ANS-100/P Anchor Amendments — Batch 1 (v2.0 → v2.1, ratified)

Drafted 30 Aug 2026 by Rater 1 · STATUS: RATIFIED 30 August 2026 (all ten: A1–A10) · Companion to Index_Mapping_and_Anchors_v2_ACTUAL

0. Governance note, stated honestly

These amendments were drafted by rater 1 after scoring three companies, which creates a known risk: anchors can be quietly fitted to the scores already given. Mitigation is procedural — the author should review each amendment against one question only: does it make two blind raters more likely to land on the same score? Not whether the resulting numbers feel right. Where an amendment changes an existing score, the change is computed and disclosed below; the three company files are updated only after sign-off, with an amendment note in each. Rater 2 scores against v2.1 from evidence packs only, never against these files.

Ratification checklist: A1 ☐ · A2 ☐ · A3 ☐ · A4 ☐ · A5 ☐ · A6 ☐ · A7 ☐ · A8 ☐ · A9 ☐ · A10 ☐

A1 — Q25 (D6): what "material RPE improvement with attribution" means

Problem. Infosys (~6% RPE growth, tools named), TCS (~3.5%, headcount-driven), HCLTech (~4.5%, partial) all landed at 2 with rater discomfort in three directions. The 2/3 line needs a test, not a feeling.

Amended anchor text (append to D6 level 3):

Level 3 requires management attribution that is quantified and connected to the reported economic metric — a disclosed number linking named AI workflows or tools to the productivity or revenue-per-employee outcome (e.g., "X% of the delivery-efficiency gain is attributable to platform Y"). Naming tools while leaving the magnitude unquantified, or citing "automation and AI" within general margin commentary, is level 2 regardless of the size of the RPE movement. RPE improvement achieved primarily through headcount reduction without workflow-level attribution is level 2 regardless of magnitude.

Effect on existing files. None. Infosys's borderline resolves cleanly to 2 (tools named, magnitude unquantified at company level); TCS and HCLTech confirmed at 2. The Infosys/TCS mirror-pair now has a written answer.

A2 — Q30 (D7): segment-level pricing disclosures, and the fixed-price trap

Problem 1. TCS's "80% outcome-based" covers one segment; the anchor is silent on segment-scope evidence. Problem 2 (from the Infosys run). Majority-fixed-price contract form can masquerade as "priced on scope, not hours."

Amended anchor text (append to D7 anchors, Q30 guidance):

Pricing-model evidence is scored at the reporting-entity level. A segment-level disclosure supports level 2 where the segment is a disclosed, material share of revenue; level 3 requires that segments with majority scope- or outcome-based pricing sum to a majority of company revenue on disclosed figures. Undisclosed segment weights resolve downward. Fixed-price contract form counts toward scope-based pricing only where the company does not simultaneously disclose labour-linked productivity pass-throughs on those contracts; disclosed pass-through behaviour indicates effort-linked pricing economics regardless of contract form.

Effect on existing files. None — TCS Q30:2 and Infosys Q30:2 confirmed; both files' reasoning becomes citable anchor text.

A3 — Q18 (D4): the gradation rewrite — THE ONE THAT MOVES SCORES

Problem. Three companies, materially different responsible-AI evidence, identical score of 2: TCS (productised client framework) < Infosys (ISO 42001 certified + open-sourced RAI toolkit) < HCLTech (ISO 42001 certified with published scope across its own delivery model). D4 currently cannot discriminate across the entire IT slice.

Amended anchor text (replaces Q18 /P guidance):

• 1 — Generic responsible-AI statement; boilerplate, no structures.
• 2 — Published responsible-AI or AI-security framework with genuine specifics, where the evidence is a client offering or an unaudited self-declaration. Productised frameworks land here regardless of sophistication.
• 3 — Independently audited or certified AI management system (e.g., ISO/IEC 42001) whose published scope covers the company's own operations or delivery model; or AI surfaces explicitly within a public bug-bounty scope; or published, enforced controls with specifics for the company's own systems.
• 4 — Not publicly establishable (P-cap holds at 3). Multi-question use of one artifact: a single evidence artifact (e.g., an ISO 42001 certification) may inform both Q18 and Q22 only because the questions test different things — Q18 the assured posture on AI surfaces, Q22 the enforced governance ownership. Where one artifact supports both, note it in the company file.

Effect on existing files (the honest part). - Infosys Q18: 2 → 3 (audited system covering own operations, plus open toolkit). D4: 58% → 67%. Total 66.4 → 67.6. - HCLTech Q18: 2 → 3 (audited system, published scope explicitly across own delivery model). D4: 50% → 58%. Total 61.9 → 63.2 (63.3 with A6/Q34). - TCS Q18: 2 unchanged (productised framework, unaudited own posture). Total 63.9 unchanged. - Revised ordering: Infosys 67.6 · TCS 63.9 · HCLTech 63.3. All AI Native; #2/#3 gap narrows from 2.0 to 0.6 points. This is D4 doing its job, not score drift — but it is exactly the kind of change the author must ratify with eyes open.

A4 — Q24 (D5): when AI investment counts as "infrastructure," with a bright line

Problem. TCS's 4 could read as 3 if "AI investment as infrastructure" means operating infrastructure rather than a data-centre subsidiary.

Amended anchor text (append to D5 guidance, Q24):

Level 3: recurring AI investment disclosed as a named line item in capital-allocation or margin guidance. Level 4 (reachable — Q24 is an O-question): AI investment structurally institutionalised in balance-sheet-visible, multi-year, board-approved form — a subsidiary, JV, or disclosed capex programme — where cumulative committed amounts are ≥5% of trailing-twelve-month revenue. A single-year platform budget never reaches 4; the question asks whether AI is treated as infrastructure, and a capitalised commitment at material scale is the strongest public form of that treatment.

Effect on existing files. TCS Q24:4 stands (HyperVault commitments ≈ 20%+ of TTM revenue). HCLTech Q24:3 stands (₹3,500 Cr initial ≈ 2.9%). Infosys Q24:3 stands. The bright line converts the flagged judgment into arithmetic.

A5 — Q27 (D6) and the evidence floor: how analyst validation counts

Problem. HCLTech's 65%/90% automation rates trace to one ISG assessment; the two-item evidence floor is ambiguous about analyst reports, and republications can masquerade as corroboration.

Amended rule text (append to §3 scoring rules):

A third-party analyst assessment is ONE evidence item, of higher grade than self-declaration: it satisfies the "independent" half of the two-item floor when paired with any company-primary disclosure of the same fact. Two documents sharing a single root source (a company page quoting the analyst report; syndicated republications) count as one item. Automation rates disclosed on client-delivery workflows are scoreable in D6 for services companies — delivery is their operations — but are held one anchor level below what the numeric rate alone would indicate unless the company discloses the rate as representative of the workflow category rather than a showcase deployment.

Effect on existing files. HCLTech Q27:2 confirmed by rule rather than caution. Rater 2's open question ("does ISG merit a 3?") now has a written answer: not on a single root source describing showcase deployments.

A6 — The N→P reclassification rule (the Q34 ruling)

Problem. Internal deflation modelling is classed N (unobservable), but HCLTech's CEO publicly disclosed both the model's existence and its output (2–3% annual services deflation). Refusing to score a volunteered disclosure punishes transparency — the inverse of the Index's incentive design.

Proposed rule (new §2a in the mapping doc):

An N-classed question becomes scoreable for a specific company when the company itself has publicly and voluntarily disclosed the normally-internal fact in attributable, dated form — its own filings, releases, or named-executive interviews. The question reclassifies to P for that company only; the P-cap of 3 applies; the reclassification and its evidence are noted in the company file; coverage counts include it. Third-party claims, leaks, or inference never trigger reclassification. Expect this to recur: voluntary disclosure of "unobservable" facts is characteristic of maximal disclosers, and the rule rewards exactly the behaviour the Index exists to encourage.

Effect on existing files. HCLTech Q34:3 admitted; coverage 25/35; total +0.15 (folded into the 63.3 above). No band changes. Infosys and TCS files re-checked: no volunteered N-disclosures found in their packs — no retroactive scoring.

A7 — The 2-vs-3 "systematic" test for P-questions (the honest-range compressor)

Problem. Q6, Q8, Q10, Q32, Q35 all sit at the P-cap of 3 with "a stricter reading gives 2" — the single largest source of the ±4-point honest ranges that straddle the band boundary in all three files.

Amended rule text (append to §3, applying to all P-questions at level 3):

Level 3 on a P-question requires all three of: (a) at least three distinct evidenced instances — functions, deployments, or deals, not three documents about one instance; (b) at least one instance in a revenue- or delivery-critical path; (c) at least one evidence item of third-party or primary-filing grade, not company marketing alone. Two instances, or marketing-grade evidence only, is level 2. This converts the "systematic practice" judgment into a countable test.

Effect on existing files. All current 3s survive the countable test (verified: Infosys — three internal agent deployments incl. finance, GitHub artifact grade; TCS — four client functions, Avasant third-party; HCLTech — three functions incl. ITOps delivery core, ISG third-party). Scores unchanged — but the honest ranges tighten from ~57–67 to ~61–68, because the downside readings are no longer available to a rater applying the test. This is the amendment that most improves the inter-rater property.

A8 — Band-boundary publication rule (from the structural finding)

Problem. All three honest ranges straddled the 61-point AI Integrated / AI Native line; the P-cap plus evidence floor concentrates maximal-discloser IT-services scores exactly there. Silent band placement would hide a known instability.

Proposed rule (publication format, methodology §7):

Where a company's post-amendment honest range crosses a band boundary, publish the band of the point score with a boundary marker — "AI Native (boundary)" — and state the range in the company note. Add to the methodology page: "For IT services companies with maximal disclosure, /P scores concentrate near the AI Integrated / AI Native line; this is a property of the public-evidence caps, not of the companies, and band placements at that line carry the boundary marker."

Effect. Post-A7, TCS and HCLTech likely still carry the marker; Infosys at 67.6 likely does not. Determined finally after blind second ratings.

A9 — Evidence-pack completeness floor (process rule)

Problem. HCLTech Q14 and Q21 were flagged "likely underscored — annual-report gap in the pack," which is an admission the pack was incomplete at scoring time. A score that its own rater suspects is a pack artifact is not defensible.

Proposed rule (new step 0 in the seven-step worked discipline):

Before any dimension is scored: the pack must contain the latest integrated annual report, the last four quarterly releases plus at least one earnings-call transcript, the company's AI/platform documentation pages, and at least one third-party analyst assessment where one exists. "Likely underscored due to pack gap" is not a permitted flag — fill the gap or leave the question unscored with the coverage consequence.

Required action before rater 2 receives the HCLTech pack: pull the FY26 Integrated Annual Report sections on AI training figures (Q21) and eval/benchmarking practice (Q14); re-score both against v2.1. Until then the HCLTech file carries "pack-incomplete" status on those two questions.

A10 — Disclosure-bias statement (methodology page insertion)

From the Infosys run, now confirmed three-for-three: 24–25/35 coverage appears to be the practical maximum, achieved only by maximal disclosers.

Proposed methodology text:

Grade-A confidence correlates with company transparency, not company quality. A Grade-C company is not worse than a Grade-A company; it is less visible. The confidence grade prices that difference; the band-range publication format carries it; and the Full assessment exists to close it.

Summary of score effects (pending sign-off)

No band changes. The #2/#3 gap narrows to 0.6 points — within blind-rating noise, which the second ratings must resolve.

On sign-off, in order

1. Author ratifies/edits each amendment (checklist above).
2. Rater 1 updates the three company files against v2.1, each carrying an "amended under Batch 1" note.
3. Run the A9 completeness pull for HCLTech Q14/Q21 (and spot-check TCS/Infosys packs against the same floor).
4. Freeze v2.1 as the anchor set rater 2 scores against — no further amendments until after the blind ratings, so the inter-rater test measures the anchors, not a moving target.

Published 9 September 2026 under CC BY-SA 4.0. Status changed from "proposed" to "ratified" on publication per the v2.1 Ratification & Freeze Record of 30 August 2026. No text is changed.