Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions client/cli/standard_session.py
Original file line number Diff line number Diff line change
Expand Up @@ -57,11 +57,11 @@ def standard_url_refusal(url):
binding pins by sha256.

A URL is not prose, so the word boundaries the scope test relies on are
not there: "vita492" carries no separator and no dot. Both sides are
not there: "examplestd10" carries no separator and no dot. Both sides are
stripped to letters and digits and matched as substrings, which is only
safe for a term distinctive on its own -- so the short ones are dropped.
"ansi", "vita", "2017" and "492" would each match half the web; what
survives is "vita492", "ansivita492", "2017r2024".
"std", "1" and "0" would each match half the web; what
survives is "examplestd10" and "2024".
"""
if not url or not STANDARD_ENGAGED_BEFORE:
return ""
Expand Down Expand Up @@ -120,7 +120,7 @@ def standard_binding_for(question):
It used to attach to EVERY prompt: the system rule, the standard.* tools
and the evidence policy came along whatever was asked. Asked to download
the RS274/NGC specification, the assistant answered "the system is bound to
ANSI-VITA-49.2, not RS274/NGC" eleven times over — the turn read as
EXAMPLE-STD-1.0, not RS274/NGC" eleven times over — the turn read as
normative because the question contained the word "standard".

Binding is announced when it happens. A frame that changes how an answer is
Expand Down
4 changes: 2 additions & 2 deletions client/cli/startup_banner.py
Original file line number Diff line number Diff line change
Expand Up @@ -191,8 +191,8 @@ def standard_row():

It decides how every normative answer is grounded, and it lives in ONE file
under the state directory, shared by every session on this machine. A
binding left behind by other work is therefore inherited in silence: a VITA
49.2 question was once answered against an RS274/NGC binding, and the only
binding left behind by other work is therefore inherited in silence: a
question for one standard was once answered against an RS274/NGC binding, and the only
thing that ever said so was the per-turn line, printed after the question
had already been asked. A session-wide property belongs with the other
session-wide properties.
Expand Down
22 changes: 11 additions & 11 deletions client/doc/source/standard_rag.rst
Original file line number Diff line number Diff line change
Expand Up @@ -92,7 +92,7 @@ data-format standard (bit, octet, offset, field, width, encoding…). The
generic words *standard*, *revision*, *rule* and *section* are excluded on
purpose: bound to every prompt, the frame turned "get the RS274/NGC PDF so we
can inject it as a standard" into eleven turns explaining that the system is
bound to ANSI-VITA-49.2. Erring towards binding remains the safe direction —
bound to EXAMPLE-STD-1.0. Erring towards binding remains the safe direction —
an unnecessary binding costs a system rule and two tool schemas, a missing one
costs grounding on a licensed source.

Expand Down Expand Up @@ -194,8 +194,8 @@ Operator flow

::

/standard ingest ./authorized.pdf --id ANSI-VITA-49.2 --revision 2017-R2024
/standard use ANSI-VITA-49.2 2017-R2024
/standard ingest ./authorized.pdf --id EXAMPLE-STD-1.0 --revision 2024
/standard use EXAMPLE-STD-1.0 2024
/standard status
/standard rebuild
/standard unbind
Expand Down Expand Up @@ -237,11 +237,11 @@ retention requires the explicit ``--retain-pdf`` ingestion option.
Re-extracting an already-ingested revision goes through a candidate rather than
over the active corpus::

/standard ingest ./authorized.pdf --id ANSI-VITA-49.2 --revision 2017-R2024 --candidate --layout
/standard candidates ANSI-VITA-49.2 2017-R2024
/standard promote-candidate ANSI-VITA-49.2 2017-R2024 cand-<id>
/standard rebuild ANSI-VITA-49.2 2017-R2024
/standard use ANSI-VITA-49.2 2017-R2024
/standard ingest ./authorized.pdf --id EXAMPLE-STD-1.0 --revision 2024 --candidate --layout
/standard candidates EXAMPLE-STD-1.0 2024
/standard promote-candidate EXAMPLE-STD-1.0 2024 cand-<id>
/standard rebuild EXAMPLE-STD-1.0 2024
/standard use EXAMPLE-STD-1.0 2024

Promotion archives the previous corpus as a recoverable generation, drops the
derived indexes, and clears the binding so existing sessions fail closed.
Expand All @@ -253,7 +253,7 @@ A standard's contents table names each clause and the page it opens on. That is
the one statement about structure that does not come from the extractor, so it is
what an operator should trust over any internal verdict::

/standard verify ANSI-VITA-49.2 2017-R2024
/standard verify EXAMPLE-STD-1.0 2024

The check measures the offset between printed and extracted page numbers from the
clauses that are not in dispute, then reports how many headings land on the page
Expand All @@ -269,7 +269,7 @@ An automated extraction review is evidence, not approval. The operator walks the
private review file row by row and records a human verdict beside the automated
one::

python -m standard_review --standard ANSI-VITA-49.2 --revision 2017-R2024
python -m standard_review --standard EXAMPLE-STD-1.0 --revision 2024

Keys are ``p`` PASS, ``w`` ACCEPTABLE_WARNING, ``f`` FAIL, ``n``
NEEDS_FOLLOWUP, ``s`` skip, ``u`` clear a verdict, ``b`` back, ``q`` save and
Expand All @@ -293,7 +293,7 @@ Completing the review does not approve the corpus. Approval is a separate
operator action, refused unless every row is human-reviewed with no ``FAIL`` and
no ``NEEDS_FOLLOWUP``::

/standard approve ANSI-VITA-49.2 2017-R2024
/standard approve EXAMPLE-STD-1.0 2024

It sets ``human_validation_status`` to ``APPROVED`` in the canonical manifest and
writes reviewer, timestamp and summary to a private ``approval.json`` beside it.
Expand Down
6 changes: 3 additions & 3 deletions client/eval/modeluse/modeluse_harness.py
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
"""MODEL_USE_V1 harness: drive the real model against the frozen normative tools.

Generic on purpose. It knows how to run a question through the live tool loop
and record everything that happened; it knows nothing about VITA. The cases and
and record everything that happened; it knows nothing about the standard. The cases and
their ground truth live beside it and stay private, because they quote the
document being evaluated.

Expand Down Expand Up @@ -34,8 +34,8 @@
# Which binding the battery drives. Overridable so the SAME cases, prompts
# and policy can be run against a second extraction of the same document --
# the store itself is selected by SPEAR_STANDARDS_ROOT.
SID = os.environ.get("SPEAR_EVAL_STANDARD") or "ANSI-VITA-49.2"
REV = os.environ.get("SPEAR_EVAL_REVISION") or "2017-R2024"
SID = os.environ.get("SPEAR_EVAL_STANDARD") or "EXAMPLE-STD-1.0"
REV = os.environ.get("SPEAR_EVAL_REVISION") or "2024"
ENDPOINT = "http://127.0.0.1:8082/v1/chat/completions"
MODEL = "qwen3"
# Deterministic as far as llama-server allows.
Expand Down
20 changes: 10 additions & 10 deletions client/evidence/conformance_guard.py
Original file line number Diff line number Diff line change
@@ -1,9 +1,9 @@
"""What a bound answer may say about compliance, checked against what it read.

Bound to ANSI-VITA-49.2, asked to validate the acknowledge path, a session ran
Bound to EXAMPLE-STD-1.0, asked to validate the acknowledge path, a session ran
twelve greps, called no normative tool of its own, and concluded:

The implementation is correct and compliant with ANSI/VITA 49.2-2017 (R2024).
The implementation is correct and compliant with EXAMPLE-STD-1.0.

It is not. Rule 8.4.1.1-3 is not implemented, Req-V is never honoured, the
execution acknowledge is unconditional. None of the guards before this one had
Expand All @@ -16,7 +16,7 @@
sentence at all, and the distinction is the one evidence_guard already lives on:

scoped Rule 8.4.1.1-2 is satisfied: the encoder sets exactly one bit
unscoped the implementation correctly follows the VITA 49.2 specification
unscoped the implementation correctly follows the EXAMPLE-STD-1.0 specification

The first names a clause and a fact. The second quantifies over the whole
standard, which no finite set of retrievals establishes. So an affirmative
Expand Down Expand Up @@ -66,7 +66,7 @@
r"(?:correctly|properly)\s+" + _ACT
# "implements it CORRECTLY": the adverb after the verb is as common as
# before it, and only the first form was caught.
# The gap is [\w.\-]+, not \w+: "implements the VITA49.2 ACK management
# The gap is [\w.\-]+, not \w+: "implements the EXAMPLE-STD-1.0 ACK management
# correctly" has a dot in the middle of it, and \w+ stops there.
+ r"|" + _ACT + r"\s+(?:[\w.\-]+\s+){0,5}?(?:correctly|properly)"
# "is satisfied" and "is met" are how a rule is reported as done, and
Expand Down Expand Up @@ -112,11 +112,11 @@
re.I)

# Naming the bound standard as the thing complied with -- "compliant with
# ANSI-VITA-49.2", "follows the VITA 49.2 specification" -- without a clause
# EXAMPLE-STD-1.0", "follows the EXAMPLE-STD-1.0 specification" -- without a clause
# narrowing it.
#
# Built from the binding, never from a list. The first version spelled VITA
# 49.2 into the pattern, which made a guard that exists to catch an unscoped
# Built from the binding, never from a list. The first version spelled the
# standard id into the pattern, which made a guard that exists to catch an unscoped
# claim work for exactly one standard and silently pass the same sentence
# about any other. standard_scope already derives a subject's names from its
# own identifier; this is the same rule, and the reason is the same: a
Expand All @@ -143,7 +143,7 @@ def bare_standard(binding=None):
def standard_names(binding=None):
"""The names this standard goes by, as one alternation, longest first.

Longest first so "ansi-vita-49.2" wins over "vita", and each name is
Longest first so "example-std-1.0" wins over "std", and each name is
spelled loosely so the writer's punctuation does not decide the match.

Public because two guards need the same spelling: this one asks whether
Expand Down Expand Up @@ -204,9 +204,9 @@ def _identity_terms(binding):
def names_standard_as_object(sentence, pattern):
"""Is the bound standard what this sentence claims conformance TO?

"conforms to ANSI-VITA-49.2, see Rule 8.4.1.1-2" claims the whole
"conforms to EXAMPLE-STD-1.0, see Rule 8.4.1.1-2" claims the whole
document and cites one rule as support; the citation does not narrow it.
"Rule 8.4.1.1-2 of ANSI-VITA-49.2 is satisfied" claims the rule, and
"Rule 8.4.1.1-2 of EXAMPLE-STD-1.0 is satisfied" claims the rule, and
names the standard only to say which rule it is.

The difference is positional and that is the whole of it: in the second,
Expand Down
2 changes: 1 addition & 1 deletion client/evidence/conformance_mode.py
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
"""A turn that asks whether code conforms is shaped, not merely filtered.

Asked to "validate the ACK implementation against VITA49.2", a bound session
Asked to "validate the ACK implementation against EXAMPLE-STD-1.0", a bound session
listed the structure registry, read all nine of its structures -- Tilt Angle,
Bandwidth, Manufacturer OUI, none of them about acknowledgement -- and rendered
them under "Validated Structures (PASS/ACCEPTABLE)". It never opened a source
Expand Down
2 changes: 1 addition & 1 deletion client/evidence/project_build.py
Original file line number Diff line number Diff line change
Expand Up @@ -103,7 +103,7 @@ def from_dict(cls, raw):
def _cmake_side(root):
"""A CMake option this tree needs before it will configure.

The VITA 49.2 converter has two sides and the target one wants a vendor
The EXAMPLE-STD-1.0 converter has two sides and the target one wants a vendor
SDK that is not on a workstation; configuring without -DV492C_SIDE=host
stops on "libusb-1.0 not found" and the turn concludes the project cannot
be built. Read from the file rather than assumed: an option with a
Expand Down
2 changes: 1 addition & 1 deletion client/normative/normative_evidence.py
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@

{"normative_checks": [{
"id": "ack-subtype",
"standard": "ANSI-VITA-49.2", "revision": "2017-R2024",
"standard": "EXAMPLE-STD-1.0", "revision": "2024",
"provisions": ["Rule 8.4.1.1-2@9f4e826caf39"],
"command": "ctest --test-dir build -R ack_subtype",
"evidence": {"kind": "test", "success": "exit_zero",
Expand Down
2 changes: 1 addition & 1 deletion client/normative/normative_precedence.py
Original file line number Diff line number Diff line change
Expand Up @@ -145,7 +145,7 @@
# standard's own names are added from the binding, never spelled here.
_GENERIC_SUBJECT = r"(?:the|this)\s+(?:bound\s+)?(?:standard|specification|spec)"

# Between subject and verb: "the standard explicitly requires", "VITA 49.2
# Between subject and verb: "the standard explicitly requires", "EXAMPLE-STD-1.0
# section 8.4 states". Bounded, so a subject in one clause of a sentence does
# not capture a verb belonging to another.
_GAP = r"(?:\s+[\w.,§\-()]+){0,4}?\s+"
Expand Down
4 changes: 2 additions & 2 deletions client/standard/standard_answer_policy.py
Original file line number Diff line number Diff line change
Expand Up @@ -51,11 +51,11 @@
# The same retrieval as the zero-tool bootstrap, made before the model has said
# anything rather than after it has finished.
#
# Bound to ANSI-VITA-49.2 and asked how ACK is managed, a session spent twelve
# Bound to EXAMPLE-STD-1.0 and asked how ACK is managed, a session spent twelve
# bash rounds in the C sources and made not one normative call. The bootstrap
# fired at the boundary, as designed, and searched the standard once -- into a
# context already holding twelve grep outputs and a finished answer. The answer
# came back saying the implementation "correctly follows the VITA 49.2
# came back saying the implementation "correctly follows the EXAMPLE-STD-1.0
# specification". It does not.
#
# A retrieval that arrives after the conclusion is a rebuttal, and the model
Expand Down
20 changes: 10 additions & 10 deletions client/standard/standard_scope.py
Original file line number Diff line number Diff line change
@@ -1,17 +1,17 @@
"""Does THIS question engage the bound standard?

The binding used to attach to every prompt of every session. Once an operator
bound ANSI-VITA-49.2, each turn carried the "BOUND STANDARD … normative claims
bound EXAMPLE-STD-1.0, each turn carried the "BOUND STANDARD … normative claims
must use retrieved canonical sources" system rule, the standard.* tools, and
the evidence policy that fires when such a turn is about to answer with nothing
retrieved.

Asked to download the RS274/NGC G-code specification, the assistant therefore
spent a whole task explaining that "the system is bound to ANSI-VITA-49.2, not
spent a whole task explaining that "the system is bound to EXAMPLE-STD-1.0, not
RS274/NGC" — correct, irrelevant, and repeated eleven times. The word that
dragged it there was "standard", in "inject it in our RAG as a standard": it
sits in the normative vocabulary, so the turn read as normative and the
orchestration fetched VITA evidence for a question about a PDF download.
orchestration fetched standard evidence for a question about a PDF download.

The binding is decided per turn now, from two kinds of sign:

Expand All @@ -33,15 +33,15 @@
about the turn's surroundings rather than its words.

A file the operator put in front of the model counts. "carry out the task in
doc/ack-task.md" names nothing; the file it points at said ANSI-VITA-49.2
2017-R2024 in its third paragraph, and the turn ran free. Only the standard's
doc/ack-task.md" names nothing; the file it points at said EXAMPLE-STD-1.0
2024 in its third paragraph, and the turn ran free. Only the standard's
NAME is looked for there — never the subject nouns, because any C file says
"bit" and "packet" and that would bind every turn after reading one.

And a session that has engaged the standard once stays engaged until /clear.
That began as a narrow rule — follow-up vocabulary, anaphora — and the narrow
rule kept being outrun. "can you validate the implementation ?" ran free and
declared the code "correct and compliant with ANSI/VITA 49.2-2017 (R2024)" on
declared the code "correct and compliant with EXAMPLE-STD-1.0" on
the strength of twelve greps; it is not. Later a stray "/quit" reached the
model as a prompt, matched no follow-up word either, and the unbound turn that
followed went looking for the specification on the public web and edited five
Expand Down Expand Up @@ -150,8 +150,8 @@ def asks_what_the_document_settles(text):
def identity_terms(binding):
"""Every way a user might name this standard, taken from the binding.

ANSI-VITA-49.2 / 2017-R2024 yields ansi-vita-49.2, vita, 49.2, vita-49.2,
vita 49.2, 2017-r2024, r2024 … Single short parts with no digit are
EXAMPLE-STD-1.0 / 2024 yields example-std-1.0, example, 1.0, example-std,
std-1.0, 2024 … Single short parts with no digit are
dropped: they are initials, and matching them would bind on noise.
"""
terms = set()
Expand All @@ -168,7 +168,7 @@ def identity_terms(binding):
terms.update(part for part in parts
if len(part) > 3 or any(c.isdigit() for c in part))

# Adjacent pairs, so "vita 49.2" and "vita-49.2" both land even though
# Adjacent pairs, so "example std" and "example-std" both land even though
# the identifier spells it with one separator.

for first, second in zip(parts, parts[1:]):
Expand All @@ -188,7 +188,7 @@ def engages(binding, question, *, engaged_before=False, context=""):
this turn without typing it -- a !read file. Only
the standard's own NAME counts there: "carry out
the task in doc/ack-task.md" engaged nothing, while
the file it named said "ANSI-VITA-49.2 2017-R2024"
the file it named said "EXAMPLE-STD-1.0 2024"
in its third paragraph. The subject nouns are not
consulted in context on purpose: any C file says
"bit" and "packet", and that would bind every turn
Expand Down
2 changes: 1 addition & 1 deletion doc/source/reasoning/standards.rst
Original file line number Diff line number Diff line change
Expand Up @@ -258,7 +258,7 @@ without a restriction.
A document with no role labels
==============================

Most documents do not label their provisions the way a VITA-style standard
Most documents do not label their provisions the way a rule-labelled standard
does, and they do not need to. Where there are no labels, the kind is read from
the **modal verb**: *must* makes a requirement, *should* a recommendation,
*may* a permission, and prose with no modal verb stays informative.
Expand Down
Loading