Sep 2026.
- The scope-adjustment direction is emitted and never read by the renderer, so a brief can
argue a title is inflated and then advise as though it were understated. 2 Sep 2026.
- The engine version stamped into the brief index is four versions stale. 2 Sep 2026.
- The prompt build is not recorded with the brief. The stored version string does not
name it, and three prompt builds since 7 September all write the same string, so a brief can be
attributed to a build only by its timestamp. 10 Sep 2026.
- The seniority chip flips on some postings: Senior Manager 22, Senior 8 across thirty
identical runs, and the chip sets the pay rung. Measured at temperature 0.3; not re-measured at 0.
1 Oct 2026.
- Capability names are close to unique from run to run (124 distinct in 140 slots), which
blocks any comparison that joins on capability names. 1 Oct 2026.
- Engagement-type words differ between the two products. 5 Oct 2026.
Profile Intel
€19
Measured 1 Oct 2026
A structured capability read of one person, from their CV, commissioned by
that person.
Closed vocabularies
A closed vocabulary means the system may only emit values from a
fixed list. It is what lets two briefs be compared. Where a vocabulary is not closed,
this page says so.
Profile types 7, closed
Builder · Operator · Specialist · Connector · Fixer · Translator · Anchor
The prompt states these must be exactly this set. A profile
carries a primary type and may carry a secondary. The type is separate from the career
archetype — a free-text positioning label capped at four words, which may never be one of
the seven types.
Career patterns 7, closed
Linear · Progression · Steady · Lateral · Crossover · Multi-track · Multi-context
A profile carries a dominant pattern and, where the shape is dual, a
foundation pattern. The dominant is always the more distinctive of the two, never the more
common one.
Capability ladder 4 bands, closed
Every brief carries four to six capability clusters. Each gets a band,
an evidence confidence and a recency class. The band is not a judgement of the person; it is
a statement about how much of the CV supports the claim.
Evidence confidence
| High | 3+ anchors, 2+ roles |
| Med | 2 anchors, or one role |
| Thin | 1 anchor, or implied |
Recency
| Current | within 2 years |
| Recent | 2–6 years |
| Historic | 7+ years |
The recency gate caps the top of the ladder, and only the top.
If the newest supporting evidence for a capability is seven or more years old and has not been
refreshed by recent work, that capability caps at Expert rather than Distinctive. It is never
dropped below Expert on recency alone. A rare credential stays a rare credential even when it is
old — currency of a capability and rarity of a credential are different axes, and the system does
not collapse them.
Legacy labels. Briefs generated before the ladder was renamed
carry Solid, Developing and Basic. These map to
Demonstrated, Emerging and Emerging. They are not emitted in new output.
One vocabulary is not closed, and you should know which. The management-tier scale used in
the experience timeline — IC / Team Lead / Manager / Senior Manager / Director / VP / C-level
— is specified as a list but not locked the way the vocabularies above are, and its seven values
do not map one-to-one onto the six-rung scale the pay engine uses. A fix is specified and not yet
shipped. Recorded 1 September 2026; see the changelog.
What is comparable today
Comparability rests on the closed value lists, not on the prose being
identical. So it holds for exactly the fields whose vocabularies are closed and which
have been shown to reproduce:
So the honest form of the claim is narrower than "briefs are
comparable". Two briefs can be compared on profile type, career pattern, capability band,
seniority, family, geography and the pay figures. They cannot currently be compared on management
tier or peak direct reports. Those two are named in the changelog as open, and this table is what
the claim is reduced to until they close.
Read the geography row with the backlog open.
“Comparable” here means two briefs use the same value consistently. It does not mean the
value is right for the market. A defect found 2 September 2026 puts Southern-European, Nordic and
Swiss profiles on a neighbouring band — reproducibly, and printed as calibrated. The pay rows
inherit it. See the open backlog below.
What is confirmed, and what is inferred
Everything on a brief comes from the submitted CV. Nothing is looked
up about the person, and no external record is consulted. But not everything on a brief has the
same standing.
Confirmed from the document
- Employers, role titles, stated dates
- Durations computed from those dates
- Stated skills, tools, named credentials
- Stated figures, as the CV states them
- Education and certifications as listed
Every evidence anchor must trace
to specific content in the CV. An empty anchor list is a correct output; an invented one is a
failure.
Inferred by the system
- Profile type and career pattern
- Capability clusters, bands, confidence, recency
- Seniority band and occupational family
- Salary bands, which follow from those two
- Rarity anchors and the misread-risk audit
- Trajectory, fit and direction
These are readings of a
document, not verified facts about a person.
What the system does not do
- It does not verify that a claim in the CV is true. It audits how well the CV
supports the claim, and marks each one substantiated, partial, weak, unverifiable or
unsupported.
- It does not look the person up. No search, no public records, no social profiles.
- It does not rank one person against another. A brief is commissioned by, and about, one
person.
- An impressive company-scale number the audit cannot substantiate is treated as context,
never as a lift to a band, a confidence, a level or a trajectory.
- Since 4 October 2026, nothing typed into the order form moves a figure or a band.
Answers to the form's questions steer what the read pays attention to. A self-estimated salary
is no longer shown to the model at all; it is compared with the read only after the read is
done.
The delivery gate
Before a brief is delivered, a separate reviewer model checks it
against numbered rules and returns pass, warn or fail on each. The gate owns the final decision.
There are 29 numbered phases — eleven K phases and sixteen L phases, plus a truncation
check and a language check. Six L phases carry lettered sub-checks. There is no K8.
The gate reports a subset of what it runs. All 29 phases execute, but the machine-readable
result carries fifteen entries: thirteen rules with a pass / warn / fail, plus two warn-only
sub-checks (seniority anchor, tenure coherence). The remaining sixteen phases act on the output
and raise warnings, but do not surface a per-phase verdict. We are stating this rather than
publishing a rule count that sounds larger than what is reported. Recorded 1 September 2026.
A brief is delivered on pass_validated or
compact_validated. On fail_validated it is held and not sent. A held
brief is a working gate, not a lost order — it means the reviewer found something the read could
not stand on.
Determinism — what we measured
The design intent is that the same CV produces the same brief. That
is not what we measured, so it is not what we claim.
Identical input, two runs · 1 Sep 2026
The classifications that set the salary reproduced
exactly. The management-scope reading did not.
Identical input, two runs · 23 Sep 2026
A different real, already-delivered candidate document from the one
used on 1 September — not a re-run of the same input. Reproduced offline: no live Tally
submission. The exact live CIA Request Builder and CIA Prompt Patch (REGRESSION-78) node code was
re-fetched from the workflow and executed standalone to reconstruct today's deployed system
prompt; the real user message was pulled from n8n execution history straight to disk, never
hand-retyped. Two independent, blind runs against that reconstructed prompt and the real user
message.
This run moved on the field the 1 September run held stable —
seniority band swung a full band (senior_manager → manager) — and the salary figures moved
with it, roughly a 29–38% swing on the same document. Peak direct reports, which moved on
the 1 September document, held stable on this one. Two single-document measurements six weeks
apart do not average into a rate; they show that which field is unstable depends on the document,
not on a fixed property of the field.
Caveat: the harness used to run this
comparison cannot pin the API's temperature parameter the way the live call does (temperature
0.3). The two runs above are independent completions at whatever sampling the runtime applied, so
this is directionally informative about reproducibility, not a claim of the exact same decoding
setting as production.
Eighteen characters changed · 1 Sep 2026
The same document with an eighteen-character phrase deleted from one
bullet. Nine of twelve timeline rows were identical. Three moved, and the peak-level reading
moved with them:
Peak management level for the whole profile: Director in one run,
Manager in the other — a two-band swing on the same document. The claims audit also differed:
twelve claims listed in one run, thirteen in the other, and the one claim marked unsupported in
the first was not audited at all in the second. That single difference propagated to the
inflation-risk label and to whether a whole block appeared on the page.
The prediction under test — that the phrase was setting
the seniority band — was wrong. The band did not move; the management tier did. We are
recording the failed prediction because it is the result.
Repeat runs, 1–2 October 2026 · added 5 Oct 2026
Thirty runs of one real CV on production settings, set beside two
general models given a simple prompt (thirty runs each). Seniority 30 of 30, profile type
Builder 30 of 30. Capability themes overlapped from run to run at 0.69, against 0.79 and 0.83 for
the general models, so Profile Intel is not more repeatable on capability wording either.
The headline capability took 14 different values in 30 runs, and one capability sat on the
Expert/Distinctive boundary, 17 against 13.
Echo test, eight runs per arm, synthetic CV. The same CV with and
without a header in which the person rates their own capabilities. The self-rated capabilities
appeared in 0 of 8 Profile Intel runs, against 4 to 8 of 8 for the general models; Profile
Intel marked the self-labelled version down for inflation instead. The CV used in the thirty-run
test labels its own “distinctive edge”: the general models rated that capability
Distinctive in 31 of 32 and 29 of 31 slots, and Profile Intel rated it Demonstrated in 22 of 30.
2 October: the evidence-quality score moved from
the model into code (object transformer 1.15.0). Ten runs of one CV then returned
the same score, 58, every time, where it had ranged from 62 to 74.
What we claim, precisely
- Generation runs at temperature 0.3. It is a language model reading a document, and it does
not repeat itself exactly.
- Two identical-input studies now sit on this page, six weeks apart, on two different real
documents. Read field by field rather than as two separate verdicts: occupational family,
geography and the inflation-risk label were stable on both documents. Two runs each — that is
evidence, not a guarantee.
- Market-value figure and the coordination-scope label were unstable on both documents.
Named, dated, and open on both studies.
- Seniority band, the likely-offer figure and peak direct reports flipped between the two
documents — stable on one, unstable on the other. That is the most important reading on this
page: instability is document-dependent, not a fixed property of a field. A field reproducing
cleanly on one CV is not evidence it will on the next one.
- No field on this page has been observed more than twice. "Stable" above means stable
across two runs on one document, not a measured rate — the same field could still move on a third
run or a different document.
- Both studies were run after 28 August 2026, so they describe the engine as it is
configured today. They are not a measurement of the briefs produced before that date, which ran
at a materially higher temperature — see the changelog.
- Comparability rests on the closed value lists, and the table above names which fields that
covers today.
EU AI Act posture
A Profile Intel brief is a capability overview commissioned by the
person it describes, and in that use it is not a high-risk AI system. Using one to evaluate,
screen, rank or compare someone else against a role is a different purpose and may fall
within high-risk classification. The mechanism, and the date it applies, is set out under the
common engine below.
Open backlog
- Management-tier vocabulary is not closed. Fix specified, not shipped. 1 Sep 2026.
- Peak direct reports does not always reproduce, and it is load-bearing on the pay figure
via a peak-management premium. Moved (1 vs 4) on the 1 Sep 2026 document; held stable (5 vs 5) on
a second, different document measured 23 Sep 2026. Document-dependent — not resolved.
- The claims audit varies in selection, not rating — which moves the inflation label and
whether a whole pay block renders. 1 Sep 2026.
- An earlier sixteen-input reproducibility study is unrecorded because its date and
fields cannot yet be sourced. 1 Sep 2026. Checked against Role Intel's own sixteen-run study,
23 Sep 2026, and confirmed not the same study — see the changelog. Still open, still
unsourced.
- Suppressed-pay wording renders in English on Romanian briefs. 2 Sep 2026.
Closed 5 Oct 2026: Romanian briefs now always use the Romanian sentence (renderer
16.92).
- Southern-European, Nordic and Swiss profiles are priced on the wrong band. The pay
engine holds separate bands for those regions and resolves the city correctly; the classification
step cannot emit them, so a profile is forced onto a neighbouring band and the figure is printed
as calibrated. Evidenced on a Rome profile priced on Central and Eastern European bands.
2 Sep 2026. Closed, recorded 5 Oct 2026: fixed by a geography override in salary
2.50/2.51; the fix shipped before 23 September and this page never
recorded it.
- The engine version stamped into a brief does not match this page. Briefs were seen
carrying
cia=3.20.0 against a single 3.27.2 figure this page stated on
2 Sep 2026. Status 23 Sep 2026: this page no longer publishes one overall number to
compare against — the header above now lists five components read separately off the live
workflow. Root cause now traced. The CIA prompt never defines a
_metadata.cia_version field — zero occurrences in the decoded system prompt.
The Quality Reviewer prompt tells its model to read that field anyway; asked to copy a value that
is not on the page in front of it, the model supplies a plausible one, most often its own version
number, which is why so many briefs carry the Quality Reviewer's changelog numbers rather than the
generation engine's. Demonstrated live 4 Sep 2026: three briefs of one person on an identical
engine carried cia=3.20.0, cia=3.20.0+ and cia=3.27.2.
Traced, not fixed — the field the model is asked to copy still does not exist.
Closed, recorded 5 Oct 2026: fixed by prompt patch 1.3.0 and object
transformer 1.16.1 (3 Oct). A brief generated on 4 October carries
cia=3.35.26, the build it actually ran on.
- One withdrawn brief identifier is a placeholder pattern, not a generated identifier.
It appears in the withdrawn-brief list as a string of literal X characters. Either it was never a
delivered brief and the list is one entry long, or a placeholder reached a delivered page —
which would be a defect in identifier generation rather than a dead page. Not yet established.
10 Sep 2026. Closed 5 Oct 2026: no Profile Intel record carries that identifier or a link
to it, so it was never a delivered brief.
- The language lock can be overridden by the model on an English-jargon-dense CV, even when
everything upstream is correct. Found live 23 Sep 2026: a Romanian CV with heavy English
professional terminology (job titles, functional-area names) got a deterministic, correctly-locked
detected_language="ro" at 0.90 confidence, and an explicit instruction naming every
narrative field the lock covers — the model still wrote the entire brief in English (0.95
confidence, 926 markers across all 181 narrative strings; not a marginal slip). The G-5 merge
guard, added 11 Sep 2026 after a different language-lock incident, caught it and held the brief
undelivered — that is the guard working as designed, not a new failure, and nothing was sent
to the candidate on the wrong-language brief. A revised instruction, naming English job-title and
industry-terminology density explicitly as vocabulary rather than a language signal, was tested
offline against the real held document: two independent runs, both fully Romanian (roughly 1,000
Romanian markers against under 60 English markers each, no individual field scoring predominantly
English). Prepared as a versioned patch. Not yet deployed. Status 5 Oct 2026: deployed
in input builder 1.8 on 23 September; not yet observed on a second jargon-dense
Romanian CV, so it stays open.
- The headline capability is unstable: 14 different values in 30 runs of one CV.
1 Oct 2026.
- The top band boundary flips: one capability Expert 17, Distinctive 13 across 30 runs.
1 Oct 2026.
Both products run on this. It is stated once here rather than twice above,
because a claim written in two places is a claim that will eventually disagree with itself.
A brief commissioned by the person it describes is not, in that use, a high-risk AI
system. Using one to evaluate, screen, rank or compare someone else against a role is a
different purpose, and may fall within high-risk classification under Annex III point 4(a) of
Regulation (EU) 2024/1689 — AI systems intended for the recruitment or selection of natural persons,
in particular to evaluate candidates — via Article 6(2). Article 25(1)(c) is the route by which
provider obligations transfer: a party who modifies the intended purpose of a system so that it
becomes high-risk is considered its provider and takes on the obligations in Article 16.
Those obligations apply from 2 December 2027, deferred from 2 August 2026 by Regulation (EU)
2026/1744 — so this describes where the line will fall, not an obligation in force today. It is a
statement of scope, not legal advice.
This states the mechanism, which is the
same whoever reads it. How it applies to Role Intel specifically has not been analysed —
Role Intel reads a role rather than a person, which is a different question and does not inherit the
Profile Intel answer.
Not defects — features that don't exist yet, named here on purpose so they
don't quietly drop off the list. Dated when work starts. Moved to the changelog, and off this list,
when they ship.
Enough to check a claim on a brief against a stated definition, and to see
the shape and pace of the work. Not enough to rebuild the engine.
The line is drawn where it is on purpose: what's shown is
what a reader needs to check a claim on their own brief. What's kept is what would let someone
reproduce the read without doing the work of building it.
Claims added, changed and withdrawn. Newest first. Every entry carries a date
and a reason. A withdrawn claim stays on this page — removing it would defeat the point of the page.
5 October 2026
ChangedA job title now caps how senior the model may read the roleRole
Titles with “manager” price at most as Manager, and titles with “lead” or “leader” at most as Senior Manager, when the level comes from the model's reading. Two family detectors were also tightened: a category label written by the calibrator model no longer counts as evidence for marketing technology unless the advert names such a tool, and “product launch” alone in a consumer-goods advert no longer reroutes to product marketing. Found on a live brief that called a published range “below” while its own commentary called it competitive; re-run, it reads “within”, as predicted before the fix. Regression set: 61 of 72 stored inputs unchanged, the 11 changes all on that advert. Salary 3.65.
5 October 2026
ChangedA delivery check that could block a correct brief was fixedRole
DG-17 looked for a piece of stylesheet text instead of a rendered pay figure; in a tested case it blocked a correct brief. It now looks for the rendered element. 66 stored inputs unchanged. Renderer 27.9.67, gate 1.14.
5 October 2026
AddedRole Intel briefs can show an ESCO occupation codeRole
Shown under the location line, linked to the ESCO page, only when the job title matches a row a person has checked and marked an exact match. The footer then adds: “This service uses the ESCO classification of the European Commission.” Lookup by table, no model call. In testing the same job description returned the same code three times out of three. Requirement terms are looked up and stored, not shown. Renderer 27.9.66, ESCO terms 1.0.0.
5 October 2026
AddedPublished pay ranges are compared with ReLoop.me's estimateRole
Below, at the low end, within or above, compared at band edges by code; no line when no range is published, when only a low floor is published, when the location cannot be anchored, or when pay is equity-led. Where a posting names no team, the brief prices the work described rather than the title, and says so. New warning rule DG-21. Salary 3.64, renderer 27.9.66, gate 1.13.
5 October 2026
ChangedRomanian Profile Intel briefs: pay wording fully in RomanianProfile
The suppressed-pay sentence and the pay-guard labels now render in Romanian on Romanian briefs, closing an item open since 2 September. The near-case line now reads “The CV evidence supports the same level”. English briefs unchanged. Renderer 16.92.
5 October 2026
ChangedOne marketing pay row aligned between the two enginesBoth
Marketing, Bucharest, senior individual contributor: 850/1,100/1,500 to 950/1,250/1,700 net EUR a month, the value Role Intel adopted the day before (one publisher, 2025). The other Romanian city rows for marketing are still ratios of the old Bucharest row. Profile Intel salary 2.77.
5 October 2026
AddedProfile Intel looks up ESCO codes for job titles, without showing themProfile
Experience titles are looked up in the checked table and the result is stored with the record. Nothing on the brief changes. It turns on when three briefs in a row each carry a validated exact code. Before building it we checked three runs of one CV against the table: 0 matches, and the capability names overlapped in only 3 of 47 between two runs of the same CV. That is why titles go first and capability terms wait.
5 October 2026
ChangedFour Profile Intel backlog items closedProfile
Suppressed-pay wording on Romanian briefs (fixed today). The Southern-European, Nordic and Swiss band defect (fixed in salary 2.50/2.51 and never recorded here). The cia= version stamp (fixed 3 Oct). The placeholder brief identifier (never a delivered brief). Each stays on the list, marked closed with its evidence.
4 October 2026
AddedA checked ESCO mapping table and a weekly update loopBoth
Terms are mapped to ESCO at two levels, exact and related. Two models from two vendors propose; a person validates; only validated rows reach a brief. Measured against the reviewer on 62 terms, the model panel's best code was in her accepted set 77% of the time. Details under the common engine.
4 October 2026
AddedEngagement type is classified before pricingRole
Six types. Only employment gets a salary figure; the others get an economics panel and a fixed line saying why. New blocking rule DG-20; these postings are kept out of the pay index. Object transformer 2.4.8, renderer 27.9.64, gate 1.12, index emit 0.3.3.
4 October 2026
ChangedForm answers and a self-estimated salary no longer move any figureProfile
Form answers steer what the read looks at. A self-estimate is removed from the model's input and compared with the read only afterwards. A new “second thread” row shows an alternative trajectory side by side, and the contract-rate block renders for the first time. Prompt patch 1.4.0, input builder 1.9, renderer 16.91.
4 October 2026
ChangedPraise for the employer is removed from Role Intel briefs by codeRole
Phrases such as “reputable employer” are a judgement about a company the brief has not researched. Removed deterministically. On 55 stored runs: 14 briefs touched, 21 adjectives and 5 sentences removed, no false removals found. Object transformer 2.4.7.
3–4 October 2026
ChangedUK pay priced by region, quarterly cells updated, the two engines alignedBoth
UK roles price on London unless a named place outside London is given, which applies a factor of 0.88 or 0.80. The quarterly pay check (1,017 dated points from 45 publishers) updated a set of cells. Profile Intel then took Role Intel's values for the shared cells, including 45 unsourced cells aligned down. Role Intel salary 3.63; Profile Intel salary 2.72 to 2.76.
3 October 2026
ChangedProfile Intel brief footer states crawlability and CV deletionProfile
“Your brief, your link. Crawlable by design — share it, or ask hello@reloop.me to remove it.” And: the CV and the text extracted from it are deleted within three months of delivery. The word “Confidential” was removed from that footer because the brief is not. Renderer 16.90.
3 October 2026
ChangedA no-team “Senior Manager” is shown and priced as SeniorRole
Where a posting gives the title no team, the chip reads Senior and the rung is priced as a senior individual contributor. Sentences promising a later calibration were removed from briefs. Object transformer 2.4.6, salary 3.61.
2 October 2026
ChangedRole Intel generation temperature set to 0Role
Set explicitly on the request and read back from the live workflow. The calibrator and reviewer stay at 0.3. Not yet re-measured; see the repeat-run card in the Role Intel column.
2 October 2026
ChangedEvidence-quality score and gap counts computed in codeProfile
They were model-emitted numbers and drifted between 62 and 74 on the same CV. Computed in code, ten runs returned 58 every time. Object transformer 1.15.0.
1 October 2026
Known gapRepeat-run test against two general modelsBoth
Thirty runs per product, set beside gpt-5-mini and gemini-3-flash-preview. On capability wording ReLoop.me was not more repeatable than either; the general models overlapped slightly more. The differences were in judgement: ReLoop.me did not echo a CV's self-ratings, and read a posting's seniority from its scope rather than its title. The claim that ReLoop.me is more consistent than general AI is not supported, and is not made. Figures in each product column.
27 September 2026
ChangedCompany and location typed into the Role Intel form are now readRole
Before this, both fields were ignored and the location could be read as the role title. Request builder 4.23.0.
26 September 2026
ChangedPay tables: Role Intel recalibrated, Profile Intel target roles cappedBoth
Role Intel's table was recalibrated against 788 dated external market points, with a provisional Consulting & Advisory family always published as Estimated (salary 3.58). Profile Intel now prices a role marked ready to move into at no more than one rung above the person's current level (salary 2.71).
25 September 2026
ChangedRole Intel picks a pay family from a closed listRole
The model chooses from 25 families. Where none fits, the brief says it is priced on the nearest one and names it. Pay fields are stored empty when the location cannot be anchored. Request builder 4.21.0, renderer 27.9.61.
23 September 2026
AddedRole Intel determinism re-measured, and the pay claim narrowedRole
This section carried a promise since 10 September that its figures would be
replaced by numbers rather than description. Four blind repeat runs of one real,
already-delivered job description against today’s deployed prompt now replace them.
Seven fields of twelve reproduced across all four runs; five moved. The two archetypes in play
were the same every time and which of them was dominant split two–two, with the
secondary the exact mirror each time — so the rubric shipped on 10 September appears to
have settled which types are candidates without settling which one is the answer. One claim
on this page has been narrowed as a result. It previously read that Role Intel’s money
reproduces and its headline word does not. The pay arithmetic does still reproduce to the euro
— but the seniority reading that feeds it moved on one run in four, which on the live table
is a 29% swing in the delivered figure on identical input, and the brief actually delivered
from this document carried the reading that came up once in four. Method mirrors the Profile
Intel re-measurement published the same day: live node code re-fetched, the real user message
pulled from execution history straight to disk and never hand-retyped, independent blind
completions. Same caveats — one document is not a rate, and the harness is not the
production caller.
23 September 2026
Known gapThe language lock can be overridden by the model on an English-jargon-dense CVProfile
A real Romanian CV, dense with English professional terminology (job titles,
functional-area names), got a deterministic detected_language="ro" lock at 0.90
confidence and an explicit instruction naming every narrative field the lock covers — and
the model wrote the whole brief in English anyway (0.95 confidence, 926 markers across all 181
narrative strings; every field, not one stray sentence). This is not the source-of-truth bug fixed
12 Sept (the lock correctly read the deterministic value, not a model echo) and not the buried-
instruction bug fixed 11 Sept (the instruction was unmissable). It is the underlying problem those
two fixes narrowed but did not eliminate: the model can still decline to follow an explicit
instruction. The G-5 merge guard, built 11 Sept for exactly this failure mode, caught it and held
the brief — undelivered, unpublished, no email sent. Working hypothesis for this document
specifically: the header and skills lines carry an unusually heavy load of English HR jargon kept
as-is in the Romanian original, which may pull the model into an English register despite the
lock. Tested offline against the real held document with a revised instruction naming that
jargon-density pattern explicitly as vocabulary, not a language signal: two independent runs, both
fully Romanian. Prepared as a versioned Build CIA Input patch; not yet deployed. See the open
backlog for the current line.
23 September 2026
AddedProfile Intel determinism re-measured on a second documentProfile
The 1 September identical-input study stood alone for three weeks. A second
measurement, on a different real, already-delivered document, is now on the page beside it,
reproduced offline (no live Tally submission) by reconstructing today's deployed prompt from the
live node code and pulling the real user message from n8n execution history. Read together rather
than as two separate verdicts: occupational family, geography and the inflation-risk label held
stable on both documents; the market-value figure and the coordination-scope label were unstable
on both; seniority band, the likely-offer figure and peak direct reports flipped — stable on
one document, unstable on the other. That last group is the finding: instability is
document-dependent, not a fixed property of a field, which the 1 September study alone could not
show. “Measured 1 Sep 2026” is now “Measured 23 Sep 2026.” The
18-character A/B study is untouched, still dated 1 September, not re-run this session.
23 September 2026
Changed“Different vocabularies” now has a definitionRole
One of the seven role types, Translator, required the role to sit between two
named groups “with different vocabularies”. That phrase was never defined, and one
advert flipped between Translator and Operator across four blind grader runs. The framing we
first reached for — whether the second group has to be named — turned out to
be the wrong question: the advert in question does name a second group, and both
dissenting graders rejected Translator anyway. Their actual objections were that the named
partners read as ordinary cross-functional colleagues, and that the deliverable named was a
customer-facing product rather than a document. Two groups now count as having different
vocabularies only when the posting requires the artifact to be restated to cross between them
— a regulatory requirement becoming a test scenario, a legal constraint becoming a product
rule. Teams that receive the same artifact in the same language are collaborators. And a
conversion artifact must be a document; an experience, a product or a business result is an
outcome, not an artifact. Deployed 23 September. This is a prompt change, so it is not yet
verified. The predicted effect was written down before deployment — the advert should
settle on Operator with no secondary — and repeat grading against a widened set has not yet
run. Until it does, the reproducibility figures above stand.
23 September 2026
Known gapA measurement we had taken could not be taken againRole
A queued fix rested on a figure measured in mid-September across roughly
twenty-two stored pipeline runs. When we went back to check that figure against a refined
definition, only three of those runs still existed — the execution store keeps
records for a short window, and nothing in the durable record held the values the figure was
built from. The figure is not reproducible and we are not shipping the fix that depended on
it. Six fields the engine already computed but discarded — the seat-leverage reading,
the scope-adjustment trace, and a new seniority coherence pair — are now written to the
durable record on every brief, so the question can be answered from stored data in future. This
entry is here because a measurement we cited and can no longer reproduce is exactly the kind of
claim this page exists to withdraw.
23 September 2026
ChangedTwo “unrecorded” studies flagged as possibly the same thing — they are notBoth
The Role Intel column flagged a question for whoever maintains this column: is the
unsourced sixteen-input reproducibility study named in the 1 September entry below the same study
as Role Intel's own sixteen-run archetype test (Builder six times, Operator six times, Fixer four
times)? It is not, on two grounds. Role Intel's study is dated 2 September 2026 — a day
after this engine's gap entry already recorded its own study as too old to date, which
rules out one being a late write-up of the other. And they measured different things: Role Intel's
study moved role_archetype, a Role Intel-only field with no Profile Intel equivalent,
where this engine's classification vocabulary is profile type, career pattern and the capability
ladder. The Profile Intel study stays unsourced and open; the two findings are separate.
23 September 2026
Changedhospitality.RO_bucharest.ic_senior.min tied to Role Intel's valueProfile
A 22 September fix on this engine raised this cell from 450 to 510 to clear an
inversion below ic_junior.min (500). Role Intel's own equivalent fix tied
ic_junior.min = ic_senior.min = 500 exactly, leaving the two engines ten EUR a month
apart on the same cell — flagged as cosmetic and offered as a low-effort follow-up, now
actioned. ic_senior.min: 510 → 500, mid 700 and max 1000 unchanged. This is a floor
tie, not an inversion, and it now matches Role Intel's live value exactly. Row re-checked at min,
mid and max: fully monotonic, zero new inversions anywhere in the table (the eight pre-existing
product-family inversions are unchanged and remain open). Shipped as CIA-salary v2.69.
23 September 2026
ChangedA stale version comment corrected — no data changedProfile
The CIA Index Emit node's decorative header comment read v0.2.0. The
constants it actually runs on — NODE_VERSION and INDEX_EMIT_VERSION,
both checked against the end-of-file marker — were already correctly at 0.3.0 /
0.3. Only the banner line was wrong, and only a human reading the source file would
ever have seen it; nothing a brief emits was affected. Corrected to read v0.3.0. No
version bump, because the real version was already right.
23 September 2026
Changedelementary.EU_west.ic_senior.min mirrored to the statutory floorProfile
Mirrors what Role Intel reports was approved on its own side. ic_senior.min:
1660 → 1422 — the French minimum wage (SMIC, 1422 EUR/month), the floor already used
at ic_junior.min for every EU_west row on this engine. mid 1660 and max 1950
unchanged. Ties ic_senior.min = ic_junior.min = 1422, a floor tie rather than an
inversion, the same pattern already used elsewhere in this table (finance_accounting,
hospitality). Zero new inversions; blast radius zero — the highest
salary_version on any of the 174 delivered briefs swept was still 2.63. Shipped as
CIA-salary v2.68.
23 September 2026
ChangedRegional Romanian pay sourced for marketing, not carried over from BucharestProfile
Four Romanian regional geographies in the marketing family — Cluj,
Timișoara, Iași, and a catch-all — priced sixty cells (ic_senior
through vp; ic_junior untouched) at sourced regional ratios against this
engine's own Bucharest anchor: Cluj ×1.20, Timișoara ×0.85, Iași ×0.95,
other ×0.70 (lower confidence). Two of the four ratios — Cluj and Iași — read
as at or above Bucharest rather than below it, which is not the intuitive direction; reviewed
against the sourced data and approved as sourced rather than adjusted to match instinct. These
ratios are anchor-independent by construction and were shared with Role Intel so the same
multipliers can be applied there if and when it adopts the same Bucharest anchor — that
adoption is a separate, not-yet-made decision on Role Intel's side. Zero new inversions. Shipped as
CIA-salary v2.67.
23 September 2026
Known gapfinance_accounting.EU_west.ic_senior.min traced, not fixed — staying openBoth
Role Intel's cross-check asked when this engine's ic_senior.min = 2500
was set, ahead of a possible reconciliation. Traced through this engine's own changelog history
rather than raw workflow-version archaeology: no change from v2.63 through v2.69 touches
ic_senior on this cell, and it already carried its current value before this session's
visible history begins. It is a legacy anchor, the same category as this page's other
documented 2021-era anchors, not something a recent fix introduced. Unlike the elementary fix
above, there is no sourced replacement value for this cell on either engine yet — Role
Intel's own figure for it is equally unsourced. Left as is, folded into the shared product-family
reconciliation work still open on both sides.
22 September 2026
ChangedSeventeen pay-table cells reconciled against the Profile Intel engineRole
The two engines keep separate pay tables with a common ancestor, and they had
drifted. A reconciliation pass moved seventeen cells in the Role Intel table. On thirteen of
them the Role Intel value was already the correct one and only the band edges moved. Two
larger proposed changes were not taken: one would have raised a senior individual
contributor floor above the manager floor in the same column, and one would have left the same
Bucharest role priced two-and-a-quarter times apart depending on whether the advert happened to
name the city. Both were sent back rather than shipped. The ladder is now checked at the bottom,
middle and top of every band rather than the middle alone — the defect that prompted this
was invisible to a middle-only check and was live on the other engine.
13 September 2026
AddedFractional and interim engagements are now recordedRole
The title map read “Head of” as a director-level title whether or not
the posting said fractional. Across twenty-one stored briefs, five had their band set from
the title alone and two of those were fractional roles priced at a full director band. The
words fractional, interim, part-time, contractor and freelance appeared nowhere in the pricing
logic. They are now detected and recorded. No figure changes. A band cap was considered and
rejected: the titled band drives what the employer is likely to offer, and capping it would make
that number wrong rather than right.
12 September 2026
ChangedA VP band can no longer be asserted over a seat with no reportsRole
Where a posting described more scope than its title implied, the engine could
raise the band by one level. Nothing stopped that raise landing on a VP band even when the engine's
own reading of the posting found no people management. Measured across twenty-two stored runs:
twelve upward adjustments, none downward, and three director-titled briefs raised to a VP band
on a seat with no reports. The raise is now blocked from landing on VP when the seat reads as
an individual contributor. Narrow on purpose — every other upward adjustment survives.
Forward-only: briefs already published keep the band they were published with.
10 September 2026
ChangedA job ad that describes no work now gets an answer that says soRole
The role-type rules covered two cases: the posting names one clear measure of
success, or it names none and two or more candidate types are genuinely in play. There was no
rule for the third case — a posting that names a title, a seniority, a company, a tool list
and a benefits list, and never describes the work the hire would be responsible for. The evidence
standard rules every one of those inadmissible, and the output still requires one of seven values,
so the model was left to supply one with nothing permitted to go on. In practice that means the
job title, which the same rules forbid. There is now a stated floor. On such a posting the
brief says the posting does not describe the work, names the closest type on whatever mandate
language does exist, assigns no secondary, and carries the lowest confidence the engine can
express. A floor call is a claim about the advert, not about the role. Deployed 10
September and not yet measured; like the change below it, the reproducibility figures above stand
until repeat runs replace them.
10 September 2026
Known gapThe engine build that produced a brief cannot be read back from the stored recordBoth
The version string stored with each brief names the renderer, the transformer,
the salary table and the index emitter — but not the prompt build that made the
classification. Every brief from 7 September onward records the same string across three
different prompt builds. This qualifies the entry below it. That entry says the stored
classification for every delivered brief is retained and the question is therefore answerable:
true of the classification, not of the build that produced it, which can currently be separated
only by the time the brief was created. Recorded 10 September 2026. The fix is one line and is
not shipped.
10 September 2026
WithdrawnTwenty-nine delivered brief pages have been taken down and no longer resolveBoth
Twenty-nine brief pages delivered during the private beta were withdrawn —
twenty Role Intel and nine Profile Intel. Every one of the twenty-nine has been checked and returns
404. If you hold a link to a brief from before the paid beta and it no longer opens, this is why,
and it is not a fault at your end. The pages were removed because they were produced by engine
versions whose claims this page can no longer stand behind; leaving them reachable would have left
unsupported figures in public. Ask at hello@reloop.me and the brief will be re-issued on the current
engine at no charge. The identifiers are retained internally so any link can be traced to what it
was. One of the nine Profile Intel identifiers is a literal placeholder pattern rather than a
generated identifier, and whether a brief was ever delivered under it has not been established
— that question is open and is recorded in the Profile Intel backlog above rather than closed
out quietly with the rest.
10 September 2026
ChangedThe seven types now carry a written evidence standard, not a bare list of namesBoth
Each type now carries a definition, the evidence necessary for it to apply
at all, the evidence sufficient to assign it decisively, and named patterns that resemble
the type and are not it. Job titles, seniority words, industry, culture language and tool lists are
excluded as evidence by rule. Above the seven sits one test — what would be missing from this
organisation if the role were left unfilled — and a stated order of resolution: the type
describing what the hire is accountable for is dominant, the type describing the method by which
that accountability is discharged is secondary, and a fixed order settles the remainder when a
posting names no single measure of success. This is deployed and not yet re-measured. The
reproducibility figures above stand until repeat runs replace them — a rubric that closes the
gap which plausibly caused the spread is a reason to expect improvement, not a result.
10 September 2026
Known gapThe change above alters some outputs, and the size of the shift is unmeasuredRole
Where a posting’s accountability and its dominant method point to different
types, the dominant type may now differ from what the same posting returned before 10 September.
The direction is intended — a role accountable for shipping a platform that also translates
regulation is a role for shipping a platform — but no corpus has been run, so we cannot say
how many briefs it affects. If you hold a brief from before 10 September and wonder whether it
would say something different today, the honest answer is that it might. The stored
classification for every delivered brief is retained, so this is answerable once repeat runs are
done, and it will be answered here.
7 September 2026
Withdrawn“Any inferred seniority forces Estimated” — as stated on 1 SeptemberRole
Seniority resolves four ways. The 1 September change caught two of them and this
page recorded the claim as complete. It missed the third: where a job title carries an ambiguous
token — lead, manager, principal, staff — the engine
deliberately declines to trust the title and takes the level from a model-generated signal instead.
Those briefs printed as Calibrated. The test looked at which fallback supplied
the level rather than at whether the level came from the title at all, and the model-generated path
matched neither string it was looking for. It now tests the basis directly, which has exactly one
non-inferred value. More published figures now read Estimated than did before.
That is the correct direction and it is what this page already promised. Briefs delivered between
1 and 7 September on an ambiguous title may carry the wrong confidence label; the figure itself is
unaffected. Ask and it will be re-issued.
7 September 2026
ChangedThe band a pay figure came from is now recorded with the figureRole
Role family, seniority band and geography were being written empty on every
delivered brief — the fields existed and were mapped, but to names three characters different
from the ones the engine emits. The effect was that no published figure could be traced back to the
table row it came from, on any brief, which made the pay claims on this page unauditable from the
stored record. They now populate, alongside the confidence tier and the raw seniority token the
lookup used. This does not fix the briefs already delivered, whose stored bands remain
empty; it fixes every brief from 7 September onward.
2 September 2026
Withdrawn“A pay figure shown as calibrated is calibrated for that market”Profile
The 1 September withdrawal covered locations the engine could not map at all —
those now show no figure. This is the other half, and it is worse, because nothing looks wrong.
The pay engine holds distinct bands for Southern Europe, the Nordics and Switzerland, and its own
city resolver identifies the location correctly. The classification step cannot emit those bands:
the permitted list has ten values and none of them is Southern Europe. So the profile is forced
onto a neighbouring band — a real band, with real data, from the wrong region — the figure prints
with a range and a gross equivalent, and the brief states it is calibrated. Found on a Rome
profile priced on Central and Eastern European bands, with the correct region sitting in the
engine's own resolution log. Any brief already delivered for a Southern-European, Nordic or
Swiss location carries a figure derived from a band that is not that market's. If you hold
one and want it re-issued or removed, ask. A fix is specified and not yet shipped.
2 September 2026
AddedThe Role Intel column is filled inRole
Vocabularies, the nineteen gate rules, measured determinism and the open backlog,
all read from the live workflow rather than from notes. Six of the eight outstanding items are now
answered. Two are answered with "not established" and stay that way: retention, and whether
the generation call ever ran without an explicit temperature — the retained workflow history only
reaches back to 1 September 2026, so there is nothing to read. The EU AI Act question is recorded
as open rather than inherited from the Profile Intel analysis, because a read of a role is not a
read of a person.
2 September 2026
Withdrawn“Deterministic code around two model calls”Both
Stated as shared architecture; it is not shared. Profile Intel runs two model
calls. Role Intel runs three — a calibrator, then generation, then the reviewer — and its
three run at 0.3, 0.3 and 0.3, where Profile Intel's reviewer runs at 0. A sentence written once
to avoid two copies disagreeing had instead made one description cover a system it did not
describe.
2 September 2026
Known gapTwo vocabularies are open on the Role Intel sideRole
gap_type has no fixed list at all — two runs of one input produced
eight distinct labels with a single overlap. And role_seniority_chip has no runtime
validator anywhere in the pipeline, in either product; the Profile Intel side has been
observed emitting C-level, which is off-vocabulary, and it reached a pay band. Named
here because the Role Intel column now claims a closed vocabulary for other fields, and that claim
is only worth something if the exceptions are named beside it.
2 September 2026
AddedMeasured Role Intel determinism, including the part that failsRole
Pay reproduces: identical inputs returned identical figures, one case checked to
the euro against hand-recomputed grid arithmetic. The role archetype does not: sixteen runs
of byte-identical input returned Builder six times, Operator six times and Fixer four times, and
six accidentally duplicated real submissions reproduced the dominant archetype in 3 of 5 pairs and
the secondary in 1 of 6. Pinning temperature to 0 made the spread worse. The cause is named — the
seven archetypes are given to the model as a bare list with no definitions — and the fix has not
been tried.
2 September 2026
ChangedA blocked Role Intel brief now records which rule blocked itRole
A delivered brief has always carried the full serialised gate findings. A
blocked one threw, and the throw replaced the output — so the reason vanished exactly when
it was needed. On 1 September a real brief was blocked and the stored reason was 77 characters of
the tail of one sentence, with every rule id cut away; identifying the rule required reconstructing
the arithmetic from the payload. The message is now a single line and the rule ids sit at the end,
where the truncation keeps them.
2 September 2026
ChangedThis page now covers both productsBoth
Restructured so Role Intel and Profile Intel sit side by side above the shared
engine. The Role Intel column is published empty and dated rather than filled with an
unverified summary: it lists the eight things that must be measured before anything is claimed
there. Setting the two columns beside each other makes the gap visible, which is the point.
2 September 2026
AddedThe retention promise has a mechanism — and one half of it does notBoth
A scheduled clear of stored CV text and the file reference was built for Profile
Intel. It is stated here with the part it does not deliver: the uploaded file itself sits
with the form provider, and removing our reference to it is not the same as deleting it there.
That step is not yet in place. Saying so is better than letting the promise run ahead of the
mechanism a second time.
1 September 2026
Withdrawn“Pay figures are calibrated for the market shown”Profile
Salary bands are held per occupational family and geography. A location that
did not resolve to one of the mapped regions fell through to the Central and Eastern European
band and printed a real euro figure from it, under a one-line caveat that the region was not
calibrated. A caveat above a specific number is not the same as declining to give one — the
number is what a reader takes away. From today those profiles show no pay figure at all, and say
why. This is forward-only. Briefs delivered before 1 September 2026 for unmapped
locations carry a figure derived from a European band that was not calibrated for their market,
and it stands on the page as it was published. If you hold one of those briefs and want it
re-issued or removed, ask.
1 September 2026
AddedHow much of the delivered corpus predates the temperature fixProfile
Counted against the engine fingerprint stored with each brief: 109 briefs
were generated before the fix and 9 after it, with 13 produced too early to carry a
fingerprint at all. Roughly five in six briefs delivered to date were produced at the API
default temperature of 1.0. An independent count by record date agrees to within one row.
This matters twice over. Both determinism studies on this page were run after the fix,
so they measure the engine as it is configured now and say nothing about how the earlier briefs
would reproduce. And it applies to any downstream use of the corpus, including as evaluation or
training material.
1 September 2026
AddedThis pageProfile
Published the vocabularies, the numbered gate rules and the measured
determinism status in one place, so a claim on a brief can be checked against a stated
definition rather than taken on trust.
1 September 2026
ChangedAI Act citation correctedBoth
Earlier wording cited Article 25 alone. Article 25 is the mechanism that
transfers provider obligations to whoever repurposes a system; it is not what makes the system
high-risk. The classification route is Annex III point 4(a) — AI used to evaluate candidates —
via Article 6(2). Both are now cited, along with the date those obligations actually apply,
which the earlier wording omitted and which materially changes what a reader should infer.
1 September 2026
ChangedGate-rule count stated as reported, not as runProfile
Twenty-nine numbered phases execute; fifteen entries are reported in the
machine-readable result. Publishing the larger number without that distinction would overstate
what the gate demonstrably checks per brief.
1 September 2026
Withdrawn“All classification vocabularies are closed”Profile
The management-tier scale in the experience timeline is specified as a list but
not locked, and its seven values do not map cleanly onto the six-rung scale the pay engine uses.
Found while investigating a two-band swing in the peak-level reading between two runs of the
same document. A fix is specified. The claim comes back when the fix ships and a regression test
holds it — narrowed in the meantime to the comparability table above, which names the fields the
claim still covers.
1 September 2026
Withdrawn“The same input produces the same output”Profile
Withdrawn on the strength of two controlled studies run today. The first put the
same document through the pipeline twice: seniority, family, geography and the pay figures
reproduced exactly, but peak direct reports moved from 1 to 4 and the coordination-scope label
came back as a different string. The second changed eighteen characters and moved the peak
management level two bands, from Director to Manager, along with the length of the claims audit.
The claim was made because the pipeline is deterministic code around a language model, and the
code is deterministic — but the classification step is not, and the claim did not
distinguish between them. Replaced by the measured status above, which names which fields hold
and which do not.
1 September 2026
AddedMeasured determinism figures, with the failed predictionProfile
Two runs of an identical document, and an eighteen-character A/B on the same
document. The A/B was designed to show a phrase was driving the seniority band; it showed the
opposite — the band held and the management tier moved instead. Recorded as it came out.
1 September 2026
Known gapAn earlier study is missing from this logProfile
A controlled reproducibility study of sixteen identical inputs was run before
the two above and found non-deterministic classification. It is not recorded here because we
cannot yet state its date or name the specific fields it moved, and this page does not carry
claims it cannot source. It will be added when both are established. We are naming the gap
rather than leaving it out, because a changelog that quietly omits an earlier finding is worth
less than one that says where it is incomplete.
28 August 2026
ChangedGeneration temperature pinned to 0.3Profile
Today the generation call runs at temperature 0.3 and the reviewer at 0.
Before 28 August 2026 the generation call set no temperature at all and ran at the API default
of 1.0 — the only model call without an explicit setting, while the reviewer beside it already
sat at 0. It writes the whole brief, structured fields included. Briefs produced before this date
were generated at a materially higher temperature than briefs produced after it, and are
correspondingly less likely to reproduce. This is the largest known determinism change in the
system, and it means briefs from either side of 28 August 2026 are not directly comparable
artefacts. The engine fingerprint recorded with each brief is what distinguishes them.